Methods, apparatuses, and systems for continuous muscle monitoring and health status prediction

Abstract

Various examples of the present disclosure provide example methods, apparatuses and systems for continuous muscle monitoring and health status prediction. For example, an example wearable muscle monitor may include an inertial measurement unit (IMU) configured to generate global motion signals associated with a user, at least one near-infrared spectroscopy (NIRS) sensor configured to generate local activity signals associated with the user, and a wireless microcontroller communicatively coupled to the IMU and the at least one NIRS sensor. Additionally, or alternatively, an example wearable muscle monitor may include a sensor array that includes a plurality of sensing pixels and a computing device communicatively coupled to the sensor array.

Description

[0001] METHODS, APPARATUSES, AND SYSTEMS FOR CONTINUOUS MUSCLE MONITORING AND HEALTH STATUS PREDICTION

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS.

[0003] [1] This application claims the benefit of and priority to United States Provisional Patent Application No. 63 / 717,662, filed November 7, 2024, the entire content of which is incorporated by reference in its entirety.

[0004] FEDERALLY SPONSORED RESEARCH AND DEVELOPMENT

[0005] [2] This invention was made with government support under Grant Nos. ECCS-2139659 and ECCS-2443105 awarded by National Science Foundation and Grant No. EB034332 awarded by National Institutes of Health. The government has certain rights in the invention.

[0006] BACKGROUND OF THE INVENTION

[0007] [3] There are many technical challenges and difficulties associated with muscle monitoring and health prediction.

[0008] SUMMARY OF THE INVENTION

[0009] [4] In accordance with various embodiments of the present disclosure, a wearable muscle monitor is provided. In some embodiments, the wearable muscle monitor comprises an inertial measurement unit (IMU) configured to generate global motion signals associated with a user; at least one near-infrared spectroscopy (NIRS) sensor configured to generate local activity signals associated with the user; and a wireless microcontroller communicatively coupled to the IMU and the at least one NIRS sensor. In some embodiments, the wireless microcontroller is configured to: generate one or more global motion data points based on the global motion signals, generate one or more local activity data points based on the local activity signals, and transmit the one or more global motion data points and the one or more local activity data points to a computing device.

[0010] [5] In some embodiments, the wearable muscle monitor further comprises a flexible substrate. In some embodiments, the IMU, the at least one NIRS sensor, and the wireless microcontroller are disposed on the flexible substrate. In some embodiments, the flexible substrate comprises a main island and at least one daughter island. In some embodiments, the main island and the at least one daughter island are connected through at least one stretchable serpentine hinge.

[0011] [6] In some embodiments, the IMU, the at least one NIRS sensor, and the wireless microcontroller are disposed on the main island. In some embodiments, the at least one NIRS sensor is disposed on the at least one daughter island.

[0012] [7] In some embodiments, the at least one daughter island comprises a left daughter island and a right daughter island. In some embodiments, the left daughter island is connected to a left side of the main island. In some embodiments, the right daughter island is connected to a right side of the main island.

[0013] [8] In some embodiments, the at least one NIRS sensor comprises: a central NIRS sensor disposed on a top portion of the main island; a bottom NIRS sensor disposed on a bottom portion of the main island; a top left NIRS sensor disposed on the left daughter island; and a top right NIRS sensor disposed on the right daughter island.

[0014] [9] In some embodiments, the one or more global motion data points indicate motions associated with a head of the user. In some embodiments, the one or more local activity data points indicate activities associated with a neck muscle of the user.

[0015]

[0010] In accordance with various embodiment of the present disclosure, a computerimplement method is provided. In some embodiments, the computer-implemented method comprises: receiving, by a processor, one or more global motion data points and one or more local activity data points associated with a user; and generating, by the processor, one or more predicted health status data points associated with the user based at least in part on the one or more global motion data points, the one or more local activity data points, and a dual-channel recurrent neural network (RNN).

[0016]

[0011] In some embodiments, the one or more global motion data points are based on global motion signals from an IMU. In some embodiments, the one or more local activity data points are based on local activity signals from at least one NIRS sensor.

[0017]

[0012] In some embodiments, the dual-channel RNN is based on gated recurrent units (GRU). In some embodiments, the dual-channel RNN comprises two parallel hidden layers. In some embodiments, the two parallel hidden layers comprise a first hidden layer and a second hidden layer. In some embodiments, the computer-implemented method comprises inputting the one or more global motion data points to the first hidden layer and inputting the one or more local activity data points to the second hidden layer. In some embodiments, outputs from the two parallel hidden layers are concatenated and sent to a fully connected (FC) layer.

[0018]

[0013] In some embodiments, the one or more global motion data points indicate motions associated with a head of the user. In some embodiments, the one or more local activity data points indicate activities associated with neck muscle of the user.

[0019]

[0014] In some embodiments, the one or more predicted health status data points indicate one or more predicted coughing events associated with the user.

[0020]

[0015] In accordance with some embodiments of the present disclosure, a wearable muscle monitor is provided. In some embodiments, the wearable muscle monitor comprises a sensor array and a computing device communicatively coupled to the sensor array. In some embodiments, the sensor array comprises a plurality of sensing pixels and configured to generate a plurality of optical sensing signals. In some embodiments, each of the plurality of sensing pixels comprises a nearinfrared light-emitting diode (NIR LED) and a photodiode (PD). In some embodiments, the computing device is configured to generate at least one predicted tissue morph data object based on the plurality of optical sensing signals.

[0021]

[0016] In some embodiments, the at least one predicted tissue morph data object comprises a muscle activation factor (MAF).

[0022]

[0017] In some embodiments, the at least one predicted tissue morph data object comprises a predicted gesture classification. In some embodiments, the computing device is configured to generate the predicted gesture classification based on inputting the plurality of optical sensing signals to a trained machine learning model. In some embodiments, the trained machine learning model comprises an encoder network. In some embodiments, the trained machine learning model comprises a contrastive learning model. In some embodiments, the trained machine learning model comprises a classification model.

[0023]

[0018] In accordance with some embodiments of the present disclosure, a computer-implemented method is provided. In some embodiments, the computer-implemented method comprises: receiving, by at least one processor, a plurality of optical sensing signals from a sensor array, wherein the sensor array comprises a plurality of sensing pixels; and generating, by the at least one processor, at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals.

[0019] In accordance with some embodiments of the present disclosure, a computer program product is provided. In some embodiments, the computer program product comprises at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising an executable portion configured to receive, by at least one processor, a plurality of optical sensing signals from a plurality of sensing pixels of a sensor array; and generate, by the at least one processor, at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals.

[0024]

[0020] The foregoing illustrative summary, as well as other exemplary objectives and / or advantages of the disclosure, and the manner in which the same are accomplished, are further explained in the following detailed description and its accompanying drawings.

[0025] BRIEF DESCRIPTION OF DRAWINGS

[0026]

[0021] The description of the illustrative embodiments may be read in conjunction with the accompanying figures. It will be appreciated that, for simplicity and clarity of illustration, elements illustrated in the figures have not necessarily been drawn to scale, unless described otherwise. For example, the dimensions of some of the elements may be exaggerated relative to other elements, unless described otherwise. Embodiments incorporating teachings of the present disclosure are shown and described with respect to the figures presented herein.

[0027]

[0022] FIG. 1A provides a schematic illustration of a Laryngeal Health Monitor (LaHMo) patch deployed onto the laryngeal anterior area for muscular monitoring, symptom tracking, and recovery evaluation. The diagram on the right of FIG. 1A illustrates working flowline of the LaHMo on the laryngeal anterior area and the targeted applications. The LaHMo platform features onsite continuous muscle tracking over the laryngeal anterior area and wireless communication coupled with artificial intelligence (Al) boosted predicting analysis that classifies neck movement, swallow behavior, respiratory symptoms, and others, serving as diagnostic basis for clinical decision-making and precision treatment.

[0028]

[0023] FIG. IB provides an exploded view of an example LaHMo patch. In the example shown in FIG. IB, the example patch uses a serpentine hinge to enable mechanical foldability, facilitating both ergonomic fit and function. The hinge connects the main and daughter islands into a single flexible printed circuit board (fPCB), with the former running integrated circuits, including a microcontroller, an analog-to-digital converter, an operational amplifier, two of the four NIRS sensors, and an IMU, while the latter hosts the other two NIRS sensors.

[0029]

[0024] FIG. 1C provides an image of an example LaHMo patch deployed onto the neck.

[0030]

[0025] FIG. ID provides a block diagram showing the operational flow of an example LaHMo system. In some embodiments, a Bluetooth low energy (BLE) client (e g. a smartphone) receives wirelessly transmitted data from an example LaHMo patch and presents it in an accessible format for immediate review. In some embodiments, the data is further analyzed with a pre-trained RNN designed to detect and classify physiological events indicative of laryngeal health. This AL powered analysis can be conducted in embedded systems or cloud servers. Finally, the users utilize the analyzed data for various medical and health applications. The LaHMo platform could facilitate a bidirectional flow of dynamic and effective interactions between caregivers and receivers.

[0031]

[0026] FIG. IE illustrate example computer-implemented methods in accordance with some embodiments of the present disclosure.

[0032]

[0027] FIG. 2A provides an example schematic illustration indicating the corresponding sensing locations of the NIRS sensors in a LaHMo patch, highlighted with the central area 202, bottom area 204, top left area 206, and top right area 208, respectively.

[0033]

[0028] FIG. 2B provides a schematic illustration indicating dimensional parameters acquired from the embedded IMU, including the yaw, pitch, and roll Euler angles, respectively.

[0034]

[0029] FIG. 2C provides an image of an example LaHMo patch showing the corresponding position of the NIRS sensors.

[0035]

[0030] FIG. 2D provides a visual depiction of workflow in data preprocessing including data storage, analysis, and visualization. The process starts with the storage of raw data and signal peaks, followed by continuous processing and peak analysis of acquired data, and then completes the real-time visualization of processed data.

[0036]

[0031] FIG. 2E provides a representative preprocessed data from a repetitive test, on physiological events including deep breath, swallowing, dry cough, throat clearing, aerobic workout, and anaerobic workout. The preprocessed data feeds into Al models for further analysis.

[0037]

[0032] FIG. 3A provides a flow diagram of Al models highlighting three candidate architectures of RNN. Their distinguishment resides in utilizing mono- variants and dual -variants (represented by mono-RNN and dual-RNN respectively) of three different RNN architectures, including fully recurrent neural network (FRNN), long-short term memory (LSTM), and GRUs, respectively. For FRNN and GRU, there is only a hidden state (h_t) in every hidden layer cell; for LSTM, there is a hidden state (h t) and a cell state (c t) in each hidden layer cell. Four NN layers are demonstrated, with the dimension labeled underneath each diagram block. The preprocessed input and the output of a 1 -second sample are displayed at the start and end of the flow diagram.

[0038]

[0033] FIG. 3B provides loss curves comparison over the course of 400 epochs (performance of GRU) among training of the eight RNNs, including PV-biased RNN 301, PV-biased BiRNN 303, IMU-biased RNN 305, IMU-biased BiRNN 307, mono-RNN 309, mono-BiRNN 311, dual-RNN 313, and dual-RNN 319.

[0039]

[0034] FIG. 3C provides accuracy curves of training and testing over the course of 400 epochs for the dual-GRU and dual-BiGRU networks, including training dual-GRU network 321, testing dual-GRU network 323, training dual-BiGRU network 325, and testing dual-BiGRU network 327.

[0040]

[0035] FIG. 3D provides a confusion matrix of the dual-GRU and other networks in the preprocessed data.

[0041]

[0036] FIG. 3E provides a representative test on physiological events detection, involving continuous dry coughing. In particular, reference number 329 refers to the circles that indicate dry cough events in the test set. Reference number 331 refers to crosses that indicate dry cough in the training set. Reference number 333 refers to crosses that indicate predicted events with the dual-GRU model. Reference number 335 indicates signals from the central NIRS sensor, reference number 337 indicates signals from the bottom NIRS sensor, reference number 339 indicates signals from the top left NIRS sensor, and reference number 341 indicates signals from the top right NIRS sensor. Reference number 343 indicates yaw angles, reference number 345 indicates pitch angles, reference number 347 indicates roll angles.

[0042]

[0037] FIG. 3F provides standard deviations of the accuracy curve 349 of the dual-GRU network and the accuracy curve 351 of dual-BiGRU network for the training dataset over 400 epochs.

[0043]

[0038] FIG. 3G provides the p-values and t-statistics of the null hypothesis (i.e., the dual-GRU does not outperform the dual-BiGRU during the training process) over 400 epochs. Within the shaded area, as the training approaches the ES point, the p-values of the hypotheses for both the testing and training datasets drop to the rejection region (at a significance level of 0.05). As such, at the training ES point, statistical analysis suggests a rejection of the null hypothesis, due to the dual-GRU being more robust when encountering different inputs.

[0044]

[0039] FIG. 4A provides training-process visualization of the domain-adaptation network (adap-GRU). To enable Al-model individualization without undergoing a similar lengthy training process, a lightweight adaptation dual-GRU network (adap-GRU) is utilized, highlighted in the dashed box labeled “Adaptation”. This leads to a minimal training requirement from the target user, with only four datasets needing to be fed to the generalized, pre-trained dual GRU to stand up the adap-GRU. The main training process, with embedded data from a sample subject, establishes a dual-GRU network.

[0045]

[0040] FIG. 4B provides graphical representation of the four inputs required to train the adap-GRU. Photovoltage and Euler angle data over the course of the four physiological activities of interest are displayed, with gray lines representing raw data from tests and the dark lines presenting the averages.

[0046]

[0041] FIG. 4C provides the accuracy and loss curves of training and testing for the dual- and adap-GRU models. In particular, reference number 402 indicates the loss curve of training the adap-GRU model, reference number 404 indicates the loss curve of testing the adap-GRU model, reference number 406 indicates the loss curve of training the dual-GRU model, reference number 408 indicates the loss curve of testing the dual-GRU model, reference number 410 indicates the accuracy curve of training the adap-GRU model, reference number 412 indicates the accuracy curve of testing the adap-GRU model, reference number 414 indicates the accuracy curve of training the dual-GRU model, reference number 416 indicates the accuracy curve of testing the dual-GRU model. The results show that the adap-GRU network (96.1% accuracy at ES = 30 epochs) can be trained significantly faster than a brand-new dual-GRU network (65.4% accuracy at ES = 30 epochs).

[0047]

[0042] FIG. 4D provides the confusion matrix of the test data shows the algorithm prediction of the event type and the actual event type (1 -swallowing, 2-dry cough, 4-throat clearing, 5-aerobic workout).

[0048]

[0043] FIG. 5 A provides the representative central measurements (photovoltage) 501 and yaw measurements (angles) 503 during a 30-minute test during a basketball game. The markers show the manually labeled cough (circles) and cough-like (rectangles) events. The latter contains deep breaths, aerobic and anaerobic exercises during the long test. After data preprocessing, the finetuned adap-GRU model produces a real-time prediction based on the 3 s slices.

[0049] [441 FIG. 5B provides photographs (top row) and corresponding 3s slices of preprocessed data (bottom row, including central measurements (photovoltage) 505 and yaw measurements (angles) 507) showing the highlighted activities in the classification algorithm. From left to right: deep breath during rest, dry cough, anaerobic workout (shooting ball), and aerobic workout (dribbling and jogging). The black dashed boxes show the unique features that are considered during the manual label process.

[0050]

[0045] FIG. 5C provides example training curves during model fine-tuning (including the accuracy curve 509 and the loss curve 511 during model training, as well as the accuracy curve 513 and the loss curve 515 during model testing). The model reaches a loss of 0.34 / 0.36 for training / testing and an accuracy of 0.87 / 0.83 for training / testing at ES (epoch = 50).

[0051]

[0046] FIG. 5D provides the confusion matrix of the test data (0-deep breath, 1-dry cough, 2-anaerobic workout, 3 -aerobic workout).

[0052]

[0047] FIG. 5E provides the prediction score of classification for the physiological events occurred during the test period. Each temporal pixel represents a 3s data slice extracted from the real-time preprocessor (from top to bottom: deep breath, cough, anaerobic workout, and aerobic workout).

[0053]

[0048] FIG. 5F provides a calculated respiratory health indicator as a function of time. The datapoint color indicates the number of observed coughs in a 5-second window, that ranges from no cough to 5 coughs per time window.

[0054]

[0049] FIG. 5G provides a calculated exercise intensity indicator as a function of time. The datapoint color indicates the number of observed exercise events in a 5-second window, that ranges from no workout to intensive workouts.

[0055]

[0050] FIG. 6A provides a photographic depiction of the laryngeal prominence (LP), the targeted sensing area of the LaHMo patch. The anterior neck during various activities (deep breath, dry cough, throat clearing, and swallowing) yield measurements of its contour movements for comparison with simulation results. FIG. 6A further illustrates a cross-sectional model of the layered tissue at the sensing site, which consists of the epidermis, dermis, subcutaneous fat, and skeletal muscle (from outside inwards).

[0051] FIG. 6B provides a 3D contour plot showing the simulation results on the logarithm of the fluence corresponding to responses from the central photodiode of a LaHMo patch. The height of the LP (the distance between the LP’s peak and the surface of the throat) and the LP’s vertical displacement (the location of the LP’s peak relative to the midpoint of the throat) over the experiment. The logarithm of fluence is the calculated brightness of the tissue in the operational area. The LP height and the vertical displacement are shown on the bottom plane.

[0056]

[0052] FIG. 6C provides simulation results on the cross-sectional profiles of LED illumination from a LaHMo patch into the neck and their accompanying measured photovoltage and calculated fluence data during the deep breath and dry cough actions. To facilitate a practical visualization of the effects of the actions, there are three cross-section profiles per physiological action: at the start point, the point of the greatest change in measured, and the endpoint, respectively.

[0057]

[0053] FIG. 6D provides example placements of the electrode of four EMG channels and the reference.

[0058]

[0054] FIG. 6E provides acquired EMG signals at three different swallow events (highlighted with orange boxes) in a 10-second interval.

[0059]

[0055] FIG. 6F provides an acquired LaHMo data at the corresponding time interval. In particular, reference number 602 illustrates example central measurements (photovoltage), reference number 604 illustrates example bottom measurements (photovoltage), reference number 606 illustrates example top left measurements (photovoltage), and reference number 608 illustrates example top right measurements (photovoltage). Reference number 610 illustrates roll, reference number 612 indicates pitch, and reference number 614 indicates yaw.

[0060]

[0056] FIG. 7 illustrates three regions of an example LaHMo patch. A main PCB hinge, consisting of 10 serpentine traces, connects control portion to the sensing portion on the main island. Two more hinges connect the sub-islands to the main sensing island.

[0061]

[0057] FIG. 8A to FIG. 8G provide example demonstrations of the durability of the PCB of an example LaHMo patch against different deformations. In particular, FIG. 8A provides the original shape of the PCB. FIG. 8B provides a 45-degree twisting around the long axis of the patch. FIG.

[0062] 8C provides a 90-degree twisting around the short axis of the patch. FIG. 8D provides a 90-degree buckling around the hinge structure. FIG. 8E provides a 180-degree buckling around the hinge structure. FIG. 8F provides the same as FIG. 8E, shown in a different view. FIG. 8G provides a further incurve around the long axis after the 180-buckling around the hinge structure.

[0058] FIG. 9A to FIG. 9D provides example results from example power consumption tests. Each subplot shows both the power consumption at connected status and disconnected status. In particular, FIG. 9A to FIG. 9D highlight the measured values of a 1 -minute slice at the 5-hour timepoint.

[0063]

[0059] FIG. 9A illustrates voltages from a voltage supply at connected statue (reference number 901) and unconnected status (reference number 903).

[0064]

[0060] FIG. 9B illustrates current consumption at connected statue (reference number 905) and unconnected status (reference number 907).

[0065]

[0061] FIG. 9C illustrates port impedance at connected statue (reference number 909) and unconnected status (reference number 911).

[0066]

[0062] FIG. 9D illustrates power consumption at connected statue (reference number 913) and unconnected status (reference number 915).

[0067]

[0063] FIG. 9E illustrates accumulated current at connected statue (reference number 917) and unconnected status (reference number 919). Inset shows the accumulated current change in the first 8 hours. The calculated last time of the commonly used power source button battery (BB), lithium-ion battery (LI) and AAA battery are labelled in the inset.

[0068]

[0064] FIG. 9F illustrates accumulated power at connected statue (reference number 921) and unconnected status (reference number 923).

[0069]

[0065] FIG. 10 provides a comparison of LaHMo data between one dry cough event while standing still (top) and while walking (bottom). In FIG. 10, reference number 1002 and reference number 1004 indicate signals from the central NIRS sensor, reference number 1006 and reference number 1008 indicate signals from the bottom NIRS sensor, reference number 1010 and reference number 1012 indicate signals from the top left senor, reference number 1014 and reference number 1016 indicate signals from the top right NIRS sensor, reference number 1018 and reference number 1020 indicate roll angles, reference number 1022 and reference number 1024 indicate pitch angles, and reference number 1026 and reference number 1028 indicate yaw angles.

[0070]

[0066] FIG. 11 A to FIG. 1 IB provide a comparison between training and testing sets during the training of dual-FRNN, dual-LSTM, and dual-GRU.

[0071]

[0067] FIG. 11A illustrates example loss curves. In particular, reference number 1101 indicates loss curve of the training sets during the training of dual-FRNN. Reference number 1103 indicates loss curve of the testing sets during the training of dual-FRNN. Reference number 1105 indicates loss curve of the training sets during the training of dual-LSTM. Reference number 1107 indicates loss curve of the testing sets during the training of dual-LSTM. Reference number 1109 indicates loss curve of the training sets during the training of dual-GRU. Reference number 1111 indicates loss curve of the testing sets during the training of dual-GRU.

[0072]

[0068] FIG. 11B provides example accuracy curves. In particular, reference number 1113 indicates accuracy curve of the training sets during the training of dual-FRNN. Reference number 1115 indicates accuracy curve of the testing sets during the training of dual-FRNN. Reference number 1117 indicates accuracy curve of the training sets during the training of dual-LSTM. Reference number 1119 indicates accuracy curve of the testing sets during the training of dual-LSTM. Reference number 1121 indicates accuracy curve of the training sets during the training of dual-GRU. Reference number 1123 indicates accuracy curve of the testing sets during the training of dual-GRU.

[0073]

[0069] FIG. 12A to FIG. 12B provide example training curves for example FRNNs. In FIG.

[0074] 12A, reference number 1202 indicates loss curve for NIRS-biased FRNN, reference number 1204 indicates loss curve for NIRS-biased BiFRNN, reference number 1206 indicates loss curve for IMU-biased FRNN, reference number 1208 indicates loss curve for IMU-biased BiFRNN, reference number 1210 indicates loss curve for mono FRNN, reference number 1212 indicates loss curve for mono BiFRNN, reference number 1214 indicates loss curve for dual-FRNN, and reference number 1216 indicates loss curve for dual -BiFRNN. In FIG. 12B, reference number 1218 indicates accuracy curve for NIRS-biased FRNN, reference number 1220 indicates accuracy curve for NIRS-biased BiFRNN, reference number 1222 indicates accuracy curve for IMU-biased FRNN, reference number 1224 indicates accuracy curve for IMU-biased BiFRNN, reference number 1226 indicates accuracy curve for mono FRNN, reference number 1228 indicates accuracy curve for mono BiFRNN, reference number 1230 indicates accuracy curve for dual-FRNN, and reference number 1232 indicates accuracy curve for dual-BiFRNN.

[0075]

[0070] FIG. 12C to FIG. 12D provides example training curves for example LSTMs. In FIG.

[0076] 12C, reference number 1234 indicates loss curve for NIRS-biased LSTM, reference number 1236 indicates loss curve for NIRS-biased BiLSTM, reference number 1238 indicates loss curve for IMU-biased LSTM, reference number 1240 indicates loss curve for IMU-biased BiLSTM, reference number 1242 indicates loss curve for mono LSTM, reference number 1244 indicates loss curve for mono BiLSTM, reference number 1246 indicates loss curve for dual-LSTM, and reference number 1248 indicates loss curve for dual-BiLSTM. In FIG. 12D, reference number 1250 indicates accuracy curve for NIRS-biased LSTM, reference number 1252 indicates accuracy curve for NIRS-biased BiLSTM, reference number 1254 indicates accuracy curve for IMU-biased LSTM, reference number 1256 indicates accuracy curve for IMU-biased BiLSTM, reference number 1258 indicates accuracy curve for mono LSTM, reference number 1260 indicates accuracy curve for mono BiLSTM, reference number 1262 indicates accuracy curve for dual-LSTM, and reference number 1264 indicates accuracy curve for dual-BiLSTM.

[0077]

[0071] FIG. 12E to FIG. 12F provides example training curves for GRUs. In FIG. 12E, reference number 1266 indicates loss curve for PV-biased GRU, reference number 1268 indicates loss curve for PV-biased BiGRU, reference number 1270 indicates loss curve for IMU-biased GRU, reference number 1272 indicates loss curve for IMU-biased BiGRU, reference number 1274 indicates loss curve for mono GRU, reference number 1276 indicates loss curve for mono BiGRU, reference number 1278 indicates loss curve for dual-GRU, and reference number 1280 indicates loss curve for dual -BiGRU. In FIG. 12F, reference number 1282 indicates accuracy curve for NIRS-biased GRU, reference number 1284 indicates accuracy curve for NIRS-biased BiGRU, reference number 1286 indicates accuracy curve for IMU-biased GRU, reference number 1288 indicates accuracy curve for IMU-biased BiGRU, reference number 1290 indicates accuracy curve for mono GRU, reference number 1292 indicates accuracy curve for mono BiGRU, reference number 1294 indicates accuracy curve for dual-GRU, and reference number 1296 indicates accuracy curve for dual-BiGRU.

