Methods and systems for physiological detection and alerting - Patents.com
The described method and system effectively reduce false alarms in seizure detection by using biosensors and advanced processing techniques, ensuring high sensitivity and low false alarm rates in both inpatient and outpatient settings, thereby improving seizure detection accuracy and reducing SUDEP risk.
Patent Information
- Application Number
- JP2024527712
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-12-01
- Filing Date
- 2022-11-30
- Publication Date
- 2024-11-15
AI Technical Summary
Existing non-electroencephalography (non-EEG)-based seizure monitoring devices in outpatient settings suffer from high false alarm rates (FAR), leading to alarm fatigue and poor adherence, despite demonstrating acceptable sensitivity in inpatient settings.