[0078]

[0072] FIG. 13 A to FIG. 13F provide a comparison of the training curves between RNN models and their bidirectional variants. In particular, FIG. 13A, FIG. 13C, and FIG. 13E provide accuracy curves for three dual-RNNs and dual-BiRNNs. FIG. 13B, FIG. 13D, and FIG. 13F provide accuracy curves for three mono-RNNs and mono-BiRNNs.

[0079]

[0073] In FIG. 13A, reference number 1301 indicates accuracy curve for training dual-FRNN, reference number 1303 indicates accuracy curve for testing dual-FRNN, reference number 1305 indicates accuracy curve for training dual-BiFRNN, and reference number 1307 indicates accuracy curve for testing dual-BiFRNN.

[0080]

[0074] In FIG. 13B, reference number 1309 indicates accuracy curve for training mono-FRNN, reference number 1311 indicates accuracy curve for testing mono-FRNN, reference number 1313 indicates accuracy curve for training mono-BiFRNN, and reference number 1315 indicates accuracy curve for testing mono-BiFRNN.

[0081]

[0075] In FIG. 13C, reference number 1317 indicates accuracy curve for training dual-LSTM, reference number 1319 indicates accuracy curve for testing dual-LSTM, reference number 1321 indicates accuracy curve for training dual-BiLSTM, and reference number 1323 indicates accuracy curve for testing dual-BiLSTM.

[0082]

[0076] In FIG. 13D, reference number 1325 indicates accuracy curve for training mono- LSTM, reference number 1327 indicates accuracy curve for testing mono-LSTM, reference number 1329 indicates accuracy curve for training mono-BiLSTM, and reference number 1331 indicates accuracy curve for testing mono-BiLSTM.

[0083]

[0077] In FIG. 13E, reference number 1333 indicates accuracy curve for training dual-GRU, reference number 1335 indicates accuracy curve for testing dual-GRU, reference number 1337 indicates accuracy curve for training dual-BiGRU, and reference number 1339 indicates accuracy curve for testing dual-BiGRU.

[0084]

[0078] In FIG. 13F, reference number 1341 indicates accuracy curve for training mono-GRU, reference number 1343 indicates accuracy curve for testing mono-GRU, reference number 1345 indicates accuracy curve for training mono-BiGRU, and reference number 1347 indicates accuracy curve for testing mono-BiGRU.

[0085]

[0079] FIG. 14A to FIG. 14H provide example confusion matrices for FRNN models on testing dataset.

[0086]

[0080] FIG. 14A provides an example confusion matrix for an example PV-biased FRNN.

[0087]

[0081] FIG. 14B provides an example confusion matrix for an example PV-biased BiFRNN.

[0088]

[0082] FIG. 14C provides an example confusion matrix for an example IMU-biased FRNN.

[0089]

[0083] FIG. 14D provides an example confusion matrix for an example IMU-biased BiFRNN.

[0090]

[0084] FIG. 14E provides an example confusion matrix for an example mono-FRNN.

[0091]

[0085] FIG. 14F provides an example confusion matrix for an example mono-BiFRNN.

[0092]

[0086] FIG. 14G provides an example confusion matrix for an example dual -FRNN.

[0093]

[0087] FIG. 14H provides an example confusion matrix for an example dual-BiFRNN.

[0094]

[0088] FIG. 15A to FIG. 15H provide example confusion matrices for LSTM models on testing dataset.

[0095]

[0089] FIG 15A provides an example confusion matrix for an example PV-biased LSTM.

[0090] FIG. 15B provides an example confusion matrix for an example PV-biased BiLSTM.

[0096]

[0091] FIG. 15C provides an example confusion matrix for an example IMU-biased LSTM.

[0097]

[0092] FIG. 15D provides an example confusion matrix for an example IMU-biased BiLSTM.

[0098]

[0093] FIG. 15E provides an example confusion matrix for an example mono-FRNN.

[0099]

[0094] FIG. 15F provides an example confusion matrix for an example mono-BiLSTM.

[0100]

[0095] FIG. 15G provides an example confusion matrix for an example dual-LSTM.

[0101]

[0096] FIG. 15H provides an example confusion matrix for an example dual -BiLSTM.

[0102]

[0097] FIG. 16A to FIG. 16H provide example confusion matrices for GRU models on testing dataset.

[0103]

[0098] FIG. 16A provides an example confusion matrix for an example PV-biased GRU.

[0104]

[0099] FIG. 16B provides an example confusion matrix for an example PV-biased BiGRU.

[0105]

[0100] FIG. 16C provides an example confusion matrix for an example IMU-biased GRU.

[0106]

[0101] FIG. 16D provides an example confusion matrix for an example IMU-biased BiGRU.

[0107]

[0102] FIG. 16E provides an example confusion matrix for an example mono-GRU.

[0108]

[0103] FIG. 16F provides an example confusion matrix for an example mono-BiGRU.

[0109]

[0104] FIG. 16G provides an example confusion matrix for an example dual-GRU.

[0110]

[0105] FIG. 16H provides an example confusion matrix for an example dual -BiGRU.

[0111]

[0106] FIG. 17A and FIG. 17B provide extended tests with a LaHMo on various activities. 16 types of physiological activities are classified into four groups to show their distinct pattern signatures in the data collected from the wearable patch.

[0112]

[0107] FIG. 18 illustrates a continuous dry cough test and photodiode responses of LaHMo to the continuous dry coughs, with minimal time and no recovery between coughs.

[0113]

[0108] FIG. 19 illustrates a random dry cough test and photodiode responses of LaHMo to random dry coughs, with varying amounts of time and recovery between coughs.

[0114]

[0109] FIG. 20 illustrates a separated dry cough test and photodiode responses of LaHMo to separate dry coughs, with enough time to allow for complete recovery between coughs.

[0115] [HO] FIG. 21A and FIG. 21B illustrates a comparison of the photodiode response of LaHMo to various swallow qualities. FIG. 21 A illustrates a response to different chewing times (same food sample) (from top to bottom: 15 chews, 12 chews, 9 chews). In FIG. 21 A, reference number 2101, reference number 2109, and reference number 2117 indicate signals from the central NIRS sensor. Reference number 2103, reference number 2111, and reference number 2119 indicate signals from the bottom NIRS sensor. Reference number 2105, reference number 2113, and reference number 2121 indicate signals from the top left senor. Reference number 2107, reference number 2115, and reference number 2123 indicate signals from the top right NIRS sensor.

[0116] [Hl] FIG. 21B illustrates an enlarged view of FIG. 21A at the swallowing. In FIG. 21B, reference number 2125, reference number 2133, and reference number 2141 indicate signals from the central NIRS sensor. Reference number 2127, reference number 2135, and reference number 2143 indicate signals from the bottom NIRS sensor. Reference number 2129, reference number 2137, and reference number 2145 indicate signals from the top left senor. Reference number 2131, reference number 2139, and reference number 2147 indicate signals from the top right NIRS sensor.

[0117]

[0112] FIG. 21C provide a continuous swallowing (top: sufficient swallow, bottom: insufficient swallow (more muscle activities)). In FIG. 21C, reference number 2149 and reference number 2151 indicate signals from the central NIRS sensor. Reference number 2153 and reference number 2155 indicate signals from the bottom NIRS sensor. Reference number 2157 and reference number 2159 indicate signals from the top left senor. Reference number 2161 and reference number 2163 indicate signals from the top right NIRS sensor.

[0118]

[0113] FIG. 22 provides a photodiode response of LaHMO to swallowing vegetable juice.

[0119]

[0114] FIG. 23 provides a photodiode response of LaHMO to swallowing water.

[0120]

[0115] FIG. 24A to FIG. 24E provide example photovoltage responses of LaHMo to various vowel phonemes.

[0121]

[0116] FIG. 24A illustrates example photovoltage responses of a close front flat / i / _

[0122]

[0117] FIG. 24B illustrates example photovoltage responses of a close back flat / ui / .

[0123]

[0118] FIG. 24C illustrates example photovoltage responses of an open back flat lai.

[0124]

[0119] FIG. 24D illustrates example photovoltage responses of a close-mid central flat / a / .

[0125]

[0120] FIG. 24E illustrates a vowel location reference. The horizontal axis shows the tongue peak anterior location, the vertical axis shows the mouth open content.

[0126]

[0121] FIG. 25 provides a photovoltage response of LaHMo to pronunciation of close back flat vowel phenome.

[0127]

[0122] FIG. 26 provides a photovoltage response of LaHMo to pronunciation of close central flat vowel phenome.

[0123] FIG. 27 provides a photovoltage response of LaHMo to pronunciation of close front flat vowel phenome.

[0128]

[0124] FIG. 28 provides a photovoltage response of LaHMo to pronunciation of close-mid back flat vowel phenome.

[0129]

[0125] FIG. 29 provides a photovoltage response of LaHMo to pronunciation of close-mid central flat vowel phenome.

[0130]

[0126] FIG. 30 provides a photovoltage response of LaHMo to pronunciation of close-mid front flat vowel phenome.

[0131]

[0127] FIG. 31 provides a photovoltage response of LaHMo to pronunciation of open-mid back flat vowel phenome.

[0132]

[0128] FIG. 32 provides a photovoltage response of LaHMo to pronunciation of open-mid central flat vowel phenome.

[0133]

[0129] FIG. 33 provides a photovoltage response of LaHMo to pronunciation of open-mid front flat vowel phenome.

[0134]

[0130] FIG. 34 provides a photovoltage response of LaHMo to pronunciation of open back flat vowel phenome.

[0135]

[0131] FIG. 35 provides a photovoltage response of LaHMo to pronunciation of open front flat vowel phenome.

[0136]

[0132] FIG. 36A provides the LaHMo data of the subject producing pitch G4. In FIG. 36A, reference number 3602 indicates signals from the central NIRS sensor, reference number 3604 indicates signals from the bottom NIRS sensor, reference number 3606 indicates signals from the top left senor, reference number 3608 indicates signals from the top right NIRS sensor, reference number 3610 indicates roll angles, reference number 3612 indicates pitch angles, and reference number 3614 indicates yaw angles.

[0137]

[0133] FIG. 36B provides the LaHMo data of the subject producing pitch G3. In FIG. 36B, reference number 3616 indicates signals from the central NIRS sensor, reference number 3618 indicates signals from the bottom NIRS sensor, reference number 3620 indicates signals from the top left senor, reference number 3622 indicates signals from the top right NIRS sensor, reference number 3624 indicates roll angles, reference number 3626 indicates pitch angles, and reference number 3628 indicates yaw angles.

[0134] FIG. 36C provides the LaHMo data of the subject producing pitch C3. In FIG. 36C, reference number 3630 indicates signals from the central NIRS sensor, reference number 3632 indicates signals from the bottom NIRS sensor, reference number 3634 indicates signals from the top left senor, reference number 3636 indicates signals from the top right NIRS sensor, reference number 3638 indicates roll angles, reference number 3640 indicates pitch angles, and reference number 3642 indicates yaw angles.

[0138]

[0135] FIG. 36D provides the corresponding audio segment of FIG. 36A.

[0139]

[0136] FIG 36E provides the corresponding audio segment of FIG 36B.

[0140]

[0137] FIG. 36F provides the corresponding audio segment of FIG. 36C.

[0141]

[0138] FIG. 37A provides the LaHMo data of the subject producing a head voice at pitch G4. In FIG. 37A, reference number 3701 indicates signals from the central NIRS sensor, reference number 3703 indicates signals from the bottom NIRS sensor, reference number 3705 indicates signals from the top left senor, reference number 3707 indicates signals from the top right NIRS sensor, reference number 3709 indicates roll angles, reference number 3711 indicates pitch angles, and reference number 3713 indicates yaw angles.

[0142]

[0139] FIG. 37B provides the LaHMo data of the subject produced a chest voice at pitch G4. In FIG. 37B, reference number 3715 indicates signals from the central NIRS sensor, reference number 3717 indicates signals from the bottom NIRS sensor, reference number 3719 indicates signals from the top left senor, reference number 3721 indicates signals from the top right NIRS sensor, reference number 3723 indicates roll angles, reference number 3725 indicates pitch angles, and reference number 3727 indicates yaw angles.

[0143]

[0140] FIG. 37C provides the corresponding audio segment of FIG. 37A.

[0144]

[0141] FIG. 37D provides the corresponding audio segment of FIG. 37B.

[0145]

[0142] FIG. 37E provides the LaHMo data of the subject whispers at pitch C3. In FIG. 37E, reference number 3729 indicates signals from the central NIRS sensor, reference number 3731 indicates signals from the bottom NIRS sensor, reference number 3733 indicates signals from the top left senor, reference number 3735 indicates signals from the top right NIRS sensor, reference number 3737 indicates roll angles, reference number 3739 indicates pitch angles, and reference number 3741 indicates yaw angles.

[0146]

[0143] FIG. 37F provides the LaHMo data of the subject produces a loud voice at pitch C3. In FIG. 37F, reference number 3743 indicates signals from the central NIRS sensor, reference number 3745 indicates signals from the bottom NIRS sensor, reference number 3747 indicates signals from the top left senor, reference number 3749 indicates signals from the top right NIRS sensor, reference number 3751 indicates roll angles, reference number 3753 indicates pitch angles, and reference number 3755 indicates yaw angles.

[0147]

[0144] FIG. 37G provides the corresponding audio segment of FIG. 37E.

[0148]

[0145] FIG. 37H provides the corresponding audio segment of FIG. 37F.

[0149]

[0146] FIG. 38A to FIG. 38E provides example comparison of the photodiode response of LaHMo to different head motions.

[0150]

[0147] FIG. 38A provides photodiode response of LaHMo to a cervical extension in 30 degrees. In FIG. 38A, reference number 3802 indicates signals from the central NIRS sensor, reference number 3804 indicates signals from the bottom NIRS sensor, reference number 3806 indicates signals from the top left NIRS sensor, and reference number 3808 indicates signals from the top right NIRS sensor. Reference number 3810 indicates ground truth values collected by IMU.

[0151]

[0148] FIG. 38B provides photodiode response of LaHMo to a cervical extension in 60 degrees. In FIG. 38B, reference number 3812 indicates signals from the central NIRS sensor, reference number 3814 indicates signals from the bottom NIRS sensor, reference number 3816 indicates signals from the top left NIRS sensor, and reference number 3818 indicates signals from the top right NIRS sensor. Reference number 3820 indicates ground truth values collected by IMU.

[0152]

[0149] FIG. 38C provides photodiode response of LaHMo to a cervical left bending in 15 degrees. In FIG. 38C, reference number 3822 indicates signals from the central NIRS sensor, reference number 3824 indicates signals from the bottom NIRS sensor, reference number 3826 indicates signals from the top left NIRS sensor, and reference number 3828 indicates signals from the top right NIRS sensor. Reference number 3830 indicates ground truth values collected by IMU.

[0153]

[0150] FIG. 38D provides photodiode response of LaHMo to a cervical left bending in 30 degrees. In FIG. 38D, reference number 3832 indicates signals from the central NIRS sensor, reference number 3834 indicates signals from the bottom NIRS sensor, reference number 3836 indicates signals from the top left NIRS sensor, and reference number 3838 indicates signals from the top right NIRS sensor. Reference number 3840 indicates ground truth values collected by IMU.

[0154]

[0151] FIG. 38E provides a continuous response to left and right bending, 30 degrees. In FIG.

[0155] 38E, reference number 3842 and reference number 3844 indicate signals from the central NIRS sensor, reference number 3846 and reference number 3848 indicate signals from the bottom NIRS sensor, reference number 3850 and reference number 3852 indicate signals from the top left NIRS sensor, and reference number 3854 and reference number 3856 indicate signals from the top right NIRS sensor.

[0156]

[0152] FIG. 39A to FIG. 39C provide training and testing standard variation during model training process when doing a k-fold cross validation.

[0157]

[0153] FIG. 39A provides training and testing standard variations during model training process for a mono-FRNN and dual-FRNN. In particular, reference number 3901 indicates training standard variations associated with a mono-FRNN, reference number 3903 indicates testing standard variations associated with a mono-FRNN, reference number 3905 indicates training standard variations associated with a mono-BiFRNN, reference number 3907 indicates testing standard variations associated with a mono-BiFRNN, reference number 3909 indicates training standard variations associated with a dual-FRNN, reference number 3911 indicates testing standard variations associated with a dual-FRNN, reference number 3913 indicates training standard variations associated with a dual-BiFRNN, and reference number 3915 indicates testing standard variations associated with a dual-BiFRNN.

[0158]

[0154] FIG. 39B provides training and testing standard variations during model training process for a mono-LSTM and dual-LSTM. In particular, reference number 3917 indicates training standard variations associated with a mono-LSTM, reference number 3919 indicates testing standard variations associated with a mono-LSTM, reference number 3921 indicates training standard variations associated with a mono-BiLSTM, reference number 3923 indicates testing standard variations associated with a mono-BiLSTM, reference number 3925 indicates training standard variations associated with a dual-LSTM, reference number 3927 indicates testing standard variations associated with a dual-LSTM, reference number 3929 indicates training standard variations associated with a dual-BiLSTM, and reference number 3931 indicates testing standard variations associated with a dual-BiLSTM.

[0159]

[0155] FIG. 39C provides training and testing standard variations during model training process for a mono-GRU and dual-GRU. In particular, reference number 3933 indicates training standard variations associated with a mono-GRU, reference number 3935 indicates testing standard variations associated with a mono-GRU, reference number 3937 indicates training standard variations associated with a mono-BiGRU, reference number 3939 indicates testing standard variations associated with a mono-BiGRU, reference number 3941 indicates training standard variations associated with a dual-GRU, reference number 3943 indicates testing standard variations associated with a dual-GRU, reference number 3945 indicates training standard variations associated with a dual-BiGRU, and reference number 3947 indicates testing standard variations associated with a dual-BiGRU.

[0160]

[0156] FIG. 40 provides a prediction of adap-GRU on target user.

[0161]

[0157] FIG. 41 A to FIG. 41D provide a basic Monte Carlo simulation of the patch applied on a tissue surface.

[0162]

[0158] FIG. 41A provides a model setup. In particular, reference number 4101 indicates epidermis, reference number 4103 indicates dermis, reference number 4105 indicates subcutaneous fatty tissue, and reference number 4107 indicates muscle.

[0163]

[0159] FIG. 4 IB provides a vertical cross-section fluence profile at the light source. In particular, reference number 4109 points to the location and direction of the light source. Reference number 4111 points to the location and size of the photodetector at coordinate (0,0).

[0164]

[0160] FIG. 41C provides a highlighted spatial elements with fluence 1 / 100 of the light source.

[0165]

[0161] FIG. 4 ID provides a horizontal cross-section fluence profile at the interface of epidermis and background.

[0166]

[0162] FIG. 42A to FIG. 42D provide a basic Monte Carlo simulation of the patch applied on a laryngeal prominence.

[0167]

[0163] FIG. 42A provides a model setup. In particular, reference number 4202 indicates epidermis, reference number 4204 indicates dermis, reference number 4206 indicates subcutaneous fatty tissue, and reference number 4208 indicates muscle.

[0168]

[0164] FIG. 42B provides a vertical cross-section fluence profile at the light source. In particular, reference number 4210 points to the location and direction of the light source. Reference number 4212 points to the location and size of the photodetector at coordinate (0,0).

[0169]

[0165] FIG. 42C provides a highlighted spatial elements with fluence 1 / 100 of the light source.

[0170]

[0166] FIG. 42D provides a horizontal cross-section fluence profile at the interface of epidermis and background.

[0171]

[0167] FIG. 43 provides a simulation results on the cross-sectional profiles of LED illumination from a LaHMo patch into the neck and their accompanying measured photovoltage and calculated fluence data during the throat clearing and swallowing events.

[0168] FIG. 44A to FIG. 44D provide a cross-correlation between MC simulation results and NIRS measurements.

[0172]

[0169] FIG. 44A provides cross-correlation between MC simulation results and NIRS measurements for a deep breath. In particular, reference number 4402 indicates signals from the central NIRS sensor, and reference number 4404 indicates signals from the bottom NIRS sensor.

[0173]

[0170] FIG. 44B provides cross-correlation between MC simulation results and NIRS measurements for a dry cough. In particular, reference number 4406 indicates signals from the central NIRS sensor, and reference number 4408 indicates signals from the bottom NIRS sensor.

[0174]

[0171] FIG. 44C provides cross-correlation between MC simulation results and NIRS measurements for a throat clearing. In particular, reference number 4410 indicates signals from the central NIRS sensor, and reference number 4412 indicates signals from the bottom NIRS sensor.

[0175]

[0172] FIG. 44D provides cross-correlation between MC simulation results and NIRS measurements for swallowing. In particular, reference number 4414 indicates signals from the central NIRS sensor, and reference number 4416 indicates signals from the bottom NIRS sensor.

[0176]

[0173] FIG. 45A to FIG. 45B provide the comparison between a two-element Gaussian distribution (as shown in FIG. 45A) and a lambertian distribution (as shown in FIG. 45B) in the range from -2 to 2. The Gaussian distribution uses a standard deviation of 1 for both x and y variables.

[0177]

[0174] FIG. 46A provides LaHMo data of the subject producing a cycle of natural musical scale from C3 to C4, and back to C3, using chest voice. In FIG. 46A, reference number 4602 indicates signals from the central NIRS sensor, reference number 4604 indicates signals from the bottom NIRS sensor, reference number 4606 indicates signals from the top left senor, reference number 4608 indicates signals from the top right NIRS sensor, reference number 4610 indicates roll angles, reference number 4612 indicates pitch angles, and reference number 4614 indicates yaw angles.

[0178]

[0175] FIG. 46B provides the corresponding audio recording of the test segment presented in FIG. 46A.

[0179]

[0176] FIG. 46C provides LaHMo data of the subject producing a cycle of natural musical scale from C4 to C5, and back to C4, using falsetto. In FIG. 46C, reference number 4616 indicates signals from the central NIRS sensor, reference number 4618 indicates signals from the bottom NIRS sensor, reference number 4620 indicates signals from the top left senor, reference number 4622 indicates signals from the top right NIRS sensor, reference number 4624 indicates roll angles, reference number 4626 indicates pitch angles, and reference number 4628 indicates yaw angles.

[0180]

[0177] FIG. 46D provides the corresponding audio recording of the test segment presented in FIG. 46C. The green and aqua dashed lines represent the C3 and C4 note samples that correspond to the videostroboscopy slices in FIG. 46E to FIG. 46H).

[0181]

[0178] FIG. 46E provides a videostroboscopy slice at chest voice C3.

[0182]

[0179] FIG. 46F provides a videostroboscopy slice at chest voice C4.

[0183]

[0180] FIG. 46G provides a videostroboscopy slice at falsetto voice C4.

[0184]

[0181] FIG. 46H provides a videostroboscopy slice at falsetto voice C5. Dashed circles in FIG.

[0185] 46E to FIG. 46H highlight the corniculate cartilages, thin arrows label the central line of the crosssection of the corniculate cartilages, pointing at anterior. Thick arrows highlight the moving trend of the corniculate cartilage in order for the subject to produce the corresponding note.

[0186]

[0182] FIG. 47 provides an example graphic interface of the BTViz. The top left panel shows the device scanning widget, used to connect to the specified device. The top right panel shows the service / characteristic selection widget. The bottom panel shows the characteristic reader that decodes the received BLE data and implements the real-time visualization.

[0187]

[0183] FIG. 48A and FIG. 48B provide LaHMo data of solid food bolus with an increasing viscosity from the top of FIG. 48A to the bottom of FIG. 48B. The left column of each of FIG.

[0188] 48A and FIG. 48B shows the example of the acquired signal at different viscosity levels, and the right column of each of FIG. 48A and FIG. 48B shows a food example for each viscosity level.

[0189]

[0184] FIG. 49A provides the LaHMo data segment presenting a nasal breath attempt. In FIG.

[0190] 49 A, reference number 4901 indicates signals from the central NIRS sensor, reference number 4903 indicates signals from the bottom NIRS sensor, reference number 4905 indicates signals from the top left senor, reference number 4907 indicates signals from the top right NIRS sensor, reference number 4909 indicates roll angles, reference number 4911 indicates pitch angles, and reference number 4913 indicates yaw angles.

[0191]

[0185] FIG. 49B provides the LaHMo data segment presenting an oral breath attempt. In FIG.

[0192] 49B, reference number 4915 indicates signals from the central NIRS sensor, reference number 4917 indicates signals from the bottom NIRS sensor, reference number 4919 indicates signals from the top left senor, reference number 4921 indicates signals from the top right NIRS sensor, reference number 4923 indicates roll angles, reference number 4925 indicates pitch angles, and reference number 4927 indicates yaw angles.

[0193]

[0186] FIG. 50A to FIG. 50G provide an overview of an example wearable muscle monitor (also referred to as “TRAiLL” or “TRAiLL system” and the like) in accordance with some embodiments of the present disclosure, including device, deployment, applications, and electronics.

[0194]

[0187] FIG. 50A provides a photorealistic rendering of the example sensor array (also referred to as “TRAiLL patch” or “patch” and the like) on the volar wrist. FIG. 50A further includes an exploded view, highlighting surface-mount NIR LEDs delivered light into tissue and co-planar silicon photodiodes collected multiply scattered photons (shown as traces 5002), within a flexible printed circuit board.

[0195]

[0188] FIG. 50B to FIG. 50D provide a representative placement sites for the TRAiLL patch and the corresponding signals.

[0196]

[0189] FIG. 50C illustrates, when placed on the limb, TRAiLL quantified muscle activation and fatigue and showed raw recordings and the derived muscle-activation factor (MAF) in a display of a computing device.

[0197]

[0190] FIG. 50D illustrates, when TRAiLL patch is positioned on the wrist, characteristic spatiotemporal patterns enable gesture recognition in accordance with some embodiments of the present disclosure.

[0198]

[0191] FIG. 50E provides schematics summarizing three human-machine-interface applications enabled by TRAiLL signals: (i) rehabilitation monitoring via time-lapse fatigue mapping during resisted elbow flexion; (ii) gesture interpretation by decoding wrist motion signatures with a deep-learning classifier; and (iii) camera-free virtual -reality interaction using real-time muscle-activity maps.

[0199]

[0192] FIG. 50F provides a block diagram of an example hardware and processing pipeline of an example wearable muscle monitor in accordance with some embodiments of the present disclosure. In some embodiments, photocurrents from photodiodes were converted by a multichannel trans-impedance amplifier, low-pass filtered, and multiplexed to an ADC on the adaptor board. In some embodiments, LED intensity was controlled by PWM through P-MOSFET drivers under microcontroller (MCU) control. In some embodiments, the MCU received digitized data via I²C, issued 4-byte configuration commands through a level shifter, and streamed time- stamped frames to a host computer for data acquisition, neural -network inference of muscle state, and closed-loop feedback.

[0200]

[0193] FIG. 50G illustrates an example photograph of a TRAiLL patch worn on the forearm, demonstrating conformal adhesion and robustness during bending.

[0201]

[0194] FIG. 51A to FIG. 5 IK illustrate example finger-specific gesture decoding from spatially resolved forearm optical myography.

[0202]

[0195] FIG. 51A illustrates a concept of near-infrared illumination and back-scatter collection with an LED / photodiode pair, alongside Monte-Carlo steady-state fluence and a schematic of flexor morphology from open hand to fist.

[0203]

[0196] FIG. 51B illustrates a schematic illustration on the cross-section of a left volar forearm showing the flexor digitorum profundus (FDP), flexor digitorum sublimis (FDS), Flexor pollicis longus (FPL), and neighboring muscles.

[0204]

[0197] FIG. 51C, FIG. 51D, FIG. 51E, and FIG. 51F illustrate example simultaneous B-mode ultrasound and TRAiLL during isometric flexion of single digits. For example, FIG. 51C, FIG.

[0205] 5 ID, FIG. 5 IE, and FIG. 5 IF illustrates example spatiotemporal tissue morph maps (also referred to as “photovoltage maps” or “2-D photovoltage maps” as the like) (such as map 5103, map 5107, map 5123, map 5129, and map 5135), stacks of consecutive optical frames (10 ms per slice) (such as stack 5105, stack 5111, stack 5127, stack 5133, and stack 5139) as well as ultrasound images (such as image 5115, image 5119, image 5143, image 5151, and image 5155).

[0206]

[0198] FIG. 51C, FIG. 51D, FIG. 51E, and FIG. 51F further show focal hemodynamic changes co-localized with tendon-muscle junctions (such as area 5101, area 5109, area 5125, area 5131, and area 5137) as well as FDP (such as area 5121, area 5141, area 5149, and area 5157) and FDS (such as area 5113, area 5117, area 5145, area 5147, and area 5153) with arrows showing motion trends.

[0207]

[0199] FIG. 51E and FIG. 51F show minimal signal when the subject is resting (muscles are annotated with a scale bars of 10 mm).

[0208]

[0200] FIG. 51G provides representative optical sensing signals (also referred to as “traces,” “channel traces,” “photovoltage traces” and the like) among five channels (for example, line 5159 for channel (7, 3), line 5161 for channel (7, 2), line 5163 for channel (6, 4), line 5165 for channel (4, 3), and line 5167 for channel (6, 3)) based on the locations shown in FIG. 51H (for example, location 5175 for channel (7, 3), location 5169 for channel (7, 2), location 5177 for channel (6, 4), location 5171 for channel (4, 3), and location 5173 for location (6, 3)), and the channel traces exhibited distinct temporal signatures for each finger.

[0209]

[0201] FIG. 51H provides spatial positions for the five channels on the TRAiLL patch.

[0210]

[0202] FIG. 51I shows example effort dependence, including photographs (on the left) at 0%, 50% and 100% maximal voluntary contraction (MVC) and responses (on the right) from the maximally responsive channel (rounded dots) overlaid on the array mean (grey area) ± s.d. with arrowheads that mark onset (left arrowhead), mid-effort (middle arrowhead) and peak (right arrowhead).

[0211]

[0203] FIG. 51J illustrates normalized population-averaged trajectories (e.g., mean ± s.e.m., n = 5 contractions per finger) across the flexion-extension phase; dashed planes denote rest.

[0212]

[0204] FIG. 51K illustrates a linear-discriminant analysis of 3,000 single-frame features (including 200 frames x 5 fingers * 3 subjects) that produced well-separated clusters (shown as ellipses with 95 % Mahalanobis confidence).

[0213]

[0205] FIG. 52A to FIG. 52J illustrate load and grip style modulated activation and fatigue of antagonistic arm muscles measured with TRAiLL.

[0214]

[0206] FIG. 52A illustrates a photograph of elbow flexion (“curl,” biceps brachii) with arrows indicate motion. A TRAiLL patch was centered on the anterior mid-arm.

[0215]

[0207] FIG. 52B illustrates muscle-activation factor (MAF) during five unloaded and 5-lb loaded curls, interleaved with related state. FIG. 52B further provides heat maps showing channelwise MAF on anterior (top) and posterior (bottom) rows.

[0216]

[0208] FIG. 52C illustrates mean ± s.e.m. MAF during a 60-s isometric biceps hold with loaded flexion decayed to -55% of peak, whereas unloaded effort showed minimal decline. FIG.

[0217] 52C further highlights the first 15 s with arrows showing contraction onset.

[0218]

[0209] FIG. 52D illustrates a photograph of elbow extension (“kick-back,” triceps brachii) with a patch centered on the posterior mid-arm.

[0219]

[0210] FIG. 52E illustrates, as in FIG. 52B, for five unloaded and 5-lb loaded kickbacks with responses predominated on posterior rows.

[0220]

[0211] FIG. 52F illustrates mean ± s.e.m. MAF during a 60-s isometric triceps hold with loaded extension decayed to -40% of peak, whereas unloaded showed little decline (as highlighted in FIG. 52C above).

[0212] FIG. 52G illustrates a schematic of the seated concentration-curl paradigm for grip testing with the patch over biceps).

[0221]

[0213] FIG. 52H illustrates MAF traces and channel-wise maps for five short curls and a static hold at maximal torque using thumb-over grip (TOG) versus thumb-under grip (TUG), where TUG elicited larger and more widespread activation.

[0222]

[0214] FIG. 521 illustrates representative photographs of TOG (top) and TUG (bottom).

[0223]

[0215] FIG. 52 J illustrates a three-day continuous monitoring of biceps MAF. In FIG. 52J, two timelines (top / middle) mark Day 1 and Days 2-3, with markers indicate the seven 30-s excerpts plotted below. Shading on the diagrams denotes a fatigue threshold above which cumulative fatigue was expected to build.

[0224]

[0216] FIG 53A to FIG. 53L illustrate three-phase self-supervised pipeline improves ASL letter classification from TRAiLL optical myography.

[0225]

[0217] FIG. 53A illustrates phase I denoising auto-encoder where noisy 3-D time series data are compressed by an encoder and reconstructed by a decoder for further training.

[0226]

[0218] FIG. 53B illustrates phase II contrastive pre-text learning where translated, rotated and mirrored augmentations are passed through the encoder and sent to a contrastive SSL to maximize the difference between each ASL-A label in the latent space.

[0227]

[0219] FIG. 53C illustrates phase III gesture-classification downstream task where time-series latent are fed to a multilayer perception that assigned one of 24 static ASL-A gestures.

[0228]

[0220] FIG. 53D to FIG. 53E illustrate representative input (as shown in FIG. 53D) and reconstruction (as shown in FIG. 53E) heat maps, demonstrating high structural fidelity while removing the high-frequency noises.

[0229]

[0221] FIG. 53F illustrates a learning curve of Phase I training. The accuracy is defined by the proportion of reconstructed samples with less than 1% of error from original signal.

[0230]

[0222] FIG. 53G illustrates Fourier spectra of original and reconstructed signals overlapped across 0-20 Hz, indicating preservation of low-frequency physiologically relevant content while high-frequency noises are suppressed.

[0231]

[0223] FIG. 53H illustrates the t-SNE result of 6500 augmented embeddings encoded by Phase II training. The data points signify augmentation invariance.

[0232]

[0224] FIG. 53I illustrates the hidden weights of each input channel of the encoder.

[0225] FIG. 53J illustrates a cross-validation accuracy (for example, mean ± std, five train-test split folds) rise from 69 % for a naive MLP to 92 % when Phase I + II weights were combined with the classifier.

[0233]

[0226] FIG. 53K illustrate receiver-operating-characteristic curves for all 24 classes yielded area-under-the-curve more than 0.98 with a corresponding class index.

[0234]

[0227] FIG. 53L illustrate a normalized confusion matrix exhibiting strong diagonal dominance.

[0235]

[0228] FIG. 54A to FIG. 54I illustrates a TRAiLL optical myography enabled bimanual music performance and camera-free 3D gesture control.

[0236]

[0229] FIG. 54A illustrates cartoon contrasts conventional guitar playing with a virtual-instrument setup driven solely by TRAiLL patches.

[0237]

[0230] FIG. 54B illustrates a photograph of a volunteer wearing two TRAiLL patches with one on the left forearm to sense chord fingering, and the other on the right wrist to detect stringstrumming motions. Meanwhile, the volunteer was interfacing with a laptop DAQ and synthesizer.

[0238]

[0231] FIG. 54C and FIG. 54D illustrate a representative raw photovoltage traces (with 48 channels that are coded by amplitude) recorded during a four-chord progression (C-F-G-C, as shown in FIG. 54C) and a repetitive right-hand plucking pattern (“t-i-m-r”, as shown in FIG. 54D).

[0239]

[0232] FIG. 54E illustrates an illustration of single-patch control of an augmented-reality object via six hand gestures.

[0240]

[0233] FIG. 54F illustrates a channel-wise muscle-activation heat-map acquired during a 50 s session that shows discrete bursts corresponding to individual gestures.

[0241]

[0234] FIG. 54G illustrates a time-stamped classification output from a pretrained model accurately labelled grab, pinch, wave, trigger, fist and thumbs-up events.

[0242]

[0235] FIG. 54H illustrates training curves for the 6-class model (mean cross-entropy loss, left axis) reached convergence within 30 epochs, yielding 96 % validation accuracy (right axis).

[0243]

[0236] FIG. 54I illustrates a normalized confusion matrix for an independent test set (n = 720 gestures) exhibited near-perfect diagonal dominance.

[0244]

[0237] FIG. 55A to FIG. 55I illustrate a mechanical conformability of the TRAiLL epidermal optical array. Photographs shown in FIG. 55A to FIG. 55I documented the flexibility and robustness of the assembled TRAiLL patch during manual deformation, demonstrating compatibility with curved anatomical surfaces.

[0238] FIG. 55A illustrates a top view of the fabricated device showing the serpentine interconnect layout and peripheral contacts.

[0245]

[0239] FIG. 55B to FIG. 55C illustrate uniaxial bending into concave (as shown in FIG. 55B) and convex (as shown in FIG. 55C) curvatures without visible cracking or delamination.

[0246]

[0240] FIG. 55D to FIG. 55E illustrate the same bending modes applied about the long axis, concave (as shown in FIG. 55D) and convex (as shown in FIG. 55E).

[0247]

[0241] FIG. 55F to FIG. 55G illustrate complex draping across multiple radii showed stable placement of surface-mount emitters and photodiodes.

[0248]

[0242] FIG. 55H illustrates a manual uniaxial stretching of the patch where the serpentine traces accommodated tensile strain without loss of integrity.

[0249]

[0243] FIG. 55I illustrates tight rolling into a small-radius cylinder that further confirms high bendability. Across all deformations, the laminate remained intact, and components stayed planar, indicating that the patch tolerates large strains required for skin-conformal wear and repeated handling.

[0250]

[0244] FIG. 56A to FIG. 56H illustrate the wearability of the TRAiLL patch during everyday motions with two mounting orientations. Photographs shown in FIG. 56A to FIG. 56H document that the TRAiLL patch remained conformal and did not impede movement when worn on the wrist or upper arm, and that it could be deployed with the sensing layer oriented either inward (toward the skin) or outward (away from the skin) to optimize adhesion and comfort.

[0251]

[0245] FIG. 56A to FIG. 56B illustrate wrist-worn with the sensor side facing inward during hand opening and an “OK” gesture, showing full range of motion without edge lifting.

[0252]

[0246] FIG. 56C to FIG. 56D illustrate the upper-arm placement with the sensor side facing outward during elbow extension and flexion, demonstrating unobstructed joint motion and secure attachment.

[0253]

[0247] FIG. 56E to FIG. 56F illustrate device views highlighting the inward orientation: the sensing layer was laminated on the inner arc of the curved patch to maximize skin contact on concave surfaces.

[0254]

[0248] FIG. 56G to FIG. 56H illustrate device views of the outward orientation: the sensing layer was positioned on the outer arc for convex surfaces or when outward routing reduced shear, maintaining adhesion during large-angle movements.

[0249] FIG. 57A to FIG. 57B illustrate schematic of a TRAiLL adaptor. The design and values of the TIA components are further shown in FIG. 57B.

[0255]

[0250] FIG. 57C illustrates a top-down photo of the flexible TRAiLL adaptor.

[0256]

[0251] FIG. 58A illustrates a schematic of the TRAiLL Pico board.

[0257]

[0252] FIG. 58B illustrates a top-down photo of the flexible TRAiLL Pico board.

[0258]

[0253] FIG. 59A illustrates (on the left) a setup of the Monte-Carlo simulation for the musclefat-skin model for the relaxed status. Layers are epidermis and dermis (1 mm + 3 mm, as shown by reference number 5901), subcutaneous fat (4-8 mm, as shown by reference number 5903), and muscle (as shown by reference number 5905). A near-infrared LED (as shown by reference number 5907) and a photodiode (PD as shown by reference number 5909) were positioned on the skin surface on opposite sides of the apex. The LED was oriented along the local skin normal. The right portion of FIG. 59A log-fluence isosurface visualize the sub-surface sampling volume between source and detector.

[0259]

[0254] FIG. 59B is based on the same illustration format as those shown in FIG. 59A but for flexed statues.

[0260]

[0255] FIG. 60A illustrates the setup for the muscle density test.

[0261]

[0256] FIG. 60B illustrates a detector readout versus the muscle scattering coefficient ( / zs) scaled from 1.0 to 5.0 relative to baseline optical properties. Each point is the mean of 5 independent simulations error bars denote 95% confidence intervals. FIG. 60B shows that increasing muscle isproduced a monotonic increase in detected signal amplitude.

[0262]

[0257] FIG. 61A to FIG. 61B illustrate photon-path density visualization for the relaxed state (as shown in FIG. 61A) and flexed state (as shown in FIG. 61B). The simulated skin surface is shown, and the surface LED and photodiode were placed symmetrically around the apex and the LED was oriented along the local skin normal. The volumes depict the photon-measurement density function, highlighting regions traversed by photons that contribute to the detected signal for a 1-mm-radius PD.

[0263]

[0258] FIG. 61C illustrates near-surface path-length distributions at the PD location (including line 6101 that corresponds to a flexed state and line 6103 that corresponds to a relaxed state), derived from time-resolved fluence. Curves are area-normalized. Flexion produced a left-shift of the distribution with a reduced long-path tail relative to the relaxed state, indicating shorter average optical paths through tissue.

[0259] FIG. 62 illustrates finger-specific activation and relaxation trajectories extracted from TRAiLL signals. In FIG. 62, five stacked panels (from thumb to pinky) summarize the temporal shape of single-finger contractions obtained from the most-informative channel per instance. For each finger, individual flex / activation segments (“up”) and relaxation segments (“down”) were min-max normalized to [0,1] and the down segments were time-reversed so both phases progress left-to-right. Dark curves show across-trial means (up and reversed-down). The x-axis denotes sample index within the padded segment.

[0264]

[0260] FIG. 63 illustrates example the spatiotemporal tissue morph maps (also referred to as “spatiotemporal TRAiLL maps” or “heat maps”) of thumb flexion across subjects and repetitions. Heat maps show normalized optical signals from the TRAiLL patch during isolated thumb flexion trials; time is on the x-axis and sensor channel on the y-axis. Columns correspond to three subjects (left to right), and rows indicate repeated trials from the same subject. Color denotes the relative change in photovoltage.

[0265]

[0261] FIG. 64 illustrates spatiotemporal TRAiLL maps of index finger flexion across subjects and repetitions. Heat maps show normalized optical sensing signals from the TRAiLL patch during isolated index finger flexion trials; time is on the x-axis and sensor channel on the y-axis. Columns correspond to three subjects (left to right), and rows indicate repeated trials from the same subject. Color denotes the relative change in photovoltage.

[0266]

[0262] FIG. 65 illustrates spatiotemporal TRAiLL maps of middle finger flexion across subjects and repetitions. Heat maps show normalized optical signals from the TRAiLL patch during isolated middle finger flexion trials; time is on the x-axis and sensor channel on the y-axis. Columns correspond to three subjects (left to right), and rows indicate repeated trials from the same subject. Color denotes the relative change in photovoltage.

[0267]

[0263] FIG. 66 illustrates spatiotemporal TRAiLL maps of ring finger flexion across subjects and repetitions. Heat maps show normalized optical signals from the TRAiLL patch during isolated ring finger flexion trials; time is on the x-axis and sensor channel on the y-axis. Columns correspond to three subjects (left to right), and rows indicate repeated trials from the same subject. Color denotes the relative change in photovoltage.

[0268]

[0264] FIG. 67 illustrates spatiotemporal TRAiLL maps of pinky finger flexion across subjects and repetitions. Heat maps show normalized optical signals from the TRAiLL patch during isolated pinky finger flexion trials; time is on the x-axis and sensor channel on the y-axis. Columns correspond to three subjects (left to right), and rows indicate repeated trials from the same subject. Color denotes the relative change in photovoltage.

[0269] [2651 FIG. 68A to FIG. 68B illustrate channel-wise photovoltage traces recorded by the TRAiLL patch placed over the biceps brachii are shown as stacked time series. FIG. 68A illustrates the unloaded curls. FIG. 68B illustrates with a handheld dumbbell (loaded condition).

[0270]

[0266] FIG. 69A to FIG. 69B illustrate channel-wise photovoltage traces recorded by the TRAiLL patch placed over the tricep brachii are shown as stacked time series. FIG. 69A illustrates the unloaded kickbacks. FIG. 69B illustrates with a handheld dumbbell (loaded condition).

[0271]

[0267] FIG. 70A to FIG. 70B illustrate channel-wise photovoltage traces recorded by the TRAiLL patch placed over the bicep brachii are shown as stacked time series. FIG. 70A illustrates the thumb-over grip (TOG). FIG. 70B illustrates the thumb-under grip (TUG).

[0272]

[0268] FIG. 71 A to FIG. 71C illustrate five 30-s excerpts per condition show the MAF from a bicep-worn TRAiLL patch. Traces correspond to the segments highlighted in FIG. 52J (Day 1). The results demonstrated task-dependent loading of the biceps during daily life and support the inference that fatigue accumulation occurs primarily during active work and exercise rather than during sleep. FIG. 71 A illustrates working state where intermittent activation with brief excursions above the fatigue threshold (as highlighted in the shading). FIG. 71B illustrates aerobic exercise state where frequent, sustained bursts with prolonged time above threshold. FIG. 71 C illustrates a sleep state where low, stable signals remaining below threshold.

[0273]

[0269] FIG. 72 A to FIG. 72C illustrate five 30-s excerpts per condition show the MAF from a bicep-worn TRAiLL patch. Traces correspond to the segments highlighted in FIG. 52J (Day 2). The results demonstrated task-dependent loading of the biceps during daily life and support the inference that fatigue accumulation occurs primarily during active work and exercise rather than during sleep. FIG. 72A illustrates working state where intermittent activation with brief excursions above the fatigue threshold (as highlighted in the shading). FIG. 72B illustrates aerobic exercise state where frequent, sustained bursts with prolonged time above threshold. FIG. 72C illustrates the sleep state where low, stable signals remaining below threshold.

[0274]

[0270] FIG. 73A, FIG. 73B, FIG. 73C, and FIG. 73D illustrate measurement examples of a TRAiLL patch during a representative contraction. As shown, raw traces are overlaid with outputs from a denoising autoencoder trained to reconstruct clean sequences and applied channel-wise. The time axis is normalized to the range (0-511).

[0271] FIG. 74A to FIG. 74C illustrate two-dimensional array frames from the TRAiLL patch that are augmented to increase geometric diversity while preserving label semantics. The columns of FIG. 74A to FIG. 74C include the original map and three independent augmented exemplars (For example, Aug 1-3). FIG. 74A illustrates translation augmentations where maps are shifted within the sensor grid, maintaining relative intensity structure. FIG. 74B illustrates rotation augmentations where maps are rotated to emulate varied patch orientations. FIG. 74C illustrates mirror augmentations where horizontal / vertical reflections generate left-right symmetry variants.

[0275]

[0272] FIG. 75A illustrates a pairwise cosine similarity heatmap of 50 augmented samples from the same original sample in the latent space. Each pixel represents the cosine similarity between two samples after passing through the encoder.

[0276]

[0273] FIG. 75B illustrates a pairwise cosine similarity heatmap of 24 augmented samples from different ASL-A gestures in the latent space.

[0277]

[0274] FIG. 75C illustrates the absolute reconstruction error across pre-augmented views. For one sequence, each of 49 augmented views (rows; view-1...49) was encoded by a trained EncoderlDCNN and decoded by a linear Reconstructor; the canonical signal (view-0) served as the target. The heatmap shows the absolute difference between the reconstructed view and the canonical signal, averaged across channels (columns are time samples; hot colormap). Prominent vertical bands mark time segments with higher error shared across views, whereas minimal horizontal structure indicates little view-specific deviation.

[0278]

[0275] FIG. 76A illustrates L2-norm of TRAiLL measurements for 10 repetitions of each right-hand pick — thumb, index, middle, and ring (left to right). In FIG. 76A, solid curves denote the mean temporal envelope while shaded bands show ±1 s.d. across trials.

[0279]

[0276] FIG. 76B illustrates corresponding channel-by-time heat map (including 48 channels where brighter represents higher activation) across the same sequence, along with vertical divisions mark the four picks.

[0280]

[0277] FIG. 77 illustrates power consumption of the TRAiLL system during continuous operation, including a power diagram (shown on the top), a supply voltage diagram (shown in the middle), and a current diagram (shown in the bottom) measured with a bench supply while the device ran continuously. For FIG. 77, the supply was held at 5 V throughout, while the current averaged ~24 mA.

[0278] FIG. 78 illustrates example methods in accordance with some embodiments of the present disclosure. In example shown in FIG. 78, the example method 7800 comprises receiving a plurality of optical sensing signals from a sensor array (step 7802) and generating at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals (step 7804).

[0281] DETAILED DESCRIPTION OF THE INVENTION

[0282]

[0279] Some embodiments of the present disclosure will now be described more fully hereinafter with reference to the accompanying drawings, in which some, but not all embodiments of the disclosure are shown. Indeed, these disclosures may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this disclosure will satisfy applicable legal requirements. Like numbers refer to like elements throughout.

[0283]

[0280] As used herein, terms such as “front,” “rear,” “top,” etc. are used for explanatory purposes in the examples provided below to describe the relative position of certain components or portions of components. Furthermore, as would be evident to one of ordinary skill in the art in light of the present disclosure, the terms “substantially” and “approximately” indicate that the referenced element or associated description is accurate to within applicable engineering tolerances.

[0284]

[0281] As used herein, the term “comprising” means including but not limited to and should be interpreted in the manner it is typically used in the patent context. Use of broader terms such as comprises, includes, and having should be understood to provide support for narrower terms such as consisting of, consisting essentially of, and comprised substantially of.

[0285]

[0282] The phrases “in one embodiment,” “according to one embodiment,” “in some embodiments,” and the like generally mean that the particular feature, structure, or characteristic following the phrase may be included in at least one embodiment of the present disclosure, and may be included in more than one embodiment of the present disclosure (importantly, such phrases do not necessarily refer to the same embodiment).

[0286]

[0283] The word “example” or “exemplary” is used herein to mean “serving as an example, instance, or illustration.” Any implementation described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other implementations.

[0284] If the specification states a component or feature “may,” “can,” “could,” “should,” “would,” “preferably,” “possibly,” “typically,” “optionally,” “for example,” “often,” or “might” (or other such language) be included or have a characteristic, that a specific component or feature is not required to be included or to have the characteristic. Such a component or feature may be optionally included in some embodiments, or it may be excluded.

[0287]

[0285] The term “electronically coupled,” “electronically coupling,” “electronically couple,” “in communication with,” “in electronic communication with,” or “connected” in the present disclosure refers to two or more elements or components being connected through wired means and / or wireless means, such that signals, electrical voltage / current, data and / or information may be transmitted to and / or received from these elements or components.

[0288] LaHMo Platform Overview

[0289]

[0286] Neuromuscular diseases (NMDs), characterized by progressive muscle function deterioration, pose significant challenges to healthcare systems worldwide. They profoundly impact patient mobility, quality of life, and economic burden, especially in the post-pandemic era. Managing NMDs is costly, encompassing direct medical costs, long-term care, lost productivity, and psychological toll on patients and families. These grand challenges draw researchers to develop advanced sensor technologies and data analytics to realize therapeutics and rehabilitation with precision and personalization. Muscle-tracking technology, combining biosensors and analytical algorithms, has emerged as a promising solution for real-time monitoring of specific muscular units. This rapidly evolving field holds immense potential for creating innovative treatments for NMD and advancing the development of user-friendly cybernetic interfaces. Despite its potential for innovative NMD treatment interfaces, current muscle-tracking technologies face challenges in leveraging high spatial resolution, minimizing motion artifacts, ensuring user comfort, and integrating multiple sensing modalities. The anterior neck region comprises both intrinsic and extrinsic laryngeal muscles, each with distinct physiological roles. The intrinsic muscles, including the thyroarytenoid and cricothyroid muscles, are primarily responsible for modulating tension and length of the vocal cords, facilitating phonation. In contrast, the extrinsic muscles, such as the sternohyoid and thyrohyoid, are involved in positioning and stabilizing the larynx during swallowing and respiration. Dysfunction in these muscle groups can manifest as conditions like dysphonia, which often stems from intrinsic muscle impairment affecting vocal cord vibration, and dysphagia, which may involve a complex interplay of both intrinsic and extrinsic muscle dysfunction affecting the coordination of swallowing. Additionally, abnormal activity in these muscles can be associated with persistent post-CO VID dry cough, where monitoring muscle function can provide insights into the extent and impact of these symptoms. Moreover, the frequent and involuntary occurrence of throat clearing, and dry cough can be indicative of dystussia, a cough dysfunction characterized by an impaired cough reflex or coordination, potentially leading to ineffective airway clearance and respiratory complications. Monitoring these symptoms with precision could be crucial for the early detection and management of dystussia, thus enhancing patient outcomes. A nuanced examination of the intrinsic and extrinsic laryngeal muscle groups is essential for accurate diagnosis and therapeutic interventions, as well as for evaluating the progression and treatment of post-CO VID conditions.

[0290]

[0287] Muscle tracking technologies primarily rely on ultrasonic sensors, NIRS sensors, electromyography (EMG) sensors, IMU, and mechano-acoustic (MA) sensors. Although promising, each technology has limitations. Ultrasonic sensors typically necessitate a specialized adhesive layer to effectively transmit ultrasound waves, which can be inconvenient for long-term monitoring. Signals collected from EMG sensors have limited spatial resolution due to inherent noise from electrode crosstalk, electromagnetic interference, and electrocardiographic artifacts. IMU-based muscle trackers offer the advantages of portability, ease of integration into wearable devices, and the ability to provide real-time motion tracking without environmental constraints. However, they may suffer from signal drift over time and require sophisticated algorithms to interpret complex muscle movements. MA sensors often require a high computational load for signal analysis and are not ideal for large area, spatially resolved signal detection. NIRS sensors are advantageous in their fast detection time, relatively low fabrication cost, and deep-tissue penetration. The rich containment of NIR-ab sorbing myoglobin in muscles lead to high sensing selectivity to muscular locomotion. However, many NIRS sensors for muscle tracking heavily rely on strong attachment between the sensor and the skin to ensure LED-tissue coupling, thus making them prone to motion artifacts and compromising the wearing comfort. Furthermore, ultrasonic and NIRS sensors collect signals in the temporal domain capable of monitoring short-term and long-term muscle activities, such as swallowing and static exercises, whereas signals from the IMU, EMG, and MA sensors contain valuable information in the frequency domain capable in distinguishing dry cough, talking, and dynamic exercises. Integrating these multi-modality sensors from a single platform may hold great promise via an Al algorithm in leveraging their respective capabilities while mitigating the limitations.

[0291] [2881 Neuromuscular diseases pose significant health and economic challenges globally, necessitating innovative monitoring technologies to perform personalized, accurate therapeutics and rehabilitation. Many muscle-tracking devices pick up muscular motions either indirectly from mechano-acoustic signatures on skin surface or via ultrasound waves that demands specialized skin adhesion. Various embodiments of the present disclosure provide a wireless wearable system, Laryngeal Health Monitor (LaHMo), designed to be conformally placed on the neck for continuously measuring movements of underlying muscles. The system uses near-infrared (NIR) light that features deep-tissue penetration and strong interaction with myoglobin to capture muscular locomotion. The incorporated inertial measurement unit sensor further decouples the superposition of signals from NIR recordings. Integrating a multimodal ALboosted algorithm based on recurrent neural network, the system accurately classifies activities of physiological events. An adaptive model enables fast individualization without enormous data sources from the target user, facilitating its broad applicability.

[0292]

[0289] Here, various embodiments of the present disclosure a wireless, wearable, multi-modal muscle-tracking system, referred to as LaHMo or wearable muscle monitor, where integrated NIRS and IMU feed a heterogenous Al model for enhanced monitoring of laryngeal muscles. In the present disclosure, the term LaHMo or wearable muscle monitor may be used interchangeably.

[0293]

[0290] In some embodiments, the device features four NIRS sensors to track muscle activity during various physiological events, and a synchronized IMU sensor to audits device global motion serving as a reference for counteracting motion artifacts. The hybrid sensing approach is enhanced with an advanced Al-driven platform and a wireless data collection system, offering improved portability and instantaneous data analysis capabilities. This integration of Al models significantly surpasses traditional methods by providing more accurate predictions and comprehensive insights into muscle behavior, thus enabling more precise monitoring and treatment strategies.

[0294]

[0291] Long-term tests and simulations suggest the potential efficacy of the LaHMo platform for real-world applications, such as monitoring disease progression in neuromuscular disorders, evaluating treatment efficacy, and providing biofeedback for rehabilitation exercises. Example LaHMo platform in accordance with the present disclosure may serve as a general non-invasive, user-friendly solution for assessing neuromuscular function beyond the anterior neck, potentially improving diagnostics and treatment of various neuromuscular disorders. For example, a Monte Carlo simulation and two simultaneous gold standard tests based on EMG and videostroboscopy validate the NIRS technology in tracking muscles of the anterior neck area, visualizing the penetration depth of the sensor. Long-term on-body tests demonstrate the capability of LaHMo in continuously monitoring laryngeal muscle activity during various physiological events, including deep breathing, coughing, swallowing, and exercise. High-level physiological indicators, including respiratory health indicators and exercise intensity indicators, draw real-time LaHMo measurements into instant clinically relevant feedback via the Al models, thus offering advanced point-of-care diagnostics. These tests showcase its potential for real-world applications, such as tracking disease progression, evaluating treatment efficacy, and providing biofeedback for physical rehabilitation and sports performance monitoring related to deep muscular tissue. The LaHMo platform, as a multi-modal wearable intelligent healthcare system, may establish a broadly applicable solution for continuous, non-invasive monitoring of muscular locomotions, with the potential to improve the diagnosis and treatment for a broad range of NMDs.

[0295] Example Designs of LaHMo Platform

[0296]

[0292] Fig. 1A illustrates the utility of an example LaHMo deployed onto the neck with realtime Al-boosted analysis of symptoms related to dysphagia, dysphonia, and other respiratory diseases (e.g. COVID), for facilitating clinical decision-making and rehabilitation in accordance with some embodiments of the present disclosure. In some embodiments, the ergonomic design of the LaHMo patch allows conformal attachment onto the anterior neck region with minimal discomfort. In some embodiments, the patch transmits data to a cloud server capable of real-time Al analysis. In some embodiments, the analysis of LaHMo signals may enable continuous monitoring of long-COVID symptoms of coughing, dysphagia, and dysphonia. In some embodiments, those collected data streaming in real-time inform clinicians and caregivers for deeper interpretations and informed decisions for personalized therapeutics.

[0297]

[0293] Fig. IB provides an example exploded view of an example LaHMo patch which features a flexible design for both ergonomic fit and integrated functionality. In some embodiments, the LaHMo patch comprises a flexible substrate 107 that comprises one or more layers. In some embodiments, the example LaHMo patch (also referred to as a wearable muscle monitor) comprises an IMU, at least one NIRS sensor, and a wireless microcontroller 105 that are disposed on the flexible substrate 107.

[0298] [2941 In some embodiments, the flexible substrate 107 comprises a main island 109 and at least one daughter island such as, but not limited to, the left daughter island 111 and the right daughter 113. In the example shown in Fig. IB, the left daughter island Ill is connected to a left side of the main island 109, and the right daughter island 113 is connected to a right side of the main island 109.

[0299]

[0295] In some embodiments, the example LaHMo patch (also referred to as a wearable muscle monitor) uses two stretchable serpentine hinges to connect two smaller daughter sensing islands with its main island that contains arrays of small holes for good air permeability (for example, as shown in Fig. 7).

[0300]

[0296] For example, as shown in Fig. IB, the sensors on the main island include:

[0301]

[0297] i) an IMU module 101 (for example, LSM6DSOX from STMicroelectronics) for motion tracking, which features a detect range of ±16 g for acceleration and ±2000 degree per second (dps) for angular rate, at a sample rate of 1.6 kHz. In some embodiments, the IMU module 101 is configured to generate global motion signals associated with a user. For example, global motion signals may indicate motions associated with a user’s head.

[0302]

[0298] ii) at least one NIRS sensor 103, each consisting of one NIR light emitting diode (LED) (SFH 4043, Osram) and one photodiode (PD) (VEMD 1060X01, Vishay) for muscle activity monitoring. In some embodiments, injection of NIR light into the skin allows absorption and scattering primarily from the muscle tissue beneath the sensor, as myoglobin richly contained in muscles shows strong absorption in NIR. In some embodiments, part of the backscattered light post to the light-tissue interaction can reach a nearby PD which generates corresponding signals that reflect muscular modulation. In some embodiments, the on-board distance between each pair of PD and LED is set to be 3.5 mm, which has shown optimized sign correspondence and signal-to-noise ratio. In some embodiments, the at least one NIRS sensors 103 is configured to generate local activity signals associated with the user. For example, local activity signals may indicate activities associated with neck muscles of the user.

[0303]

[0299] In some embodiments, the key integrated circuits (ICs) on the main island include:

[0304]

[0300] i) a Bluetooth Low Energy (BLE) microcontroller unit (MCU) (ESP32-C3FH4, Espressif) to support the data acquisition from the sensors, computation tasks, and wireless communication capabilities. In some embodiments, the BLE MCU is communicatively coupled to the IMU module 101 and at least one NIRS sensor 103,

[0305] [3011 ii) an analog-to-digital converter (ADC) (ADS 1115, Texas Instruments) featuring a 16-bit resolution for 4 channels to handle the data coming from the photodiodes before transmitting to the MCU,

[0306]

[0302] iii) two operational amplifiers (Op Amp, TLV9002IDSGT, Texas Instruments) that act as transimpedance amplifiers to preprocess the photovoltages output by the photodiodes, and

[0303] iv) two low dropout (LDO) linear regulators (ADP7118ACPZN-R7, Analog Devices) for power noise removal.

[0307]

[0304] In some embodiments, the two daughter sensing islands also comprises one NIRS sensor each.

[0308]

[0305] In some embodiments, the flexible serpentine hinges are designed to ensure not only the continuity of the electrical connectivity between the islands but also a flexible fit that conforms to the neck’s profile while maintaining structural integrity (for example, as shown in Fig. 8A to Fig. 8G). In some embodiments, a removable lithium-ion battery (Engpow, 150 mAh) is used to power the whole patch, and it can support the device running for 4 hours or longer if using intermittent sleep mode (for example, as shown in Fig. 9A to Fig. 9F). In some embodiments, the whole PCB is encapsulated by biocompatible, flexible, and adhesive polydimethylsiloxane (PDMS, Sylgard™ 182, Dow) layers on the top and bottom. In some embodiments, Fig. 1C displays an actual device worn by a subject on the anterior neck to demonstrate the true-to-size perspective of the LaHMo patch. In some embodiments, the compact and unobtrusive design shows that the patch can be used in everyday settings without hindering the normal activities of the user. A detailed block diagram illustrating the operational mechanism of the proposed LaHMo platform in accordance with some embodiments of the present disclosure appears in Fig. ID. In the example shown in Fig. ID, the wireless microcontroller 115 is communicatively coupled to the IMU 117 and the at least one NIRS sensor 119. For example, the wireless microcontroller 115 may receive global motion signals from the IMU 117 and generate one or more global motion data points based on the global motion signals by incorporating time stamps to the global motion signals. The wireless microcontroller 115 may receive local activity signals from the at least one NIRS sensor 119 and generate one or more local activity data points based on the local activity signals by incorporating time stamps to the local activity signals. The wireless microcontroller 115 may transmit the one or more global motion data points and the one or more local activity data points to the computer device 121 for real-time monitoring and further analysis.

[0309] [3061 In some embodiments, the biosensor section of the LaHMo platform serves as the foundation for data acquisition and initial processing. This includes the communication between the microcontroller and other on-board sensors. Specifically, the LEDs are programmed through the programmable analog outputs of the MCU, while the NIRS signal is sent to the external ADC module before reaching the Inter-Integrated Circuit (I2C) interface of the MCU, together with the data from the IMU, which is then processed through a Madgwick filter (additional details of which are described herein).

[0310]

[0307] In some embodiments, the Al analysis section of the LaHMo platform outlines the algorithms and computational processes that interpret the collected data. First, a BLE client, which can be either a smartphone / watch or a personal computer, acquires the wirelessly transmitted data and presents the processed data in an accessible format for immediate review. Then, further data analysis relies on a pre-trained RNN classification system. This Al-powered analysis can occur either on the edge via embedded systems or in the cloud, with the latter providing the computational power needed for more complex interpretations.

[0311]

[0308] In some embodiments, the human interface section of the LaHMo platform shows the core of the LaHMo platform, emphasizing its practical application in various scenarios ranging from muscle tracking, swallow training, vocal training, surgical recovery monitoring, and others. For patients with neck injuries at home, the LaHMo platform monitors recovery progress and offers suggestions for laryngeal muscle training and rehabilitation optimization based on the determination from the clinic end. In a clinical setting, the LaHMo platform provides healthcare professionals with smart diagnosis capabilities based on the comprehensive statistics acquired from the patient in real time. This bidirectional flow of information fosters a dynamic interaction between patients and clinicians, promoting a more engaged and informed healthcare experience with personalized precision.

[0312] Example Theory and Implementation of Madgwick Filter

[0313]

[0309] The Madgwick filter combines data from accelerometers, gyroscopes, and, optionally, magnetometers to estimate three-dimensional orientation with high precision. Generally, the orientation of a rigid body or coordinate frame can be represented by a quaternion, which is a four- dimensional complex number. For example, if the one arbitrary axis of frame A is represented by the unit vectorAr — [rxryrz, then the orientation of frame B relative to frame A can be represented by the following quaternion.

[0314]

[0315]

[0311] In equation (1), the rotation angle 0 is defined from the angle between frames A and B when they are aligned by the axisAr. With this definition, two more basic quaternion calculations with physical significance can be defined. The conjugate of the quaternion q, is the one that describes the relative orientation of frame A to frame B.

[0316]

[0317]

[0313] The outer product determines the compound orientation.

[0318]

[0319]

[0316] With that, the coordinates translation for a vector v changing from frame A to frame B may be determined. Here,Av andBv are the coordinates of the same vector v under frame A and frame B, respectively.

[0320]

[0317] Bv =Bq ⊗Av ⊗Bq* (5)

[0321]

[0318] This can also be described by a rotation pR that applies to the vectorAv

[0322]

[0323]

[0320] The same rotation can be achieved by performing a sequence of rotations from the alignment with frame A on frame B. These rotations are defined by the following equation.

[0324] ψ = atan2(2q1q2+ 2q3q4, 2q12+ 2q22— 1)

[0325]

[0321] e=- sin-1(2q2q4+ 2q1q3) (7)

[0326] (p- atan2(2q3q4— 2q4q2, 2ql + 2q4— 1)

[0327]

[0322] In equation (7), the function atan2 is the 2-argument arctangent that measures the phase angle of a complex number x + iy. Euler angles / 1, 6, (p are the rotation angles around zB, yB, xB, respectively.

[0323] In a realistic scenario where an IMU is measuring the angular speed and acceleration of itself in the earth frame, the following results may be determined:

[0328]

[0329]

[0329] In equation (8), S and E represent the sensor frame and the earth frame, respectively.sa)tis the angular rate vector [0 a>xjya)zmeasured by the gyroscope, iq^t is the relative orientation changing rate of the earth frame relative to the sensor frame at the time t. With equation (5), the calculated quaternion at time t, fq^t, can be derived from the estimated quaternion of the earth frame at time (t — 1), the orientation changing rate at time t, and the sampling interval of the sensor At (equation (9)). However, with just the observation from the gyroscope, the estimation of the quaternion will not yield a unique solution due to the unknown direction of the earth’s gravity field in the earth frame. This will eventually leave the orientation perpendicular to the gravity field undecided. A complete and unique solution can be found by solving the optimization problem of minimizing the objective function f, whose definition can be found in equation (10). In this function,Ed is a reference vector of the earth frame andSs is the measured direction in the sensor frame. Solving the optimization problem leads to equation (11),

[0330]

[0331] (a > 1, is a noise cancellation factor) is the step-size in the optimization, V / is the gradient of the solution surface, defined by its Jacobian and the optimization function, calculated in equation (12). In equation (12), the general formEd andSs have been replaced by the measurables,Eg —

[0001] , (gravity) andsd — [0 axayaz] (accelerations).

[0332]

[0330] Combining the estimation from equation (8) and equation (11) will lead to the Madgwick filter fusion algorithm.

[0333]

[0334]

[0332] This equation can be simplified by optimizing the choice of yt. Eventually, this equation may be in the following form.

[0335]

[0336]

[0337]

[0336] Here, ft is the divergence rate of

[0338]

[0339] expressed as the magnitude of the quaternion derivative with respect to the gyroscope measurement error, and

[0340]

[0341] is the direction of the estimated error. With equations (14) to (16), the quaternion E est.tat agiven time t and filter out the error item fjEqc,t inthe measurements of gyroscope and accelerometer may be calculated. In some embodiments, At may be selected based on the sample rate of the IMU, which is 0.001 s. In some embodiments, the Madgwick filter may be implemented on the MCU to avoid the error caused by the data transmission delay.

[0342] Example Data Preprocessing Approaches at Various Physiological Events

[0343]

[0337] Data preprocessing improves visualization of muscle activity and prepares for continuous condition classification, both in real time. Here, various embodiments of the present disclosure may make use of a multiprocessing strategy to analyze, visualize, and store acquired data in real time (for example, as shown in Fig. 2A to Fig. 2E). For example, Fig. 2A demonstrates the strategic placement of the NIRS sensors on the neck, highlighting the top left, top right, central, and bottom positions. In some embodiments, these locations are chosen for their proximity to key laryngeal muscles, including the sternohyoid muscle and mylohyoid muscle, involved in various physiological functions (such as, but not limited to, swallowing, cough, speech, respiration, and others). Fig. 2B details the orientation axes — pitch, yaw, and roll — utilized in the data collection process, providing a three-dimensional perspective on how neck movements are recorded. Fig. 2C showcases a LaHMo patch highlighting the relative locations of respective sensors, which correlates with the sensing areas shown in Fig. 2A. In particular, Fig. 2C illustrates a central NIRS 210 sensor disposed on a top portion of the main island 218, a bottom NIRS sensor 212 disposed on a bottom portion of the main island 218, a top left NIRS sensor 214 disposed on the left daughter island 220, and a top right NIRS sensor 216 disposed on the right daughter island 222.

[0344]

[0338] Fig. 2D illustrates the sequential flow of data processing concurrent with the detection of signals by the LaHMo patch. In some embodiments, the initialization stage includes three processes storage, analysis, and visualization. In some embodiments, the storage process creates a tabular database upon instantiation and prepares to receive and store serial data according to a known set of keys. In some embodiments, the analysis process starts acquiring serial data and stores new data in a temporary buffer. In some embodiments, the acquisition of data is accomplished using a custom software and user interface, named BTViz, to handle Bluetooth connection events, store acquired data in a buffer, and visualize acquired data. At each minute interval, this temporary buffer stores a tabular database in the storage process. To improve the signal-to-noise ratio while preserving critical frequency information, various embodiments of the present disclosure may apply a Butterworth filter (for example, 6th order bandpass filter at 0.1-5 Hz for photovoltage, and a 4th order band stop filter at 3-7 Hz or orientation data) to the temporary data buffer. This preprocessed data is then sent to a visualization process to be plotted in real time using BTViz. Simultaneously, a peak finder algorithm is applied to the processed data to extract key pulsatile features. A 3-second temporal slice is then set up centered at each peak found by the algorithm. Simultaneously, these slices are saved to the tabular database for further Al analysis. The graphs in the analysis process and visualization process sections depict the normalized data pre-filter and post-filter application, illustrating the effectiveness of these preprocessing steps. Feature localization is then conducted on the filtered data, where specific physiological characteristics are identified, highlighted by the gray shades which represent key points of interest in the muscle activity patterns.

[0345]

[0339] Fig. 2E presents a comprehensive analysis of physiological events across a spectrum of laryngeal muscle motion frequencies, each row corresponding to a specific sensor location and each graph demonstrating the signal detected during the activity at each location. The left four columns represent the low-frequency activities to be investigated: deep breathing, swallowing, dry coughing, and throat clearing. The right two columns are the high-frequency activities, categorized as aerobic and anaerobic workouts, that are to be decoupled from the low-frequency activities. The overlaid traces within each graph depict individual event cycles, with darker lines indicating the average pattern. This visualization enables the identification of unique signal patterns associated with various neck activities, allowing for further differentiation between voluntary movements and involuntary muscle activity.

[0346] Example Neural Network Classification of Preprocessed Windows

[0347]

[0340] Various embodiments of the present disclosure utilize neural network classification of preprocessed windows to decode natural physiological activities engaging the anterior neck muscle group, often accompanied by head motions. As described above in connection with at least Fig. ID, a computing device may receive one or more global motion data points and one or more local activity data points from a wireless microcontroller. In some embodiments, the one or more global motion data points may indicate motions associated with a head of the user, while the one or more local activity data points indicate activities associated with a neck muscle of the user. In some embodiments, the integration of IMU and NIRS sensors harnesses a comprehensive view, capturing both muscle activities (MA) (based on the one or more local activity data points) and head motions (HM) (based on one or more global motion data points). In some embodiments, this dual modality is central to the development of classification algorithm in accordance with some embodiments of the present disclosure, aiming to extract MA and HM information against the backdrop of ambient signals.

[0348]

[0341] Referring now to Fig. IE, example computer-implemented methods in accordance with some embodiments of the present disclosure are illustrated.

[0349]

[0342] In the example shown in Fig. IE, example computer-implemented methods comprise receiving, by a processor, one or more global motion data points and one or more local activity data points associated with a user at step / operation 123. For example, as described above, the one or more global motion data points may be generated by a wireless microcontroller based on global motion signals from an IMU, and the one or more local activity data points may be generated by the wireless microcontroller based on local activity signals from at least one NIRS sensor.

[0350]

[0343] In some embodiments, example computer-implemented methods comprise generating, by the processor, one or more predicted health status data points associated with the user based at least in part on the one or more global motion data points, the one or more local activity data points, and a dual-channel RNN at step / operation 125.

[0351]

[0344] RNNs, as a family of commonly used classification protocols for time-series data, provide technical advantages in grasping the trend and making predictions based on the development of the temporal slices. Here, various embodiments of the present disclosure develop a dual-channel RNN based on GRUs, named dual-GRU, that can adeptly distinguish between various MA and HM events and effectively leverage the nature of signals collected from both IMU and NIRS sensors.

[0352]

[0345] In some embodiments, the dual-channel structure allows for dedicated processing of IMU and NIRS data, enhancing its ability to stand against motion artifacts and extract relevant features (additional details of which are described herein). Fig. 10 shows the comparison between the LaHMo signals acquired for a cough event while standing still and walking, demonstrating the system capability to filter out motion-induced artifacts during walking. While NIRS data shows increased noise during walking due to physical motion, the IMU data captures characteristic patterns of movement that the dual-GRU model utilizes to isolate and remove those artifacts. The present disclosure provides a systematic comparison between the developed dual-GRU and various other RNN architectures based on fully recurrent neural networks (FRNN), long short-term memory (LSTM), and GRUs along with their bidirectional variants (BiFRNN, BiLSTM and BiGRU) that connect two hidden layers of opposite directions, to demonstrate the enabling capabilities offered by the dual-GRU.

[0353]

[0346] In accordance with some embodiments of the present disclosure, the preprocessed data (for example, as shown in Fig. 2D) is used to construct the dataset used for training and validating the models. Fig. 3A further elucidates the performance difference between the proposed dualchannel RNN (dual-RNN) and a normal mono-channel RNN (mono-RNN). In a mono-RNN architecture, the preprocessed data (with a dimension of [1, 3000, 7]) will be sent directly to the hidden layer (with a dimension of [1, 3000, 140]) that updates for every time point. The hidden layer is then sent to a FC layer ([1, 5]) before finally outputting the result via a SoftMax (SM) layer. In some embodiments, the FC layer refers to a neural network layer in which each neuron applies a linear transformation to the input vector based on weights associated with the inputs. Generally, to gain the best performance of the neural network, this method requires normalization in the preprocessing to avoid a biased weight towards one of the two types of sensors. This not only brings more parameters to fine-tune but also reduces the universality and robustness of the model.

[0354]

[0347] In contrast, the dual-RNN architecture in accordance with some embodiments of the present disclosure employs two parallel hidden layers, each dedicated to processing data from one of the two distinct sensor types. For example, the two parallel hidden layers may comprise a first hidden layer and a second hidden layer. In such an example, the computer-implemented method comprises inputting the one or more global motion data points to the first hidden layer and inputting the one or more local activity data points to the second hidden layer.

[0355]

[0348] As an example, the one or more global motion data points may be generated based on three-channel motion signals (i.e., yaw, pitch, and roll Euler angles from an IMU), and the one or more local data points may be generated based on the four-channel photovoltage signals (i.e., from the central NIRS sensor, the bottom NIRS sensor, the top left NIRS sensor, and the top right NIRS sensor). In such an example, the collected 7-channel data is first divided into the four-channel for photovoltage signals and the three-channel for Euler-angle information before being processed individually through the two designated RNN layers. The outputs of these two RNN layers are then concatenated and sent to the FC and SM layers.

[0356]

[0349] Fig. 3B, Fig. 3C, and Fig. 3D showcase the representative results of the trained RNNs. To emphasize the significance of combining the IMU and NIRS data, the results of the biased model utilizing data from one of the sensors are also displayed as PV-biased and IMU-biased to compare segregated and integrated data (for example, as shown in Fig. 3B). Furthermore, the performances of FRNN, LSTM, and GRU and their bidirectional variants (BiFRNN, BiLSTM, and BiGRU) are also illustrated for the optimization of the architecture, with the loss and accuracy curves of their training and testing datasets presented by Fig. 11 A and Fig. 1 IB. Finally, to prevent the potential training bias that comes from the way the involved dataset is selected from the whole data pool during the training process, a k-fold cross-validation is implemented to validate the model statistically. Fig. 3B, Fig. 3C, Fig. 3D, Fig. 3E, Fig. 3F, and Fig. 3G show a series of evaluations on 24 different RNN architectures based on FRNN, LSTM, and GRU, with a particular focus on the dual-GRU model. For example, Fig. 3B shows the reduction of the cross-entropy loss of all eight types of GRU models (PV-biased GRU, PV-biased BiGRU, IMU-biased GRU, IMU-biased BiGRU, mono-GRU, mono-BiGRU, dual-GRU, and dual-BiGRU), and the training performances for FRNN and LSTM are shown in Fig. 12A to Fig. 12F. During a 400-epoch training process, all eight models show a typical reduction-converge shape, and a good early stopping (ES) point can be observed at epoch = 200. The trends of the plots show a significant difference between the sensor-biased models and the comprehensive models that make use of both modalities of the LaHMo patch. In some embodiments, the mono-RNNs and dual-RNNs display a much steeper learning curve and converge points better than their competitors, who use only one modality (for example, as shown in Fig. 13 A to Fig. 13F). This example highlights the advantage of synergizing NIRS and IMU detectors in accordance with some embodiments of the present disclosure as compared to using only one of the two types of sensors. However, the bidirectional feature does not bring about much of a difference in the learning rate and the final loss function. Fig. 3C highlights the accuracy development among the train and test datasets between dual-GRU and dual-BiGRU. The results show that dual-GRU has a higher accuracy for both train and test datasets compared to dual-BiGRU, and the overfit problem is more severe for dual-BiGRU. This example highlights that the bidirectional RNNs in accordance with some embodiments of the present disclosure offer performance comparable with the conventional RNNs. Fig. 3D, Fig. 14A to Fig. 14H, Fig. 15A to Fig. 15H and Fig. 16A to Fig. 16H show the confusion matrices of the dual-GRU and other networks in the preprocessed data. For dual-GRU, it achieves accuracies of 1.00 for deep breathing, 0.92 for dry coughing, 0.92 for throat clearing, 0.90 for swallowing, 1.00 for aerobic exercising, and 0.91 for anaerobic exercising, with these events respectively represented by labels 0-5 in the confusion matrix.

[0357]

[0350] In some embodiments, the one or more predicted health status data points indicate one or more predicted coughing events associated with the user (for example, when the one or more global motion data points indicate motions associated with a head of the user and the one or more local activity data points indicate activities associated with a neck muscle of the user). Fig. 3E shows the event recognition during a continuous dry cough scenario, with the top half of the plot showing the NIRS readings, and the bottom half of the plot showing the IMU measurements. Reference number 329 and reference number 331 show the manually labeled MA events. In particular, the crosses and dots are manually labeled samples in the train and test datasets in one possible split, respectively. Reference number 333 indicates predicted health status data points that indicate coughing events predicted by the trained GRU model. Fig. 17A to Fig. 17B provide more related examples of other MA and HM events. Dry coughs were one such MA event, with the PD responses from various coughing patterns, including continuous, random, and separate, differentiated by the interval between cough events, successfully visualized (for example, as shown in Fig. 18, Fig. 19, and Fig. 20).

[0358]

[0351] Swallowing, another experimental MA event, was tested under various conditions, including after different numbers of chews, various time intervals between swallows, and different liquids swallowed. In some embodiments, an example LaHMo system successfully visualized the photovoltage data from the various test conditions, demonstrating a clear distinction between the MA of swallowing under different circumstances (for example, as shown in Fig. 21A to Fig. 21C, Fig. 22, and Fig. 23).

[0359]

[0352] Another MA event used for visualization testing was the performance of different vowel phonemes. In some embodiments, an example LaHMo system is capable of not only visualizing but also differentiating between 11 different vowel phonemes (for example, as shown in Fig. 24A to Fig. 24E and Fig. 35). Fig. 36A to Fig. 36F illustrate example LaHMo data during the subject’s production of different pitches, ranging from G4 to C3, along with the corresponding audio segments. Fig. 37A and Fig. 37B extend the analysis by comparing head and chest voice production at G4, and further shows the LaHMo data for whispered and loud sounds at C3, all with their respective audio recordings. HM events were used the LaHMo’ s PD response to motion and the subsequent visualization of the response. Test subjects moved their heads such that their cranial pitch angles were altered, with the LaHMo system successfully detecting and visualizing these motions (for example, as shown in Fig. 38A to Fig. 38E).

[0360]

[0353] To demonstrate the independence of the dual-GRU model in the selection of the training and test sample, a k-fold cross-validation (k = 5) is implemented during the training of the models. The validation is visualized by tracking the standard deviations (STDs) of the loss and accuracy of both the training and the testing datasets (the shaded area in Fig. 3B and Fig. 3C). In some embodiments, a large STD around the training ES point would suggest a poor reproducibility of the model. Fig. 3F shows the development of the accuracy STD of dual-GRU and dual-BiGRU in the 5 folds of training, and the accuracy STD and loss STD plots of other mentioned models can be found in Fig. 39A to Fig. 39C. The STD tracking plot shows a significant reduction for both models, suggesting an increase in prediction robustness. Both models also show a steady and low STD around the 200thepoch, suggesting this point is a decent ES point. To compare the performance of the two models, a statistical examination over various data splits (for example, as shown in Fig. 3G) shows the trajectory of the p-value and t-statistic of the null hypothesis, indicating that the dual-GRU shows consistent performance with the dual-BiGRU during the training process. As seen in the shaded area in Fig. 3G, as the training approaches the ES point, the p-values of the hypotheses for both the training and testing datasets drop to the rejection region (at a significance level of 0.05). Therefore, at the training ES point, it is statistically confirmed that the dual-GRU is more robust when encountering different divisions of inputs, with the data ultimately rejecting the null hypothesis.

[0361] Example Reduction of Motion Artifact with LaHMo Patch and Dual-GRU Model

[0354] In some embodiments, an example LaHMo platform integrates two key strategies to reduce the motion artifact that is ubiquitous in NIRS measurements: i) the use of an IMU to capture motion signals and ii) the application of the dual-GRU model to analyze and differentiate the signals from both the IMU and NIRS modalities.

[0362]

[0355] In some embodiments, the IMU is incorporated into the LaHMo platform to detect and measure motion signals, which is used to calibrate NIRS signals. In some embodiments, the IMU provides real-time data on the subject’s movement, represented as a time series M(t) where t denotes time. In some embodiments, this data helps identify and account for motion artifacts that might otherwise be misinterpreted as physiological signals in the NIRS data IV(t). By capturing M(t), the IMU enables correlating motion data with NIRS data and effectively filter out components of I(t) that are caused by motion rather than true physiological activity. To demonstrate how the dual-GRU model processes both the IMU and NIRS data to isolate physiological signals from motion artifacts, the mechanism of the dual-GRU architecture is described as follows. First, the hidden state of NIRS, / iN(t) can be given by the following expressions:

[0363]

[0356] hN(t) = (1 - zN(t)) ° hN(t - 1) + zN(t) ° h̃N(t)

[0364]

[0357] = tanh(ihwAT(t) + rw(t) ° U tlN^N (t - 1)) + bh)

[0365]

[0358] In the above expressions, zw(t) and rw(t) are the update and reset gates,

[0366]

[0367] and UhNare weight matrices, and bhis the bias term. ° denotes the Hadamard (elementwise) product. The same applies to the hidden state of IMU modality,

[0368]

[0369] After the two hidden states are acquired, the two hidden states are then concatenated to form a combined hidden state before passed through an output layer:

[0370]

[0371]

[0360] y(t) = a W0■ hc(t) + bo)

[0372]

[0361] Here, Woand b0are the weight matrix and the bias term of the dense layer, respectively, a is the rectified linear unit (ReLU) activation function. The output y(t) represents the filtered physiological signal. This layer allows the model to effectively exclude the influence of the motion by assigning different weight for each signal in each modality, ensuring that y(t) primarily reflects true physiological activities. Example Domain Adaptation for System Individualization

[0373]

[0362] Training neural networks can be time-consuming and inconvenient for users. A large neural network, while highly capable, would be limited to a single user, requiring every target user to undertake an extensive training procedure. The main training process demonstrated in the right portion of Fig. 4A uses 99 samples from the sample subject. Having the same number of samples to train the model every time a target user is introduced can bring significant inconvenience and substantial cost of time and resources. To make the platform easily accessible and adaptable to a broad range of users, the network must be small and localized, capable of harnessing a larger remote network while remaining customized to each individual. Domain adaptation is a technique in machine learning that allows a model trained on one domain (source domain) to be applied to another domain (target domain) by leveraging the knowledge learned from the source domain. This approach is particularly useful when the target domain has limited labeled data, as it enables the model to learn from the abundant data in the source domain and adapt to the target domain with minimal additional training.

[0374]

[0363] Various embodiments of the present disclosure overcome the above technical challenges and difficulties. For example, in the context of the LaHMo platform, various embodiments of the present disclosure introduce an adapted version of the dual-GRU model, adaptive-GRU (adap-GRU), which utilizes domain adaptation to facilitate individualization of the Al analysis. The adap-GRU model is designed to be a lightweight adaptation network that can be quickly trained on a small dataset from a target user while leveraging the knowledge gained from a pre-trained dual-GRU model on a comprehensive dataset from a sample subject. The domain adaptation strategy in accordance with some embodiments of the present disclosure shows two significant advantages over its pre-trained base: 1) a significantly smaller size that requires less time to train, and 2) high personalization where users can customize the selection of available actions for a model tailored for individual needs (e.g. speech, swallow, or others).

[0375]

[0364] As shown in Fig. 4A, a small adaptation network, based on the adapted dual-GRU (adap-GRU), employs a series of fast executable additional layers, as an adaptation model, which is preceded by the core neural network architecture previously trained on a comprehensive dataset, a dual-GRU, that serves as the base model. In some embodiments, the adap-GRU network incorporates an adaptation network that comprises two linear transformation layers for each input channel, augmented by rectified linear unit (ReLU) activations, effectively projecting the input data into a space amenable to the domain of the original model without necessitating extensive retraining. Compared with training a dual-GRU, the training for adap-GRU also requires fewer types of samples requiring only four datasets from the user to be fed to a generalized, pre-trained dual-GRU. Fig. 4B shows the data from the target user different from the source user shown in Fig. 2E, thus exhibiting difference in their signal characteristics (which is an expected behavior due to the natural variations between individual characteristics). Data from four physiological activities (dry cough, swallowing, throat clearing, and aerobic workout) is necessary for this training process. The predicted result of adap-GRU is shown in Fig. 40.

[0376]

[0365] To validate the efficacy of this approach, the present disclosure compares the training curve of the adap-GRU and dual-GRU (for example, as shown in Fig. 4C). The results show that adap-GRU is trained much faster with a pre-trained dual-GRU (the curve indicated by reference number 410), compared to the training process of starting a new dual-GRU network. The ES point can be chosen to be epoch = 30, where the training dataset reaches an accuracy of 90.3% and the testing dataset reaches an accuracy of 85.7%, whereas the accuracies at this stage for dual-GRU are less, the training dataset at 65.4% and the testing dataset at 42.9%. Upon completion of the training, the adap-GRU was tested, predicting physiological events executed by an unknown user (for example, as shown in Fig. 40). The subsequent confusion matrix of the four training activities shows a similar performance to that of the dual-GRU trained on the sample subject, confirming the effectiveness of adap-GRU in accordance with some embodiments of the present disclosure on target user system individualization (for example, as shown in Fig. 4D). The accuracy of adap-GRU in incident classification was 0.81 for dry coughing, 1.00 for swallowing, 0.90 for throat clearing, and 0.90 for aerobic exercising, with the physiological events respectively represented by the labels 1, 2, 4, and 5. Because it yields a similar performance of the adap-GRU to the source user-trained dual-GRU, with fewer data inputs and fewer training epochs, various embodiments of the present disclosure provide technical benefits and advantages such as, but not limited to, enabling facile adaptability to be inclusive in both user adoption and activity categorization.

[0377] Example Long-Term LaHMo Test

[0378]

[0366] The long-term performance and stability of wearable devices are crucial for their practical utility. To validate the LaHMo platform’s efficacy in real-world scenarios, the present disclosure includes an example 30-minute on-body test during a basketball game. In such an example, the test subject wore the LaHMo patch on the anterior neck while performing various activities, including deep breathing, coughing, anaerobic exercises (e.g., shooting the ball), and aerobic exercises (e.g., dribbling and jogging). Fig. 5 A presents a subset of the raw data collected from this on-body test, with the line indicated by reference number 503 representing the yaw angle as the representative Euler angle detected by the IMU sensor and the line indicated by reference number 501 representing the photovoltage detected by a representative NIRS sensor. During the 2000-second recording period, 230 moments were manually marked as occurrences of coughs (71 occurrences shown as circles in Fig. 5 A) or non-cough events (159 occurrences shown as squares in Fig. 5A). Fig. 5B presents 3-second intervals of the raw data showcasing the correspondence with certain activities executed by the subject during the long-term test, including running, shooting, dribbling, and attempting a layup. Fig. 5C displays the accuracy and loss curves for the model’s training (solid line) and testing (dashed line), with the model shown to become increasingly accurate and confident in predicting physiological events as the test progressed over 100 epochs. The confusion matrix of the model’s predictions of activity compared to the actual activities is presented in Fig. 5D, further showing the model’s efficacy over the experiment. It reports event identification accuracies of 1.00 for deep breaths, 0.99 for dry coughs, 0.81 for anaerobic workouts, and 0.83 for aerobic workouts, labeled from 0 through 3, respectively. Fig.

[0379] 5E presents the performance of the Al model in action identification over time. For each 3-second slice that the preprocessor decides that a feature occurs, the model will attempt to identify the current action by computing the score for each label. If the score for a certain label is the most significant among all the other labels and is over 0.6, a prediction will be made. If the maximum score for a given 3-second slice is below 0.6, the model will terminate the attempt and output as an “unknown” label. Within each vertical space are pillars representing the possibility of the model classifying the event as a deep breath, cough, anaerobic workout motion, and aerobic workout motion. A higher pillar shows a higher score on the corresponding label. These predictions enable the calculation of high-level health indicators (for example, as shown in Fig. 5F and Fig. 5G). Fig.

[0380] 5F presents the respiratory health indicator, defined as the average prediction score of cough events within a 5-second window, accompanied by the number of observed cough events in that period. Fig. 5G illustrates the exercise intensity indicator, determined by the combined average prediction scores of anaerobic and aerobic workouts within a 5-second window, along with the number of observed workout events. These results enable quantitatively assessment of the relationship between the subject’s respiratory health condition and their exercise intensity during the workout. By comparing the respiratory health indicator with the exercise intensity indicator, various embodiments of the present disclosure evaluate whether increased exercise intensity correlates with a decline in respiratory health or vice versa. This quantitative analysis provides valuable insights into how physical activity impacts respiratory function, helping to identify potential issues and optimize training and health monitoring protocols for individuals based on their specific needs and conditions.

[0381] Example Validation with gold-standard and Monte Carlo simulation

[0382]

[0367] To illustrate the underlying correspondence between the signals collected from LaHMo and the muscular movements of the neck, a time-resolved Monte Carlo (MC) simulation is utilized to present a view of both internal muscular movements and corresponding optical modulations in the laryngeal. In addition to the simulation, the present disclosure provides gold-standard validation using EMG, where LaHMo was compared against EMG recordings of the thyrohyoid membrane, cricothyroid muscle, and suprahyoid muscles during controlled swallowing events. This comparison revealed a high degree of signal alignment, confirming LaHMo’ s ability to track swallowing movements in real time, corroborating the simulation with physiological measurements. The present disclosure further provides a videostrobescopy (VSS) test to highlight the alignment of the LaHMo data with the activity of vocal folds during the production of natural musical scales using chest voice and falsetto, respectively.

[0383]

[0368] In each MC simulation experiment, 107photon wave packets are launched into a predefined voxel space, and the trajectory and weight of each wave packet along all the time steps are recorded (additional details are described herein). Fig. 6A shows the laryngeal region with contour outlines defining the voxel space emulated in the simulation. The image depicts the LP of the throat, that defines sensing areas of the LaHMo patch. Accompanying this is a cross-sectional model of the layered tissue at the site: the epidermis, dermis, subcutaneous fat, and skeletal muscle. A close-up video is first recorded to show a frame-by-frame profile change of the laryngeal area. The contour of the laryngeal prominence (LP) is manually extracted from the video for every frame and is modeled in the voxel space. Finally, a series of parallel boundaries inside the voxel space is generated to represent the borders between the epidermis, dermis, subcutaneous fat, and skeletal muscles. The LaHMo placement and light emission, both on a tissue surface and the LP, are respectively represented by Fig. 41 A to Fig. 41D and Fig. 42A to Fig. 42D. During each physiological activity (deep breath, dry cough, throat clearing, and swallowing) in this simulation, the motion will be mainly induced by the vertical and horizontal displacement of the LP.

[0384] Example Monte Carlo Simulation Setup for Device Validation and Data Visualization

[0369] Monte Carlo simulation implementation. Monte Carlo (MC) methods may be used in connection with statistical computing in complex systems. For example, in an optical simulation in a biological tissue, a large group of photons are launched into the test area, namely, a voxel space, and can migrate in the voxel space under the boundary condition regarding each voxel. In this example, a time-stepped simulation is used to show the maximum scatter profile in a short period (nanosecond-level) after the launching of the photons, given the assumption that the photons shot from the light source are continuous. The main consideration of the voxel boundary conditions includes reflection, scattering, and attenuation.

[0385]

[0370] (1) Reflection. This calculation is based on the reflection coefficient of the current medium. According to Fresnel’s equation, the photon wave packet vector is mirrored when a different medium is detected at the next voxel along the vector.

[0386]

[0371] (2) Scattering. This behavior is determined by the scattering coefficient of the current medium. A scattering length and scattering direction vector are calculated before the scattering event and is implemented after the event is committed by comparing the traveled distance from the last event and the scattering length.

[0387]

[0372] (3) Attenuation. This behavior is determined by the absorption coefficient of the current medium. For every time step of the simulation, the photon packet weight is reduced by this coefficient.

[0388]

[0373] System setup considerations. In the configuration of the MC optical simulation in human tissue, the most important factors are light source setup and voxel space construction.

[0389]

[0374] (1) Light source. The light source needs to be fine-tuned in a MC setup to guide the initial orientations of the photon wave packet vectors. Most wearable NIRS devices use an NIR LED as their light sources who has a spatial light intensity distribution known as the Lambertian distribution.

[0390]

[0391]

[0376] Here, 6 denotes the angle between the irradiance and the normal axis of the light source. In real scenarios, to accelerate the computation of photon trajectories, it is possible to use an angular Gaussian distribution to approximate the LED intensity distribution.

[0392]

[0393]

[0378] In equation (18), a denotes the variance in the zenith angle of an initial photon wave packet vector launched from the light source. Fig.45 A and Fig. 45B compare the angular Gaussian distribution and Lambertian distribution, confirming the feasibility of this approximation.

[0394]

[0379] (2) Voxel space construction. According to the above, the MC simulation requires a pre-defined behavior at every voxel boundary. Thus, this step requires setting the scattering coefficient [s, the absorption coefficient.a, and the refractive index n of every voxel.

[0395]

[0380] Parameters used in the MC simulation are demonstrated in the Table 1 below.

[0396]

[0397] Table 1 Parameters used in MC simulations.

[0398] *Only used in MC simulations involving LP positioning.

[0381] Fig. 6B presents a three-dimensional graph of three dependent variables, describing the neuromuscular activity in the laryngeal region: the logarithm of the fluence at the LaHMo patch’s CP (central photodiode), the height of the LP (mm) and the vertical displacement (mm) of the LP from a set baseline throughout the Monte Carlo simulation. In some embodiments, the logarithm of the fluence is the calculated brightness of the tissue at the site of the CP during the simulation. In some embodiments, the LP height is the height of the LP’s peak relative to the baseline set at the surface of the throat. In some embodiments, the vertical displacement, finally, is the location of the LP’s peak relative to the vertical midpoint of the throat. A color bar enables a more streamlined interpretation of the figure, indicating the value of the logarithm of the fluence corresponding to the colored region on the three-dimensional graph. The cross-sectional profiles in Fig. 6C and Fig. 43 show the locations (show as triangles) and corresponding emitted nearinfrared light (shown as extrusions) of 2 LEDs (central and bottom ones corresponding to Fig. 2A) from the deployed LaHMo. Fig. 6C shows three cross-section profiles of the emitted light per physiological activity, representing the initiation stage, peak stage, and ending stage, respectively. Fig. 6C also provides a comparison between the average photovoltage response of the central and bottom PDs and their simulated counterparts, with the former represented by solid lines and the latter represented by dashed lines and crosses. In Fig. 6C, the vertical dashed lines designate the times lining up with the respective cross-sectional profiles of the Monte Carlo simulation.

[0399]

[0382] To quantitatively assess the agreement between the Monte Carlo (MC) simulated results and the actual measured signals for the deep breath event, a cross-correlation analysis using a sliding window approach is performed. The MC simulated signals and the measured signals were compared for both the central and bottom channels (for example, as shown in Fig. 44A to Fig.

[0400] 44D). In this analysis, the signals from both measurements and MC simulations are first normalized before being matched on the time axis to compensate for the unit mismatch between fluence (results of MC simulation) and photovoltage (measurements of PD). In some embodiments, the cross-correlation is calculated after the normalization and best matching of the segment and the MC simulated signal. Overall, the central channels for the four activities show higher best correlations at around zero lag (0.77 for deep breath, 0.80 for dry cough, 0.75 for throat clearing, and 0.63 for swallowing) indicating a strong similarity. The best correlations (0.79 for dry cough, 0.29 for dry cough, 0.36 for throat clearing, and 0.50 for swallowing) for bottom channels are less significant but still ample to show a similarity. These examples validate the accuracy of the MC simulation in capturing the underlying muscle activities during different physiological events and support the feasibility of using MC simulations to investigate the relationship between the detected signals and the corresponding physiological events.

[0401]

[0383] To validate the LaHMo’s data with gold-standard, a simultaneous EMG test is performed where the activity patterns of the thyrohyoid muscle, cricothyroid muscle, and suprahyoid muscles during a swallowing event are recorded (for example, as shown in Fig. 6D, Fig. 6E, and Fig. 6F). The validation experiment involved three distinct swallowing events monitored simultaneously by both EMG and LaHMo over a 10-second interval. The EMG data and LaHMo data are shown in Fig. 6E and Fig. 6F, in highlighted boxes for each swallow. The results demonstrate a high level of signal alignment between the EMG and LaHMo across all monitored muscle groups, confirming that LaHMo can track the muscle activities associated with swallowing. To further extend this validation, a proof-of-concept videostroboscopy (VSS) experiment is performed that measured LaHMo’s ability to track vocalization, particularly pitch modulation in chest voice and falsetto. The VSS recording, presented in Fig. 46E to Fig. 46H, was obtained from an authentic external source, and provided a visual confirmation of the vocal fold dynamics. Fig. 46A and Fig. 46C depict LaHMo data as the subject produced natural musical scales ranging from C3 to C4 using chest voice, and C4 to C5 using falsetto. The audio recordings of these events, shown in Fig. 46B and Fig. 46D, correlate the vocal performance with the muscle activity captured by LaHMo. Fig. 46E, Fig. 46F, Fig. 46G, and Fig. 46H provide VSS slices during these vocalizations, showing the distinct movement of the corniculate cartilages (highlighted by dashed circles). The VSS confirms the posterior movement of the corniculate cartilages during falsetto, which aligns with the optical data recorded by LaHMo, further validating the platform’s ability to monitor vocal fold dynamics.

[0402]

[0384] This combination of MC simulations and gold standard measurements, including EMG and VSS, validates the operational robustness of LaHMo, demonstrating that example devices in accordance with some embodiments of the present disclosure accurately captures physiological events such as swallowing and vocalization, while offering insight into both swallowing-related muscular movements and dynamic changes during phonation. Additional Details of Example LaHMo Platform

[0403]

[0385] In the present disclosure, an example LaHMo platform provides a wearable system that leverages NIRS and IMU sensors with a dual-modality Al algorithm to continuously monitor laryngeal muscle activity with high precision. In some embodiments, an example LaHMo platform may exhibit significant potential in providing a direct assessment of the physiological activities of extrinsic laryngeal muscles, such as the sternohyoid, thyrohyoid, and suprahyoid, which are integral to laryngeal positioning and stabilization. Additionally, an example LaHMo platform offers indirect insights into the behavior of intrinsic laryngeal muscles, such as the thyroarytenoid and cricothyroid, which are crucial for vocal fold modulation and phonation. This dual capability enables the example LaHMo platform to enhance the understanding and treatment of conditions such as dysphonia, dysphagia, post-COVID dry cough, and dystussia by offering a comprehensive view of laryngeal muscle function and its impact on these disorders. The integration of NIRS and IMU sensors in the LaHMo patch effectively records and distinguishes muscle activities and head motions. This dual sensing modality is crucial for developing accurate classification algorithms that can distinguish between various physiological events including normal deep breath, cough, swallowing, and two forms of exercise, anaerobic (low frequency) and aerobic (high frequency). The dual-GRU model developed in accordance with some embodiments of the present disclosure exhibited superior performance compared with the other 24 conventional RNN architectures, demonstrating its ability to effectively synergize NIRS and IMU data and accurately classify muscle activity and head motion events. Both the long-term on-body tests and Monte Carlo simulations further validate the efficacy of the LaHMo platform, highlighting its practical potential for real-world applications. The developed adaptation network (adap-GRU) allows for Al-model individualization without extensive training for target users, making the LaHMo platform more accessible and user-friendly.

[0404]

[0386] In some embodiments, the example LaHMo platform is designed to offer valuable insights into daily behaviors of laryngeal muscles, such as swallowing and breathing, by continuously monitoring collective muscle activities in a non-invasive manner. Its long-term wearability and non-invasive tracking lead to its promising utility more on preliminary clinical screening for hospital admission and proactive diagnosis for critical early-stage indicators, but the limited spatial quantification and structural correspondence in the analyzed signals make the system less useful for deep mechanistic analysis of a singular muscle function. Specifically, during the examination of the swallowing process, the LaHMo platform effectively tracks hyolaryngeal elevation and tongue movements, and the preliminary evidence from the IDDSI level tests (for example, as shown in Fig. 48A and Fig. 48B) also suggests that the platform can monitor and differentiate tongue movement across various bolus viscosities, providing valuable real-time biofeedback for initial clinical screening and rehabilitation tracking. Additionally, the integration of an example LaHMo platform in accordance with some embodiments of the present disclosure with other diagnostic tools and treatment modalities may hold great promise to improve patient outcomes and enhance clinical decision-making in treating muscular disorders.

[0405] a. Example Device design and components

[0406]

[0387] The outline of the four flexible PCB (fPCB) islands and three serpentine hinges may be defined in Autodesk AutoCAD 2023 before being incorporated into the fPCB design. In some embodiments, the fPCB schematic and board layout may be finished using Altium Designer (version 24.1.2). The bill of materials (BOM) can be found in Table 2.

[0407]

[0408]

[0409] Table 2. Bill of materials for LaHMo patch

[0410] b. Example Device Fabrication, Programming, and Encapsulation

[0411]

[0388] In some embodiments, panels of fPCB are manufactured according to international standards ISO 9001:2005 and IPC. In some embodiments, the device firmware may be customized from the Arduino framework on the espressif32 platform (Espressif). In some embodiments, the preparation of the PDMS encapsulation layer starts with mixing the PDMS base (97.5 wt%) with the curing agent (2.5 wt%) at 2000 rpm in a mixer (ARE-310, Thinky). In some embodiments, the mixer then runs a deaeration procedure at 2200 rpm. In some embodiments, the product is dropcast to the surface of the fPCB in a petri dish before being cured in the oven (AT09, Across international) at 70 °C for 12 hours. In some embodiments, once cured and cooled to room temperature, any excess PDMS on the back side of the fPCB is cleaned before applying one drop of light cure adhesive (AA 349, Loctite) to the center of each island. In some embodiments, the fPCB is then folded and clamped to avoid movement. In some embodiments, the adhesive layer is cured under a shuttered ultraviolet floodlight (IntelliRay 600, Uvitron) to complete the encapsulation.

[0412] c. Example Data Collection

[0413]

[0389] In some embodiments, an example device is secured to make sure the top and bottom NIRS sensors are above and below the subject’s laryngeal prominent, respectively. In some embodiments, an example 150-mA lithium-ion battery is used to power the device and is taped on the device with a piece of transparent film dressing (5 cm x 5 cm, Tegaderm, 3M). Three of the authors participated in one or more of the 4 tests (for example, as shown in Table 3 below). (1) A short test lasts for 6 minutes. The participant performs the following activities for 1 minute each: deep breathing, dry cough, swallowing, throat clearing, running in place, and push-ups. The participant will perform each activity at his or her own pace for 1 minute. (2) An extended short test lasts for 8 minutes. The participant performs the following activities for 30 seconds each: sitting still, deep breathing, throat clearing, swallowing, yawning, dry cough, randomly talking, sneezing, cervical flexion, cervical extension, cervical rotation, cervical side-bending, sit down and stand up, walking in place, jumping, and biking. (3) A repeating test lasts for 1 to 2 minutes, the participant repeats one of the following activities: dry cough, swallowing, or pronouncing a specific phoneme. (4) An exercise test lasts for 1 hour. The participant performs the following activities on a basketball court with a basketball: dribbling around the court, switchback run, shooting, and layups. The data is collected via BTViz.

[0414]

[0415] Table 3 Detailed information of the activity (ies) each participant performed

[0416] d. Example Device Firmware Operating Principle

[0417]

[0390] Sensor Initialization. In some embodiments, a four-channel ADC (ADS1115) is initialized with the following settings: measurement mode - continue; gain level - 4.096V; data rate - 860 samples per second. In some embodiments, this setup will enable continuous reading from four PDs. In some embodiments, the IMU sensor is initialized with the following settings: accelerator measure scale - ±2g; gyroscope measure scale - 250 degrees per second; output data rate (ODR) - 1000 Hz. These parameters are tailored to capture high-precision movement data. In some embodiments, the Madgwick filter is employed with the IMU, providing a real-time approximation of the device’s orientation in space (Euler angles: yaw, pitch, and roll). In some embodiments, the sampling rate of the Madgwick filter is set to 1000 Hz, which is the same sampling rate as the IMU.

[0418]

[0391] Data Acquisition and Processing. Since the ambient environment can be unpredictable, wireless data transmission can sometimes jam the on-board data acquisition. To resolve this, an example system in accordance with some embodiments of the present disclosure utilizes a hardware timer to manage data acquisition intervals, ensuring consistent sampling rates for accurate signal processing, independent of the data transmission speed. In some embodiments, the hardware timer is set to send a trigger every millisecond. In some embodiments, the trigger signal enables one data update at all seven channels. In one iteration, the MCU first measures the current time from the device boot. In some embodiments, the IMU data then sends the six measurements to the Madgwick filter, which updates the real-time Euler angles. Finally, in some embodiments, the ADC accesses the photovoltages from four PDs in the following sequence: central, bottom, top left, and top right. In some embodiments, the data is compiled into a comma-separated string to minimize the size of the whole data package.

[0419]

[0392] Wireless Communication and user-interactivity. In some embodiments, BLE technology facilitates real-time data transmission to a connected client, such as another MCU, a smartphone, or a PC. In some embodiments, the system dynamically adjusts its operation based on the connection status, ensuring data are broadcast when a client is connected and conserving resources when in standby. Additionally, in some embodiments, BLE connection callbacks manage the system’s advertising state, ensuring it remains discoverable by potential clients after disconnected. A real-time, multiprocessing Bluetooth app, BTViz, is used to read, preprocess, and visualize data (for example, as shown in Fig. 47). For example, the user may first press the “Scan for Devices” button for the app to scan for all advertising Bluetooth devices in the environment. Then, the user may select the “LaHMo” sensor from the list and select connect to device. From there, the user may select the corresponding service and characteristics from the list to read using “Read Service” button and double click at the characteristic. Finally, BTViz may show the realtime plotting of the selected measurement.

[0420] e. Example Data Analytics

[0421]

[0393] In some embodiments, the real-time data preprocessing procedures during the tests are deployed on a personal computer running Linux (Manjaro), including two Butterworth filters, one peak finding function, and one slice stacking algorithm. The detailed calculations are described herein. All the above-mentioned neural networks may be built with the PyTorch package (version 2.0.0, based on CUDA 12.1 platform). In some embodiments, the network dimensional parameters of the 24 tested RNN variants and the adap-GRU can be found in Table 4 below. In some embodiments, example training and validation were conducted in Visual Studio Code (version 1.86) environment embedded with Python 3.10.11.

[0422]

[0423]

[0424]

[0425] Table 4 List of the layer composition and dimensions of all tested networks

[0426] f. Example Data Preprocessing Principle

[0427]

[0394] Butterworth filter. Butterworth filter is an infinite impulse response (IIR) filter with a transfer function H(s) = - where / / (s) is the transfer function in the complex frequency i+(— )

[0428] domain (s-plane), <ucis the cutoff frequency (in radians per second), and N is the order of the filter. To implement the Butterworth filter in a digital system, the transfer function needs to be converted from the continuous-time domain (s-plane) to the discrete-time domain (z-plane). This is done using the bilinear transformation, which maps the left half of the s-plane to the inside of the unit circle in the z-plane.

[0429] I3’5's= r

[0430]

[0431]

[0396] Here, T is the sampling period. Applying the bilinear transformation to the Butterworth transfer function yields the discrete-time transfer function.

[0432]

[0433]

[0398] Here, btare the numerator coefficients, atare the denominator coefficients, M and N are the order of the numerator and denominator polynomials. After this, the fdter transfer function is further factored into a series of second-order sections (SOS), each with its own numerator and denominator coefficients. This transfer function of SOS is given by:

[0434]

[0435]

[0400] where bi0, btl, bi2are the numerator coefficients of the i-th SOS, and atl, ai2are the denominator coefficients of the i-th SOS. With this, the overall transfer function of the filter can be written as H(z~) = W1(z) ■ W2(z) ■... ■ Hn(z), where n is the number of SOSs.

[0436]

[0401] After the SOS arrays are obtained, the input signal is first padded on both ends to minimize startup and ending transients. Then the SOS filters are applied to the input signal in the forward direction using the difference equation y[n] = biox [n] + b;ix[n - 1] + bi2x[n - 2] -atly [n — 1] — ai2y[n — 2], where x[n] is the input signal at the time step n, and y [n] is the output signal at time step n. With that, all the SOS arrays are reversed before reapplied to the previously filtered signal. The padded length is trimmed to recover the initial signal length before the final output.

[0437]

[0402] Peak locating. The implementation of peak locating is based on the fmd_peaks() function from the SciPy library. The algorithm used by find_peaks() can be separated into several parts. First, the difference between adjacent elements in the input array x is calculated to form a new array dx. Second, in the array dx, the point where it changes from positive to negative is located. The found point(s) indicate a critical point of the original array x. Specifically, find_peaks() allow additional criteria, and the one used in this protocol is the distance, namely, the minimum distance by which peaks are considered. If one peak is detected within the minimum distance from the last peak, it will not be counted.

[0438]

[0403] Peak expansion. This step gives a window of specified size n for final classification. The algorithm first picks the peaks from the previous step, then for each peak, the algorithm creates an empty window of size n before filling it with data from original signal starting from

[0439]

[0440] p — where p is the index of the peak. Repeating this process will eventually result in an expanded array with a dimension (Np, ri), where Npis the number of peaks.

[0441] g. Monte Carlo simulation

[0442]

[0404] In the example MC simulation, the platform used is MCXLAB v2020 (1.8 - Furious Fermion), an open-source light transport simulator. In some embodiments, two MC optical simulations are performed: normal status (NS) and bump translate (BP). Within the NS portion, the skin geometry was initialized as a voxel space containing a set of parallel layers: the epidermis, dermis, subcutaneous adipose, and muscle, whose thicknesses were found to be 1.5 mm, 3.5 mm, 10 mm, and 35 mm, respectively. In some embodiments, the simulation itself was configured by establishing the seed (i.e., the starting point for random number generation), the number of photons used, and the level of debug output of the MC simulation. The configurations were then used to define the light detector and source: the near-infrared photon packets emitted from the source would reflect perpendicularly, following an angular Gaussian distribution to approximate the LED intensity distribution. After configuring the plots in which the results would be graphed, this information was used by the BP portion for data visualization. In a set 60-by-10 array, several inputs were established for the tracking and representation of the movement of the LP over the course of the physiological activities identified for the experiment. The simulation began at the timepoint designated as 0, with the time step and end respectively defined as 1 and 500 picoseconds. The voxel space is modified from the parallel pattern used in the first simulation. Here, a 2D Gaussian function (GF) on the layer boundary is used to approximate the shape of the LP. The mean value of the GF defines the vertical location of the LP’s peak. The amplitude of the GF defines the horizontal location of the LP’s peak relative to the surface of the skin. The variance of the GF defines the size of the LP. In addition to those, the tissue thicknesses, offset, and box size, are utilized to generate the Gaussian approximation of the LP. This generation was looped 600 times over the simulation — with one run per 0.5 nanoseconds — recording the fluence at every voxel at the end of each run, the movement of the LP, and the trajectory and weight loss of each emitted near-infrared photon packet as it traveled along its path (for example, as shown in Fig. 6B and Fig. 6C).

[0443] h. Example EMG validation test

[0444]

[0405] An example, EMG validation test may start by placing 9 electrodes (Kendall™, CardinalHealth) onto the subject’s neck at the labeled locations. Eight electrodes are connected to channel 1-4 of a PowerLab 16 / 35 data acquisition system (ADInstruments), and the last electrode is connected to the external ground port. The data acquisition is performed by the LabChart 7 software. For all the input channels, the sample frequency is set to 100 kHz. On all the input differential amplifiers, a bandpass filter is added, with a 1 Hz low cutoff and a 50 Hz high cutoff. A LaHMo patch is applied after all electrodes are placed. The subject uses a straw to drink a mouthful of water to prevent redundant mouth movement before swallowing the liquid naturally.

[0445] TRAiLL System Overview

[0446]

[0406] Seamlessly interpreting human motor intent comfortably and over a long term is a central challenge in developing next-generation human-machine interfaces and personalized medicine. Existing wearable sensors based on electromyography, ultrasound, and camera imaging are often limited by motion artifacts, cumbersome setup, or environmental dependencies.

[0447]

[0407] Various embodiments of the present disclosure provide TRAiLL (Tracking and Reconstructing Array of Near-infrared Sensing for Limb Locomotion) system that includes a skin-conformal, near-infrared spectroscopic sensor array that provides high-resolution mapping of subcutaneous muscle dynamics without conductive gels or external cameras. By pairing this hardware with a self-supervised learning (SSL) pipeline, various embodiments of the present disclosure translate the complex spatiotemporal data streams into robust, high-fidelity classifications of physiological state and motor intent. For example, TRAiLL system can resolve fine motor control by differentiating muscle activity from individual finger flexions and quantify complex physiological states, such as load-dependent activation and neuromuscular fatigue, during resistance exercises. In a complex classification task, TRAiLL system can achieve 92% cross-validation accuracy in interpreting 24 static gestures of the American Sign Language alphabet. Furthermore, TRAiLL system enables dynamic human-computer interaction through the bimanual control of a virtual musical instrument and camera-free manipulation of 3D objects, achieving 96% accuracy for a six-gesture control task. The TRAiLL system represents a versatile platform for creating robust, unobtrusive bio-interfaces, paving the way for next-generation applications in personalized rehabilitation, assistive communication, and immersive virtual interaction.

[0448]

[0408] Continuous, precise monitoring of subcutaneous-muscle movement holds substantial clinical and biomechanical significance, including the accurate assessment of motor impairments, tracking recovery trajectories in rehabilitation therapy, and enabling early detection of neuromuscular disorders such as Parkinson’s disease and stroke-related dysfunction. Moreover, biomechanical analyses derived from precise muscle activity monitoring can inform personalized interventions and training protocols, potentially enhancing athletic performance and preventing injury through optimized movement patterns. These capabilities are also paramount for engineering a new class of human-machine interfaces capable of interpreting subtle motor intent directly from the user’s underlying physiology.

[0449]

[0409] Recent advancements in wearable biosensors have significantly enhanced the capabilities for monitoring muscular locomotion, enabling sophisticated interactions between humans and computing devices. Among these technologies, Near-Infrared Spectroscopy (NIRS) has emerged as a promising method due to its non-invasiveness and sensitivity to physiological changes in muscle tissues. However, many NIRS devices are typically rigid and cumbersome, limiting their practical usability during extended periods of motion. Furthermore, the nonlinear correlation between sensing signal and measurement targets may further complicate the calibration metrics and device design. To overcome these limitations, the TRAiLL system in accordance with some embodiments of the present disclosure provides a flexible, breathable NIRS-based wearable sensor array designed explicitly for unobtrusive, long-term, and continuous mapping of local muscular activities (for example, as shown in FIG. 50A to FIG. 50G). To interpret the complex, high-dimensional data streams generated by TRAiLL and mitigate motion artifacts without requiring extensive labeled datasets, various embodiments of the present disclosure employ a selfsupervised learning (SSL) pipeline. This computational approach enables the model to first learn robust, denoised representations of the underlying muscular activity before being fine-tuned for high-accuracy gesture classification.

[0410] Unlike electromyography (EMG) and ultrasound sensors that require conductive gels and suffer from susceptibility to motion artifacts, TRAiLL’s gel-free optical approach ensures robust performance during dynamic movements. By directly imaging chromophore variations related to the physical and hemodynamic changes in the muscle, TRAiLL also offers higher spatial specificity than surface EMG, which is prone to electrical crosstalk between adjacent muscles. Furthermore, its low-cost, simple hardware provides a spatiotemporal map of muscle activity that is more directly suited for continuous classification tasks than the complex structural images generated by ultrasound. In addition to that, compared to camera-based systems, which rely on ambient lighting or dedicated IR cameras that not only increase complexity and cost but also introduces constraints from the surrounding environment, TRAiLL utilizes integrated LEDs and photodiodes in a highly wearable and breathable form directly interfaces with the local skin, eliminating external dependencies. Pressure and inertial measurement unit (IMU)-based wearables require placement on specific joints or fingers, which can restrict natural hand movements, whereas TRAiLL’s flexible NIRS array can monitor muscle activity from unobtrusive locations on the limb.

[0450]

[0411] For example, TRAiLL in accordance with some embodiments of the present disclosure provides versatility across three applications, including rehabilitation exercise evaluation, realtime interpretation of American Sign Language (ASL), and intuitive camera-free interactions in virtual reality environments (for example, as shown in FIG. 50E). For rehabilitation exercise evaluation, TRAiLL provides quantitative measures of muscle load and fatigue, offering physiological insights beyond the kinematic data provided by IMU or camera-based systems. In ASL interpretation, it enables a private, non-optical approach that is robust to environmental lighting and occlusions. Finally, for virtual interactions, it facilitates seamless, hands-free control by decoding motor intent directly from the forearm, offering a more intuitive alternative to handheld controllers. These applications collectively highlight both the robustness of the TRAiLL system and its broad applicability in human-machine interaction.

[0451] Example Design and Characterization of the TRAiLL System

[0452]

[0412] Various embodiments of the present disclosure provide example wearable muscle monitors (that provides technical improvements and advantages over other devices.

[0413] In some embodiments, an example wearable muscle monitor comprises a sensor array (for example, the sensor array 5004 shown in FIG. 50A ad FIG. 50F) and a computing device (for example, the computing device 5006 shown in FIG. 50F) communicatively coupled to the sensor array. In some embodiments, the sensor array is a fully integrated, skin-conformal, and designed for robust, long-term monitoring of subcutaneous muscle activity (for example, as shown in FIG.

[0453] 50A to FIG. 50G).

[0454]

[0414] In some embodiments, the sensor array (for example, the sensor array 5004 shown in FIG. 50A ad FIG. 50F) comprises a plurality of sensing pixels (for example, the sensing pixel 5008 shown in FIG. 50A) which are the base sensing units. In some embodiments, each of the plurality of sensing pixels for example, the sensing pixel 5008 shown in FIG. 50A) comprises a nearinfrared light-emitting diode (NIR LED) (for example, the NIR LED 5010 shown in FIG. 50A) and a photodiode (PD) (for example, the PD 5012 shown in FIG. 50A).

[0455]

[0415] In some embodiments, the NIR LED is a high-efficiency near-infrared light-emitting diode (for example, with a peak wavelength λ = 950 nm) that injects light into the tissue. In some embodiments, the matched silicon PD is positioned 3.5 mm away to capture backscattered photons (for example, as shown in FIG. 50A).

[0456]

[0416] When NIR light penetrates the subcutaneous layers, it undergoes multiple scattering and absorption events. As muscles contract, its morphology, density, and local hemodynamics are altered, which in turn modulates the optical path of these backscattered photons, resulting in a measurable change in the light intensity captured by the photodiode.

[0457]

[0417] In some embodiments, these sensing pixels are fabricated on a thin (for example, 240 μm), flexible and stretchable polyethylene terephthalate (PET) substrate (for example, as shown in FIG. 55A to FIG. 551). As demonstrated in-situ, this construction allows the device to adhere conformally to the skin and maintain robust functionality during complex muscular movements, such as bending of the forearm (for example, as shown in FIG. 50G and FIG. 56A to FIG. 56H).

[0458]

[0418] Deployments on representative anatomical locations, such as the forearm and wrist (for example, as shown in FIG. 50B), confirmed the device’s capacity to generate high-resolution spatiotemporal tissue morph maps of tissue reflectance. These maps directly correlate with the activation patterns of underlying muscle groups, such as the flexor digitorum superficialis, enabling the quantification of muscle activation and fatigue (for example, as shown in FIG. 50C) and the recognition of complex gestures (for example, as shown in FIG. 50D). As summarized schematically (for example, as shown in FIG. 50E), these high-fidelity spatiotemporal tissue morph maps form the basis for applications such as quantitative rehabilitation evaluation, realtime interpretation of signed gestures, and intuitive, camera-free interaction within virtual environments.

[0459]

[0419] FIG. 50F details an example system block diagram of an example wearable muscle monitor.

[0460]

[0420] In some embodiments, the 48 sensing pixels are arranged in a 6x8 array, light emission intensity is controlled via pulse-width modulation (PWM), and the resulting photocurrent from each photodiode is converted to a voltage by a trans-impedance amplifier (TIA, for example, TLV9002IDDFR by Texas Instruments®) array (for example, as shown in FIG. 57A to FIG. 57B). These analog signals are then low-pass filtered, sequentially routed through an analog multiplexer (MUX, for example, CD74HC4067 by Texas Instruments®), and digitized by the built-in 12-bit analog-to-digital converter (ADC) of a microcontroller unit (MCU for example, RP2040 by Raspberry Pi, as shown in FIG. 58A and FIG. 58B). In some embodiments, the MCU streams time-stamped data frames from all 48 channels synchronously at 100 Hz to a host computer for realtime processing.

[0461]

[0421] In some embodiments, the computing device (for example, the computing device 5006 shown in FIG. 50F) is communicatively coupled to the sensor array (for example, the sensor array 5004 shown in FIG. 50F) and configured to generate a spatiotemporal tissue morph map based on the plurality of optical sensing signals (for example, step 7802 of the example method 7800 shown in FIG. 78), and generate at least one predicted tissue morph data object based at least in part on the spatiotemporal tissue morph map (for example, step 7804 of the example method 7800 shown in FIG. 78), additional details of which are described herein.

[0462] Example Spatially Resolved Tracing of Finger-Specific Muscle Activity

[0463]

[0422] To validate TRAiLL’s ability to resolve fine motor control, example embodiments of the present disclosure targeted the deep and superficial flexor muscles of the forearm, which are responsible for individual finger flexion. The underlying principle relies on detecting morphological changes in these muscles during contraction. As a finger flexes, the corresponding muscle belly thickens, altering the tissue density and optical path of back-scattered near-infrared light, a phenomenon corroborated by Monte-Carlo simulations of photon scattering (for example, as shown in FIG. 51 A, FIG. 59A and FIG. 59B). These simulations confirmed that tissue compression during muscle flexion enhances light propagation toward the photodiode, increasing the received signal intensity. This effect is twofold: the increased density of the flexed muscle belly enhances photon scattering (for example, as shown in FIG. 60A to FIG. 60B), while the concurrent thinning of the overlying skin and subcutaneous fat shortens the mean optical path length to the detector (for example, as shown in FIG. 61A to FIG. 61C). The primary muscles contributing to these signals are the flexor digitorum profundus (FDP), flexor digitorum superficialis (FDS), and flexor pollicis longus (FPL), located in the volar forearm (for example, as shown in FIG. 5 IB). The FDS primarily flexes the middle phalanges of the four fingers, while the deeper FDP is responsible for flexing their distal phalanges. The FPL performs the analogous function for the thumb, flexing its interphalangeal joint. Together, these muscles enable the fine motor control necessary for individuated finger and thumb movements.

[0464]

[0423] To directly correlate TRAiLL’s optical signals with physiological events, example embodiments of the present disclosure utilize simultaneous B-mode ultrasound imaging during isometric flexion of individual fingers (for example, as shown in FIG. 51C to FIG. 5 IF). The spatiotemporal heat maps generated by TRAiLL revealed focal hemodynamic changes that spatially and temporally corresponded with the contraction of specific tendon-muscle junctions identified in the ultrasound frames. For instance, flexion of the index finger produced a distinct activation pattern localized to the index FDS and FDP (for example, area 5117 and area 5121 as shown in the index finger portion of FIG. 51C), whereas thumb flexion engaged the FPL (for example, area 5113 as shown in the thumb portion of FIG. 51C). In contrast, the device registered minimal signal change during periods of rest, confirming the signals are directly related to muscle activity (for example, as shown in FIG. 5 IE and FIG. 5 IF).

[0465]

[0424] Analysis of the temporal dynamics from individual detector channels revealed that flexion of each finger produced a unique and repeatable temporal signature (for example, as shown in FIG. 51G and FIG. 51H). For example, channel (4, 3) showed the highest response to index finger flexion, while channel (7, 2) was most sensitive to motion of the ring finger. This fingerspecific channel response scaled proportionally with the level of effort, with the photovoltage from the maximally responsive channel for each finger increasing systematically from rest (0%) to 50% and 100% of maximal voluntary contraction (for example, as shown in FIG. 511). When averaged across multiple contractions (n=5), the normalized signal trajectories for each finger followed stereotyped, finger-specific time courses through the flexion-extension phase, indicating high repeatability (for example, as shown in FIG. 51 J and FIG. 62).

[0466] [4251 To quantify the classification potential of these optical sensing signals, various embodiments of the present disclosure apply linear discriminant analysis (LDA) to a dataset comprising 3000 single-frame feature vectors from three independent subjects (for example, as shown in FIG. 63 to FIG. 67). The two-dimensional projection of these feature vectors showed five well-separated clusters, one for each finger, with 95% Mahalanobis confidence ellipses demonstrating minimal overlap (for example, as shown in FIG. 5 IK). This separability confirms that TRAiLL can robustly differentiate the subtle muscle activation patterns associated with individual finger movements, providing a reliable basis for high-resolution gesture interpretation.

[0467] Example Linear Discriminant Analysis for Finger Gesture Classification

[0468]

[0426] To quantitatively assess the separability of the high-dimensional signals generated by the TRAiLL sensor during discrete finger movements, various embodiments of the present disclosure utilize Linear Discriminant Analysis (LDA).

[0469]

[0427] LDA is particularly well-suited for this task as it is a powerful supervised learning method for dimensionality reduction and classification of high-dimensional data. The method seeks to find a linear combination of features — for example, a projection from the 48-channel sensor space to a lower-dimensional space (2D for visualization) in accordance with some embodiments of the present disclosure — that maximizes the separation between the means of different classes. For example, various embodiments of the present disclosure maximize the ratio of the between-class variance (how far apart the class means are) to the within-class variance (how spread out each individual class is). By projecting the data onto axes that best separate the finger gestures, LDA provides a low-dimensional visualization that inherently emphasizes the discriminatory information within the temporal signal, making it an excellent tool for assessing class separability, particularly in the context of biosignal-based interfaces.

[0470]

[0428] In some embodiments, the dataset used for this analysis is compiled from recordings of three healthy subjects, each performing five repetitions of five discrete, isometric finger flexion gestures (thumb, index, middle, ring, and pinky). In some embodiments, to create a single, representative data point for each gesture instance, a feature extraction process was applied to the continuous 48-channel time-series data. First, the data stream was segmented into individual gesture repetitions based on the event labels recorded during the experiment. For each segment, the point of maximum muscle exertion was identified. For example, various embodiments of the present disclosure utilize the L2 norm, L(t), of the 48-channel sensor vector, s(t), at each time point, t:

[0471]

[0472]

[0429] This calculation yields a one-dimensional time series representing the total signal magnitude. The single time point, tpeak, corresponding to the maximum value of this L2 norm trace was identified as the moment of “peak activation”. The 48 raw sensor values at this specific time point, s(tpeak), were then extracted to form the 48-dimensional feature vector for that gesture repetition. This process resulted in a final dataset of 75 feature vectors (3 subjects x 5 fingers x 5 repetitions), with each vector representing the sensor snapshot at the peak of a specific finger movement.

[0473]

[0430] In some embodiments, the LDA model is implemented using the LinearDiscriminantAnalysis class from the scikit-learn library in Python. For example, the model was trained (or “fitted”) on the data from a single subject (Subject 1), which consisted of 25 feature vectors (5 fingers x 5 repetitions). The goal is to find a transformation matrix, W, that projects the original 48-dimensional data, X, onto a lower-dimensional space, Y, such that class separability is maximized.

[0474] Y = WTX

[0475]

[0431] For visualization, the number of components was set to two, resulting in a projection onto a 2D plane. LDA computes the within-class scatter matrix (Sw) and the between-class scatter matrix (Sg):

[0476]

[0477]

[0432] Here, C is the number of classes (5 fingers), Dcis the set of samples for class c, μcis the mean of class c, Ncis the number of samples in class c, and μ is the overall mean. The optimal transformation matrix, W, is composed of the eigenvectors corresponding to the largest eigenvalues of the matrix Sw-1SB. Once the model was fitted, this single transformation matrix was applied to the entire dataset of 75 feature vectors from all three subjects to project them onto the same 2D plane for comparative visualization.

[0478]

[0433] The resulting 2D projections of all 75 data points are displayed as a scatter plot in FIG.

[0479] 5 IK. Each point represents a single gesture instance, with its color indicating the finger class and its marker shape denoting the subject. To statistically quantify the distribution of these clusters, various embodiments of the present disclosure calculate 95% Mahalanobis confidence ellipses for each finger class, aggregating the data from all three subjects. The equation for a Mahalanobis ellipse is given by:

[0480]

[0481] where x is a point on the ellipse, μcis the mean vector, and Xcis the covariance matrix of the class c in the 2D projected space. The value on the right-hand side is the chi-squared statistic for the desired confidence level (p=0.95) and degrees of freedom (df' 2). The minimal overlap observed between these confidence ellipses provides strong statistical evidence that the muscle activation patterns for individual finger movements, as captured by TRAiLL, are highly distinct and robust across different subjects.

[0482] Example Quantification of Muscle Load and Fatigue in Antagonistic Pairs

[0483]

[0434] In some embodiments, the at least one predicted tissue morph data object generated by an example wearable muscle monitor comprises a muscle activation factor (MAF).

[0484]

[0435] For example, to evaluate TRAiLL’ s utility in monitoring gross motor movements and physiological states relevant to rehabilitation and exercise science, various embodiment of the present disclosure measure the activity of antagonistic arm muscles during standard resistance exercises. For example, a single TRAiLL patch is positioned over the mid-humerus to simultaneously capture signals from the biceps brachii during elbow flexion (curls) and the triceps brachii during elbow extension (kickbacks) (for example, as shown in FIG. 52A and FIG. 52D).

[0485]

[0436] To quantify the intensity of muscle engagement from the optical signals, various embodiments of the present disclosure develop a normalized MAF. This factor is calculated by first taking the L2 norm of the baseline-corrected, low-pass filtered signals, Si(t) (for example, as shown in FIG. 68A to FIG. 69B) across all N sensor channels: L(t) =

[0486]

[0487] ■ This aggregate signal is then normalized, p(t) = clip(-^, 0,1), and passed through a soft bilateral R

[0488] logarithmic transform, defined as:

[0489]

[0490] where the constants R, fc, and p0are empirically chosen to normalize the output between 0 and 1 while enhancing sensitivity to both weak and strong muscle contractions. This straightforward, one-time calibration for a given user and sensor location ensures robust performance and does not impede the system’s broad, long-term applicability. In some embodiments, the MAF reflects the principle of motor unit recruitment: to generate greater force, the nervous system recruits an increasing number of motor units. This widespread fiber contraction leads to larger morphological and hemodynamic changes in the muscle tissue, which in turn increases the backscattering of nearinfrared light detected by TRAiLL and results in a higher MAF value.

[0491]

[0437] During sets of consecutive curls and kickbacks, the MAF showed a clear loaddependent response, with significantly higher activation when lifting a 5-lb load compared to unloaded movements (for example, as shown in FIG. 52B and FIG. 52E). This reflects greater recruitment of motor units to meet the increased force demand, resulting in a larger L2 norm of the sensor signals. Beyond simple load detection, the temporal dynamics of the MAF during 60-second sustained isometric holds revealed distinct neuromuscular strategies. At the onset of contraction, both exercises exhibited a characteristic overshot in the MAF, where a large burst of motor unit activity is required to overcome inertia and stabilize the joint before settling into a steady hold (for example, as shown in FIG. 52C and FIG. 52F). Following this initial phase, the loaded bicep curl, a biomechanically stable posture, exhibited a gradual decay to approximately 70% of its peak value, consistent with classic neuromuscular fatigue, defined as a progressive, exercise-induced reduction in the ability of a muscle to generate force (for example, as shown in FIG. 52C). In contrast, the loaded triceps kickback, a less stable posture that subjects reported as more difficult to maintain, showed a more complex profile: after an initial decay to -50%, the MAF began to increase again (for example, as shown in FIG. 52F). This secondary rise suggests a

[0492] 'll compensatory recruitment strategy; as the primary triceps motor units’ fatigue, the nervous system engages additional, less-efficient synergist muscles to maintain the joint angle, thereby increasing the total muscle activity captured by the sensor array and elevating the MAP. The ability to continuously track such detailed neuromuscular dynamics opens possibilities for objectively quantifying exercise quality, optimizing rehabilitation protocols by monitoring muscle fatigue in real-time, and developing adaptive training programs that adjust intensity based on physiological feedback.

[0493]

[0438] Various embodiments of the present disclosure illustrate TRAiLL’s sensitivity to subtle biomechanical changes by comparing a thumb-over grip (TOG) with a thumb-under grip (TUG) during a seated concentration curl protocol (for example, as shown in FIG. 52G, FIG. 521, FIG.

[0494] 70A and FIG. 70B). This comparison was designed to test whether subtle differences in thumb orientation, known to significantly alter muscle recruitment patterns during gripping tasks, could be detected by TRAiLL. The MAF traces and corresponding heat maps revealed that the TUG elicited a substantially higher and more spatially widespread activation of the biceps compared to the TOG (for example, as shown in FIG. 52H). This difference arises from the dual function of the biceps brachii as both an elbow flexor and a powerful forearm supinator. The thumb-under grip enforces a greater degree of forearm supination, a position where the bicep has optimal leverage for flexion. Consequently, maintaining this posture against the dumbbell’s torque while simultaneously performing the curl requires a greater overall recruitment of motor units within the biceps compared to the less supinated thumb-over grip, demonstrating that TRAiLL can resolve fine-grained differences in muscle engagement arising from minor variations in grip posture.

[0495]

[0439] Beyond analyzing discrete, high-intensity exercises, the skin-conformal design and long-term comfort of the TRAiLL patch facilitate continuous, ambulatory monitoring of muscle dynamics in real-world settings. To demonstrate this, a TRAiLL patch was worn on the biceps for three consecutive days to track activation patterns across a wide range of daily activities (for example, as shown in FIG. 52J). The MAF successfully differentiated between high-exertion tasks (such as working with and without a load), different stages of aerobic exercise, and different periods of sleep. By establishing a fatigue-threshold MAF, this continuous data stream offers the potential to quantify the accumulation of muscle fatigue throughout the day (as shown in the shade areas above the threshold lines). This capability to unobtrusively monitor physiological states over extended periods is crucial for applications in personalized rehabilitation, where understanding the impact of daily routines on recovery is paramount, and for developing adaptive interfaces that respond to a user’s changing physiological capacity.

[0496] Example Self-Supervised Learning for American Sign Language Classification

[0497]

[0440] In some embodiments, the at least one predicted tissue morph data object generated by an example wearable muscle monitor comprises a predicted gesture classification. For example, example embodiments of the present application may generate the predicted gesture classification based on inputting the plurality of optical sensing signals to a trained machine learning model.

[0498]

[0441] For example, to demonstrate TRAiLL’s utility in complex human-computer interaction, various embodiments of the present disclosure provide a pipeline to classify 24 static gestures of the American Sign Language alphabet (ASL-A) from optical myography data. For example, various embodiments of the present disclosure implement a self-supervised learning (SSL) approach, as it enables a model to learn robust feature representations from complex, unlabeled bio-signals, thereby minimizing the need for extensive manual data annotation. The three-phase SSL pipeline began with a denoising autoencoder. In this phase, an encoder network compressed the high-dimensional input data (48 channels x 512 temporal samples) into a lowdimensional latent space, and a decoder network reconstructed the original data from this compressed representation (for example, as shown in FIG. 53A). This phase effectively removed high-frequency noise while preserving the structural fidelity of the underlying physiological signals, as shown by comparing representative original (for example, as shown in FIG. 53D) and reconstructed (for example, as shown in FIG. 53E) heat maps and raw data (for example, as shown in FIG. 73 A to FIG. 73D). The model demonstrated efficient learning, as the training loss exhibited an exponential decay that plateaued within approximately 20 epochs, resulting in a final reconstruction accuracy exceeding 70%; for this phase, accuracy was defined as the percentage of data points reconstructed with less than a 2% error relative to a pre-constructed low-noise baseline (for example, as shown in FIG. 53F). Fourier analysis confirmed that the autoencoder selectively suppressed high-frequency noise above 20 Hz while preserving the low-frequency content characteristic of muscle activity (for example, as shown in FIG. 53G).

[0499]

[0442] In the second phase, various embodiments of the present disclosure utilize a contrastive learning model to refine the latent space. The trained encoder from Phase I was used to generate embeddings for augmented versions of the input data, including random translations, rotations, and mirroring (for example, as shown in FIG. 53B and FIG. 74A to FIG. 74C). The model was then trained in a contrastive pretext task, where it learned to differentiate between “positive pairs” (i.e., two different augmentations originating from the same gesture sample) and “negative pairs” (augmentations from different gesture samples). Using a contrastive loss function, the training objective is to pull the latent space representations of positive pairs closer together while pushing the representations of negative pairs farther apart. A t-SNE visualization of 6,500 augmented embeddings revealed well-separated, class-distinct clusters, indicating that the model successfully learned augmentation-invariant features for each ASL gesture (for example, as shown in FIG.

[0500] 53H). This augmentation invariance was further confirmed by high intra-class and low inter-class cosine similarity between embeddings (for example, as shown in FIG. 75 A and FIG. 75B). Moreover, the latent representation of any augmented view could accurately reconstruct the original, canonical signal, demonstrating that the encoder captured the core gesture features irrespective of the transformation (for example, as shown in FIG. 75C). The learned weights of the encoder show a non-uniform contribution from the input channels, highlighting the model’s ability to identify the most informative sensor locations for this task (for example, as shown in FIG. 531).

[0501]

[0443] Finally, in Phase III, the pretrained encoder was frozen and its latent representations were fed into a multilayer perceptron (MLP) for the downstream classification task (for example, as shown in FIG. 53 C). The complete SSL pipeline achieved a five-fold cross-validation accuracy of 92%, a significant improvement over a naive MLP trained on the raw data (69%) (for example, as shown in FIG. 53 J). The model’s high performance was further evidenced by the receiver-operating-characteristic (ROC) curves, which yielded an area-under-the-curve (AUC) greater than 0.98 for all 24 classes (for example, as shown in FIG. 53K). The resulting normalized confusion matrix showed strong diagonal dominance, with minimal off-target misclassifications, confirming the model’s robustness and high accuracy for ASL-A gesture recognition (for example, as shown in FIG. 53L).

[0502] Example Bimanual Musical Performance and 3D Gesture Control

[0503]

[0444] To explore the system’s potential for more dynamic and complex human-computer interfaces, the present disclosure demonstrated the system’s use in applications such as i) bimanual control of a virtual musical instrument, and ii) camera-free gesture control of a 3D object in an augmented reality environment. The first application, bimanual control of a virtual musical instrument, addresses the long-standing challenge of creating expressive, unencumbered digital musical instruments. This application opens an avenue for accurate musical expression in comparison with other camera-based gesture instruments, which often suffer from occlusion, limited tracking accuracy under variable lighting, and high computational overhead, constraining their expressivity and reliability in performance contexts, particularly for individuals who may lack the finger strength or dexterity required for traditional instruments or performing standardized gestures, by translating gross motor movements of the forearms into complex musical output.

[0504]

[0445] As an example, a volunteer was fitted with two TRAiLL patches, with one on the left forearm to decode chord-fretting gestures and another on the right wrist to detect strumming motions, enabling the volunteer to play a virtual guitar synthesized on a laptop (for example, as shown in FIG. 54A and FIG. 54B). While other wearable sensors have been explored for musical expression, they face significant challenges in bimanual applications. Systems based on electromyography (EMG) are susceptible to motion artifacts and often require heterogeneous sensor setups for two-handed control. Data gloves, while offering high fidelity, can be physically cumbersome and restrict the natural dexterity needed for playing. Vision-based approaches, conversely, are vulnerable to line-of-sight and occlusion issues, where one hand can block the camera’s view of the other. The raw photovoltage traces recorded during a four-chord progression (C-F-G-C) and a standard finger-picking pattern revealed distinct, highly repeatable spatiotemporal signatures for each discrete action (for example, as shown in FIG. 54C, FIG. 54D, FIG. 76A and FIG. 76B). Critically, the signals from the two limbs were captured simultaneously without any apparent interference, confirming the system’s capacity for robust, bimanual read-out.

[0505]

[0446] In the second application, a single TRAiLL patch on the forearm was used to control a virtual 3D object via six distinct hand gestures (for example, as shown in FIG. 54E). A continuous 50-second recording session showed discrete bursts of muscle activity in the channel-wise heat map, corresponding to the performance of individual gestures (for example, as shown in FIG. 54F). A pretrained classification model accurately identified and time-stamped each “grab,” “pinch,” “wave,” “trigger,” “fist,” and “thumbs-up” event in real-time (for example, as shown in FIG. 54G). The six-class model was trained efficiently, reaching convergence and a validation accuracy of 96% within 30 epochs (for example, as shown in FIG. 54H). The model’s robustness was confirmed on an independent test set of 720 gestures, where it achieved near-perfect diagonal dominance in the normalized confusion matrix, indicating minimal misclassification between the different gestures (for example, as shown in FIG. 541).

[0506] Additional Details of Example TRAiLL System

[0507]

[0447] The TRAiLL system introduced here leverages a wearable optical myography mapping and a self-supervised learning framework to continuously trace subcutaneous muscle dynamics for a broad range of applications. By directly imaging the physical and hemodynamic changes in contracting muscles, TRAiLL provides a high-resolution spatiotemporal map that circumvents key limitations of established technologies; it avoids the need for conductive gels and the susceptibility to electrical crosstalk common in EMG, and offers a more practical solution for continuous, ambulatory monitoring than complex ultrasound imaging. The present disclosure demonstrated versatility across three key applications that span physiological monitoring and human-computer interaction. First, in a rehabilitation context, the system successfully quantified complex physiological states, such as load-dependent activation and neuromuscular fatigue during resistance exercises, offering objective insights for personalizing physical therapy. Second, for assistive communication, it translated high-dimensional data into accurate, real-time classifications of 24 American Sign Language gestures with high fidelity. Finally, it enabled intuitive virtual interaction by decoding motor intent for the bimanual control of a virtual musical instrument and camera-free manipulation of 3D objects, highlighting its potential to replace cumbersome handheld controllers.

[0508]

[0448] The TRAiLL system overcomes many of the limitations of existing wearable technologies, offering a gel-free, motion-robust, and unobtrusive solution that is independent of external infrastructure like cameras. This work represents a significant step toward the development of practical, long-term wearable sensors capable of seamlessly integrating with the human body to both monitor health and augment our ability to interact with the digital world. Future work will focus on the clinical validation of these metrics in rehabilitation settings, expanding the gesture library for more complex control schemes, and engineering a fully wireless, embedded version of the system for real-world, out-of-lab deployment.

[0509] i. TRAiLL Device Fabrication and System Integration

[0449] The TRAiLL system consists of three main components: the flexible TRAiLL sensor array (for example, as shown in FIG. 55 A to FIG. 551), a flexible TRAiLL adaptor for signal conditioning (for example, as shown in FIG. 56A to FIG. 56H), and a rigid TRAiLL Pico board for data processing and control (for example, as shown in FIG. 57A to FIG. 57C). The sensor array is constructed on a thin, flexible substrate designed for skin-conformal contact. The flexible printed circuits (FPCs) are designed in Altium Designer (v24.5.1) and fabricated to IPC-6013 Class 2 standards (ILCPCB) using their transparent polyethylene terephthalate (PET) protocol, resulting in a 240 μm thick, optically clear base. Each sensor array houses an 8x6 array of sensing pixels. For each pixel, a near-infrared micro-LED (SFH 4043, OSRAM, X = 950 nm) and a silicon photodiode (VEMD 1060X01, Vishay) are surface-mounted with a center-to-center separation of 3.5 mm. The assembly is performed using a standard reflow process with a low-temperature solder paste to protect the flexible substrate. Following assembly, the entire electronic array is encapsulated in a -500 pm layer of biocompatible, soft silicone (Smooth-On Ecoflex 00-30), which is degassed prior to curing to ensure optical clarity and mechanical robustness. The final device is mounted onto the skin using a 3M Tegaderm transparent film dressing.

[0510]

[0450] The sensor array interfaces via a flexible flat cable (FFC) with the flexible TRAiLL adaptor, which is responsible for signal conditioning. On the adaptor, photocurrent signals from the 48 photodiodes are first converted to voltage and amplified by a 48-channel transimpedance amplifier (TIA). The analog signals are then routed through analog multiplexers (for example, model CD74HC4067 by Texas Instruments). The adaptor board connects to the rigid TRAiLL Pico board, which houses a Raspberry Pi Pico microcontroller unit (MCU). The MCU uses its built-in 12-bit analog-to-digital converters (ADCs) to sample the incoming signals sequentially at an effective rate of 100 Hz per channel. The MCU also controls the MUX switching, modulates the LED intensity using pulse-width modulation (PWM), and streams the digitized, time-stamped data to a host computer via a USB serial connection. The entire system is powered via the USB connection, with a typical power consumption of less than 150 mW during operation (for example, as shown in FIG. 77).

[0511] j. Example Applications of Wearable Muscle Monitors

[0512] i. Participant Preparation for Applications

[0451] Prior to each data collection session, the skin at the target location of example particants of applications was cleaned with a 70% isopropyl alcohol wipe to ensure a clear and stable optical interface for the sensor array.

[0513] ii. Example Application Protocols

[0514] 1. ASL-A Alphabet and Finger Tracing Test

[0515]

[0452] Sensor Placement: A single TRAiLL patch was positioned over the volar forearm, centered approximately one-third of the distance from the wrist to the elbow (for example, as shown in FIG. 56A and FIG. 56B). This placement was chosen to optimally target the flexor digitorum profundus (FDP) and superficialis (FDS) muscles, which are critical for finger and hand articulation.

[0516]

[0453] Protocol: Participants were seated comfortably with their forearm resting on a table. For the ASL-A data collection, participants were cued via a computer screen to perform a series of 24 static ASL-A gestures, with each gesture repeated 5 times. Each repetition was held for a duration of 1 second, followed by a 2-second rest period. For the finger tracing protocol, participants were instructed to perform isometric flexion of each individual finger against a fixed surface to determine their Maximal Voluntary Contraction (MVC).

[0517] 2. Exercise and Grip Style Test

[0518]

[0454] Sensor Placement: A single TRAiLL patch was centered over the mid-humerus to simultaneously capture activity from the biceps brachii (anteriorly) and triceps brachii (posteriorly) (for example, as shown in FIG. 52A, FIG. 52D, FIG. 56C, and FIG. 56D).

[0519]

[0455] Protocol: The experimental session was divided into three sequential blocks: a dynamic test, a static test, and a grip style comparison.

[0520]

[0456] Dynamic Test: Participants first performed one set of five repetitions of elbow flexion (bicep curls), followed by one set of five repetitions of elbow extension (triceps kickbacks). This sequence was initially performed without a load and was then repeated with a 5 lb. dumbbell after a 5-minute break. Each repetition consisted of a 1 -second concentric / eccentric motion, followed by a 1 -second rest.

[0457] Static Test: Participants held a single isometric contraction with a 25 lb. dumbbell in the elbow flexion (curl) position until task failure, defined as the point when the participant could no longer maintain the correct posture. The same procedure was repeated for the elbow extension (kickback) position.

[0521]

[0458] Grip Style Test: Participants performed one set of five repetitions of concentration curls using the dynamic test timing protocol with two distinct grips: (i) a thumb-under grip (TUG), the standard grip where the thumb wraps around the handle to oppose the fingers, creating a stable, supinated forearm posture; and (ii) a thumb-over grip (TOG), where the thumb is positioned on the same side of the handle as the fingers, creating a less stable “thumbless” grip.

[0522] 3. Three-Day Continuous Monitoring

[0523]

[0459] Sensor Placement: A single TRAiLL patch was affixed to the skin overlying the belly of the biceps brachii using a biocompatible, medical-grade double-sided adhesive.

[0524]

[0460] Protocol: The participant provided informed consent for the study. The sensor was connected to a lightweight, portable data acquisition unit that continuously recorded optical signals for approximately 72 hours. During this period, the participants were instructed to perform their normal daily routines without restriction, which included periods of desk work, structured aerobic exercise, and sleep. The participant also maintained a time-stamped activity log to enable the correlation of specific events with the recorded MAF data (for example, FIG. 52J).

[0525] 4. Virtual Guitar Test

[0526]

[0461] Sensor Placement: Two TRAiLL patches were used for bimanual control. The first was placed on the left volar forearm to sense chord-fretting gestures. The second was placed on the right wrist to detect strumming and finger-picking motions.

[0527]

[0462] Protocol: The participants were seated in a chair with their forearms resting on the armrests to ensure a relaxed posture (for example, FIG. 54B). The task involved playing a sequence of single musical notes. Each note was produced by the synchronous combination of a left-hand gesture and a right-hand finger-picking motion. Participants were instructed to first form one of the four target chord gestures (ASL-A signs for ‘C’, ‘F’, or ‘G’) with their left hand, and then, while holding the sign, perform a single finger-pick with their right hand to trigger the note. The specific sequence performed was a C-F-G-C chord progression, where each chord was triggered by one finger-pick from a repetitive “t-i-m-r” (thumb, index, middle, ring) pattern. The classified bimanual signals were mapped in real-time to a software synthesizer to produce the corresponding musical audio.

[0528] 5. VR Command Test

[0529]

[0463] Sensor Placement: A single TRAiLL sensor array was positioned on a participant’s dominant forearm, in the same location as the ASL-A test.

[0530]

[0464] Protocol: Participants were instructed to perform six distinct hand gestures used for virtual or augmented reality control: grab, pinch, wave, trigger, fist, and thumbs-up. Each gesture was cued on a screen and performed multiple times in a continuous 50-second session to test the real-time classification accuracy of the system.

[0531] k. Data Acquisition and Signal Processing

[0532] i. Real-time Data Acquisition and Visualization

[0533]

[0465] Data were acquired from the TRAiLL Pico board using a custom Python (3.12.9) script. The script launched three parallel processes using Python’s multiprocessing library to ensure nonblocking operation: one for reading the serial data stream, one for saving data to disk, and a main process for real-time visualization. Communication with the device was established using the pyserial (3.5) library. The serial process continuously read incoming ASCII data, which was formatted as a 6x8 matrix of integer values per time point, and parsed it into a NumPy array. A baseline correction was immediately applied in real-time by subtracting the first complete 48-channel data sample from all subsequent samples. This baseline-corrected data was then passed to the saving and visualization processes via separate multiprocessing queues. A dedicated saving process ran concurrently to log the data without interrupting acquisition. For each session, a unique CSV file was created with a timestamp in the filename (e.g., test-data- YYYY-MM-DD-HH-MM-SS.csv). The saving process continuously retrieved data from its queue and appended it to this file. Each row in the CSV corresponds to a single time point and contains a high-precision timestamp, a text label for the current gesture status (e.g., ‘open’, ‘fist’), and the 48 flattened, baseline-corrected sensor values. For real-time monitoring, a visualization process used “matplotlib. animation” to display the 8x6 sensor array data as a heat map, which was interpolated using “scipy. ndimage. zoom” for smoother visuals.

[0534] ii. Data Preparation for Denoising

[0535]

[0466] All subsequent analyses were performed offline on the saved CSV files to prepare the data for the SSL training pipeline. The first step was to create paired noisy input and clean target signals for the denoising autoencoder. The baseline-corrected data, as saved in the CSV files, served as the “noisy” input. To create the corresponding “clean” target signal, the physiological signal was isolated from noise using a second-order Butterworth low-pass filter. This filter type was chosen for its maximally flat frequency response in the passband, ensuring that the underlying muscle activity signals are not distorted. A cutoff frequency of 5 Hz was selected, as most of the signal power from voluntary muscle contractions resides below this frequency, while this effectively attenuates high-frequency noise from electronic sources and motion artifacts. The filtering was implemented using a zero-phase filtering approach (scipy. signal. sosfiltfilt), which processes the signal in both forward and reverse directions to eliminate phase shifts that would otherwise distort the temporal relationship between the signal and the experimental events.

[0536] iii. Data Augmentation and Feature Preparation for Contrastive Learning

[0467] For the second phase of SSL training, the “noisy” input data was first passed through the trained encoder from the denoising stage to generate low-dimensional latent space vectors. To create robust, invariant features, a set of data augmentations was applied to these vectors before they were used in the contrastive learning framework. Three types of transformations were applied probabilistically: random translation, which shifts the feature map along both spatial axes by a random offset; random rotation, which rotates the feature map by an angle between -10 and 10 degrees; and random horizontal and vertical mirroring. This augmentation strategy creates varied “views” of each gesture, which are essential for the contrastive task of learning to identify different augmentations of the same underlying gesture while distinguishing them from others. This two-step process ensures the contrastive learning phase operates on a robust and meaningful representation of muscle activity.

[0537] 1. Feature Engineering and Statistical Analysis i. Statistical Methods

[0538]

[0468] All statistical analyses were performed using Python’s scipy (1.15.2) and numpy (2.2.4) libraries. Descriptive statistics, including the mean and standard error of the mean (s.e.m.), were used to summarize time-series data across multiple trials. The shaded error bands accompanying time-series plots in the figures represent the s.e.m. calculated at each time point across all repetitions and / or channels of a given task. For classification results, accuracy and standard deviation (std) were calculated across the five folds of the cross-validation procedure.

[0539]

[0469] For the analysis presented in FIG. 511, the “maximally responsive channel” for each finger was identified programmatically. For each finger-specific dataset (e.g., all trials corresponding to “thumb” flexion), the variance of the signal was calculated for each of the 48 channels across all time points. The channel exhibiting the highest variance was selected as the maximally responsive channel for that specific finger, as this indicates the sensor location that captures the greatest dynamic range of signal change during the movement.

[0540] ii. Dimensionality Reduction and Feature Analysis

[0541]

[0470] To quantitatively assess the class separability of the muscle activation patterns for the analysis presented in FIG. 5 IK, Linear Discriminant Analysis (LDA) was performed using the LinearDiscriminantAnalysis implementation in the scikit-learn library (for example, based on the linear discriminant analysis for finger gesture classification). The input dataset for the LDA was constructed by extracting a single, representative feature vector for each gesture repetition. For each repetition, the L2 norm of the 48-channel sensor signal was calculated at every time point to find the point of maximum muscle activation. The 48 pre-processed sensor values at this single peak time point were then used as the feature vector for that repetition. This resulted in a total dataset of 75 feature vectors (3 subjects x 5 fingersx5 repetitions).

[0542] m. Model Training Pipeline

[0543] i. Dataset Curation

[0544]

[0471] The dataset for the ASL-A classification task was constructed from the recordings of 3 subjects performing 24 static gestures, each repeated 5 times, resulting in a total of 360 gesture instances. Each instance was epoched into a 48-channelx512-temporal sample array. For training and evaluation, the dataset was split into training (80%) and testing (20%) sets, ensuring that all repetitions from a single subject were kept within the same set to prevent data leakage. No explicit data balancing techniques were employed due to the uniform number of repetitions for each class.

[0545] ii. Phase I: Denoising Autoencoder Architecture

[0546]

[0472] In some embodiments, the trained machine learning model comprises an encoder network (such as autoencoder). For example, the autoencoder is a fully convolutional neural network designed for time-series data. The encoder consists of five sequential ID convolutional blocks that progressively downsample the temporal dimension. The number of filters in these layers are 128, 64, 32, 16, and 32, respectively. Each convolutional layer uses a kernel size of 3, a stride of 2, and is followed by a Rectified Linear Unit (ReLU) activation function. This compresses the input into a low-dimensional latent space vector of size 32x16. The decoder mirrors this architecture using ID transposed convolutional layers to upsample the latent vector back to the original dimensions. Crucially, the decoder utilizes skip connections, concatenating the output of each upsampling layer with the corresponding feature map from the encoder. This allows the network to retain fine-grained temporal information during reconstruction. The final layer of the decoder uses a Sigmoid activation function to map the output to the [0, 1] range, consistent with the min-max normalization applied to the input data.

[0547] iii. Phase II: Contrastive Learning Framework

[0548]

[0473] In some embodiments, the trained machine learning model comprises a contrastive learning model. For example, the contrastive learning phase utilizes a SimSiam framework to refine the feature representations learned by the encoder. This approach is designed to maximize the similarity between two distinct, stochastically augmented views of the same NIRS signal segment, a process that does not require the use of explicit negative sample pairs.

[0549]

[0474] The core architecture consists of the pretrained encoder ( / ), a projection head (g), and a prediction head (q). For two augmented views, x1and x2, the model generates latent representations h = (Xi) and h2= f (x2). These are then passed through the projection head to create embeddings zx= 5r( / i1) and z2= ^( / i2). One branch’s embedding, zr, is further transformed by the prediction head into pr— q^z^. The model was trained to minimize the negative cosine similarity between the prediction from one branch and the projection from the other (for example, based on a contrastive learning framework described herein). A crucial “stop- gradient” operation was applied to the second branch’s projection (z2) to prevent the model from converging to a trivial solution. This objective forces the model to learn consistent and discriminative embeddings for different augmentations of the same underlying signal, making them suitable for downstream classification tasks.

[0550]

[0475] To learn robust and discriminative features from the NIRS signals, a contrastive selfsupervised learning framework was employed. This approach is designed to create an embedding space where augmented versions of the same signal are grouped closely together, while signals from different underlying gestures are pushed apart. The model’s core architecture is initialized using the pretrained encoder from the denoising autoencoder stage, leveraging the foundational spectral-temporal patterns already learned. The encoder is frozen to retain these general features after being trained for the specific contrastive task. It maps each input signal to a 32-dimensional latent vector, which is subsequently processed by two randomly initialized, two-layer multilayer perceptrons (MLPs): a projection head that maps it to a 64-dimensional embedding space, and a prediction head used for the final loss calculation.

[0551]

[0476] To generate the paired data views essential for contrastive learning, each raw NIRS signal segment is stochastically augmented to produce two distinct, yet correlated, versions. This process simulates realistic data variability and includes a suite of transformations to enhance the model’s invariance to common signal perturbations. Geometric operations such as random temporal translation, rotation, and mirroring are applied to account for shifts in gesture timing and sensor orientation. Concurrently, signal-level perturbations are introduced through the addition of Gaussian noise, random channel dropout, and frequency-domain masking.

[0552]

[0477] The optimization follows the SimSiam contrastive learning paradigm, which aims to maximize the similarity between the two augmented views of the same NIRS signal without requiring explicit negative sample pairs. Let x represent a raw NIRS sequence and let tx(-) and t2(•) be two stochastic augmentation functions. The encoder, f0, maps each augmented view into a latent representation:

[0553]

[0554]

[0478] These latent vectors are then passed through a projection head, g^, and a prediction head, q^

[0555]

[0556]

[0479] The training objective is to minimize the negative cosine similarity loss between the prediction from one branch and the projection from the other. A crucial “stop-gradient” operation is applied to prevent the model from converging to a trivial solution. The final loss function, £, is defined as:

[0557]

[0558]

[0480] This approach enforces representational consistency across augmented views, yielding compact and discriminative embeddings suitable for downstream classification tasks.

[0559]

[0481] The quality of the representations learned was assessed through three methods (for example, as shown in FIG. 53H as well as FIG. 73A to FIG. 73D).

[0560]

[0482] First, the cosine similarity between two embedding vectors, ztand Zj, is calculated as their dot product divided by the product of their magnitudes (for example, as shown in FIG. 73A and FIG. 73B):

[0561]

[0562]

[0483] A value close to 1 indicates high similarity, while a value close to 0 indicates low similarity. This metric was used to compare positive pairs (two augmented views of the same NIRS segment) against negative pairs (views from different segments). A significantly higher level of similarity for positive pairs confirms the model learned to group correlated signals effectively.

[0563]

[0484] Second, t-SNE is used to project the high-dimensional embeddings into a 2D space for qualitative assessment (for example, as shown in FIG. 53H). It models the similarity of highdimensional points Z( and z, as a conditional probability, Pj t, and the similarity of the lowdimensional map points ytand y as a conditional probability, <7 / | / - The algorithm minimizes the Kullback-Leibler (KL) divergence between the joint probability distributions P and Q.

[0564]

[0565]

[0485] The resulting plot, when color-coded by gesture class, provides a clear visual representation of cluster separability.

[0566]

[0486] Finally, To ensure that the augmentations did not corrupt essential signal features, the absolute error between a decoded augmented view and the original canonical signal was calculated. The error for a given view v at time t, averaged over all C channels, is:

[0567] err

[0568]

[0569]

[0487] A heatmap of this error (for example, as shown in FIG. 73C) shows that reconstruction errors are primarily dependent on the temporal segment rather than the specific augmentation, indicating that the encoder preserves the core signal structure.

[0570] iv. Phase III: Downstream Classifier Architecture

[0571]

[0488] In some embodiments, the trained machine learning model comprises a classification model. For example, for the final classification task, the pretrained and frozen encoder from Phase II was used as a feature extractor. The latent space vectors generated by the encoder were fed into a Multilayer Perceptron (MLP). The MLP consists of two hidden layers with 128 and 64 neurons, respectively, both using ReLU activation functions. The output layer contains 24 neurons, corresponding to the 24 ASL-A classes, and uses a Softmax activation function to produce a probability distribution over the classes.

[0572] v. 3D Gesture Control Model Architecture

[0573]

[0489] The model for the 6-class 3D gesture control task leverages the same SSL-pretrained encoder from the ASL-A task. The encoder weights were frozen, and only the final downstream MLP classifier was retrained. This new MLP has an identical architecture to the ASL-A classifier but with an output layer of 6 neurons (with Softmax activation) to match the number of 3D gesture classes.

[0574] n. Model Training and Validation

[0575] i. Training Procedures

[0576]

[0490] All models were implemented in PyTorch and trained on an NVIDIA A100 GPU. For Phase I, the denoising autoencoder was trained for 60 epochs using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. The Smooth LI loss function is used to compute the reconstruction error between the model’s output and the clean target signal. For Phase II, the contrastive learning framework was trained using the SimSiam framework, also with the Adam optimizer and a learning rate of 0.001. For the downstream classification tasks (Phase III), only the final MLP classifier was trained. The pretrained encoder weights from the SSL phases were frozen. The MLP was trained using the Adam optimizer with a learning rate of 0.001 and the Cross-Entropy loss function until the validation accuracy plateaued.

[0577] ii. Hyperparameter Tuning

[0578]

[0491] No systematic hyperparameter tuning (e.g., grid search) was performed. The learning rates, batch sizes, and architectural parameters were selected empirically based on common practices in the field and their performance on a preliminary validation set.

[0579] iii. Model Evaluation Metrics

[0580]

[0492] To ensure a robust and unbiased estimate of the models’ performance, a five-fold cross-validation procedure was used. The dataset was partitioned into five equal sized folds, and the model is trained five times, each time using a different fold as the test set and the remaining four as the training set. The final reported accuracy is the mean and standard deviation of the performance across these five folds.

[0581]

[0493] The performance of the final classifiers was evaluated using three primary metrics:

[0582]

[0494] Receiver Operating Characteristic (ROC) and Area Under the Curve (AUC): For each class, a one-vs-rest ROC curve was generated by plotting the true positive rate against the false positive rate at various discrimination thresholds. The AUC was then calculated to provide a single scalar value summarizing the model’s ability to distinguish between classes, where a value of 1.0 indicates perfect classification.

[0583]

[0495] Normalized Confusion Matrix: A confusion matrix was generated to visualize the classification performance on the test set. The matrix was normalized by the number of true instances in each class to show the proportion of correct and incorrect predictions for each gesture. A strong diagonal indicates high accuracy, while off-diagonal elements reveal specific misclassifications between gestures.

[0584] o. Application-Specific Implementations

[0585] i. Virtual Musical Instrument

[0586]

[0496] The real-time bimanual musical performance was implemented within a single, unified Python environment using the SCAMP (Suite for Computer-Assisted Music in Python) library. A custom script continuously received the classification output for both the left-hand chord gestures and the right-hand finger-picking motions. Upon detecting a valid combination (e g., the ‘C’ gesture followed by a ‘thumb’ pick), the script directly interfaced with a SCAMP session. This session managed a virtual instrument part, triggering the playback of the corresponding musical chord or note via MIDI messages sent to a software synthesizer. This integrated approach streamlines the signal processing pipeline, removing the latency and complexity associated with inter-process communication protocols like Open Sound Control (OSC) and external sound synthesis environments.

[0587] ii. Augmented Reality 3D Control

[0588]

[0497] The camera-free control of the 3D object was implemented using the Blender Python API (v4.5.2 LTS). The real-time classification results from the TRAiLL system were streamed from the main Python script to the Blender environment via a local Web Socket connection. A Python script within the Blender environment acted as a WebSocket client, continuously listening to incoming gesture labels. Upon receiving a label (e g., ‘grab’, ‘pinch’), a state machine within the script triggered the corresponding action on the target 3D object. For example, a ‘pinch’ gesture scales the object, a ‘grab’ gesture attaches it to the virtual hand for translation, and a ‘wave’ gesture initiates a rotation, allowing for intuitive, real-time manipulation of the virtual object.

[0589] iii. Statistics and reproducibility

[0590]

[0498] All statistical analyses were performed using custom scripts written in Python (v3.12.9). Key scientific computing and machine learning libraries utilized include NumPy (v2.2.5), SciPy (v1.15.2), Pandas (v2.2.3), Matplotlib (v3.10.1), and PyTorch (v2.6.0). A complete list of packages and specific versions used is available in the accompanying source code. Data are presented as mean ± standard error of the mean (s.e.m.) unless otherwise specified. The specific statistical tests, sample sizes (n), and outcomes are detailed below for each experiment. No statistical methods were used to predetermine sample sizes; instead, our sample sizes are like those reported in previous publications in the field of wearable biosensors and human-computer interaction.

[0591]

[0499] For the spatially resolved tracing of finger-specific muscle activity (for example, as shown in FIG. 51A to FIG. 5 IK), the separability of muscle activation patterns was quantified using Linear Discriminant Analysis (LDA). The analysis was performed on a dataset of 3,000 single-frame feature vectors, collected from 3 independent subjects performing 5 repetitions of each finger flexion, where each repetition was treated as a biological replicate. The 95% Mahalanobis confidence ellipses were calculated to visualize cluster separability in the two-dimensional projected space. For the quantification of muscle load and fatigue (for example, as shown in FIG. 52A to FIG. 52 J), the Muscle Activation Factor (MAF) was calculated and plotted over time. The error bands shown for average MAF traces represent the s.e.m. across 5 repeated trials for each condition from a representative subject. For the classification of American Sign Language gestures (for example, as shown in FIG. 53A to FIG. 53L), the model performance was evaluated using a five-fold cross-validation procedure. The dataset consisted of 6,500 gesture samples collected from 3 subjects. For each fold, the data was split into training (80%) and testing (20%) sets, ensuring that data from a single gesture performance was not split across sets. The reported accuracy (92%) represents the mean classification accuracy across the five folds. The Receiver Operating Characteristic (ROC) curves and the Area Under the Curve (AUC) were calculated for each class to assess model performance on a per-class basis. For the bimanual musical performance and 3D gesture control (for example, as shown in FIG. 54A to FIG. 541), the classification model was evaluated on an independent test set of 720 gestures (120 per class) from a single subject, with each gesture performance serving as a technical replicate. The normalized confusion matrix was computed from the predictions on this test set to visualize classification accuracy and inter-class confusion. All experiments involving representative data were successfully repeated on at least three different occasions with similar results.

[0592] p. Data and Code Availability

[0593]

[0500] The raw and processed data and the custom code used for the analyses presented in this study have been deposited in the Zenodo repository and can be accessed via a Digital Object Identifier (DOI). The repository includes all custom Python code for data acquisition, signal processing, statistical analysis, and machine learning.

[0594]

[0501] It is to be understood that the disclosure is not to be limited to the specific embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation, unless described otherwise.

Claims

CLAIMS1. A wearable muscle monitor comprising:an inertial measurement unit (IMU) configured to generate global motion signals associated with a user;at least one near-infrared spectroscopy (NIRS) sensor configured to generate local activity signals associated with the user; anda wireless microcontroller communicatively coupled to the IMU and the at least one NIRS sensor, wherein the wireless microcontroller is configured to:generate one or more global motion data points based on the global motion signals, generate one or more local activity data points based on the local activity signals, andtransmit the one or more global motion data points and the one or more local activity data points to a computing device.

2. The wearable muscle monitor of claim 1, further comprising:a flexible substrate, wherein the IMU, the at least one NIRS sensor, and the wireless microcontroller are disposed on the flexible substrate.

3. The wearable muscle monitor of claim 2, wherein the flexible substrate comprises a main island and at least one daughter island.

4. The wearable muscle monitor of claim 3, wherein the main island and the at least one daughter island are connected through at least one stretchable serpentine hinge.

5. The wearable muscle monitor of claim 3, wherein the IMU, the at least one NIRS sensor, and the wireless microcontroller are disposed on the main island.

6. The wearable muscle monitor of claim 3, wherein the at least one NIRS sensor is disposed on the at least one daughter island.

7. The wearable muscle monitor of claim 3, wherein the at least one daughter island comprises a left daughter island and a right daughter island, wherein the left daughter island is connected to a left side of the main island, wherein the right daughter island is connected to a right side of the main island.

8. The wearable muscle monitor of claim 7, wherein the at least one NIRS sensor comprises:a central NIRS sensor disposed on a top portion of the main island;a bottom NIRS sensor disposed on a bottom portion of the main island;a top left NIRS sensor disposed on the left daughter island; anda top right NIRS sensor disposed on the right daughter island.

9. The wearable muscle monitor of claim 1, wherein the one or more global motion data points indicate motions associated with a head of the user.

10. The wearable muscle monitor of claim 9, wherein the one or more local activity data points indicate activities associated with a neck muscle of the user.

11. A computer-implemented method comprising:receiving, by a processor, one or more global motion data points and one or more local activity data points associated with a user; andgenerating, by the processor, one or more predicted health status data points associated with the user based at least in part on the one or more global motion data points, the one or more local activity data points, and a dual-channel recurrent neural network (RNN).

12. The computer-implemented method of claim 11, wherein the one or more global motion data points are based on global motion signals from an inertial measurement unit (IMU).

13. The computer-implemented method of claim 11, wherein the one or more local activity data points are based on local activity signals from at least one near-infrared spectroscopy (NIRS) sensor.

14. The computer-implemented method of claim 11, wherein the dual-channel RNN is based on gated recurrent units (GRU).

15. The computer-implemented method of claim 11, wherein the dual-channel RNN comprises two parallel hidden layers.

16. The computer-implemented method of claim 15, wherein the two parallel hidden layers comprise a first hidden layer and a second hidden layer, wherein the computer-implemented method comprises inputting the one or more global motion data points to the first hidden layer and inputting the one or more local activity data points to the second hidden layer.

17. The computer-implemented method of claim 15, wherein outputs from the two parallel hidden layers are concatenated and sent to a fully connected (FC) layer.

18. The computer-implemented method of claim 11, wherein the one or more global motion data points indicate motions associated with a head of the user.

19. The computer-implemented method of claim 18, wherein the one or more local activity data points indicate activities associated with a neck muscle of the user.

20. The computer-implemented method of claim 19, wherein the one or more predicted health status data points indicate one or more predicted coughing events associated with the user.

21. A wearable muscle monitor comprising:a sensor array comprising a plurality of sensing pixels and configured to generate a plurality of optical sensing signals, wherein each of the plurality of sensing pixels comprises a near-infrared light-emitting diode (NIR LED) and a photodiode (PD); anda computing device communicatively coupled to the sensor array and configured to generate at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals.

22. The wearable muscle monitor of claim 21, wherein the at least one predicted tissue morph data object comprises a muscle activation factor (MAF).

23. The wearable muscle monitor of claim 21, wherein the at least one predicted tissue morph data object comprises a predicted gesture classification.

24. The wearable muscle monitor of claim 23, wherein the computing device is configured to:generate the predicted gesture classification based on inputting the plurality of optical sensing signals to a trained machine learning model.

25. The wearable muscle monitor of claim 24, wherein the trained machine learning model comprises an encoder network.

26. The wearable muscle monitor of claim 24, wherein the trained machine learning model comprises a contrastive learning model.

27. The wearable muscle monitor of claim 24, wherein the trained machine learning model comprises a classification model.

28. A computer-implemented method comprising:receiving, by at least one processor, a plurality of optical sensing signals from a sensor array, wherein the sensor array comprises a plurality of sensing pixels; andgenerating, by the at least one processor, at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals.

29. The computer-implemented method of claim 28, wherein the at least one predicted tissue morph data object comprises a muscle activation factor (MAF).

30. The computer-implemented method of claim 28, wherein the at least one predicted tissue morph data object comprises a predicted gesture classification.

31. The computer-implemented method of claim 30 further comprising: generating, by the at least one processor, the predicted gesture classification based on inputting the plurality of optical sensing signals to a trained machine learning model.

32. The computer-implemented method of claim 31, wherein the trained machine learning model comprises an encoder network.

33. The computer-implemented method of claim 31, wherein the trained machine learning model comprises a contrastive learning model.

34. The computer-implemented method of claim 31, wherein the trained machine learning model comprises a classification model.

35. A computer program product comprising at least one non-transitory computer-readable storage medium having computer-readable program code portions stored therein, the computer-readable program code portions comprising an executable portion configured to:receive, by at least one processor, a plurality of optical sensing signals from a plurality of sensing pixels of a sensor array; andgenerate, by the at least one processor, at least one predicted tissue morph data object based at least in part on the plurality of optical sensing signals.

36. The computer program product of claim 35, wherein the at least one predicted tissue morph data object comprises a muscle activation factor (MAF).

37. The computer program product of claim 35, wherein the at least one predicted tissue morph data object comprises a predicted gesture classification.

38. The computer program product of claim 37 further comprising:generating, by the at least one processor, the predicted gesture classification based on inputting the plurality of optical sensing signals to a trained machine learning model.

39. The computer program product of claim 38, wherein the trained machine learning model comprises an encoder network.

40. The computer program product of claim 38, wherein the trained machine learning model comprises a contrastive learning model.

Images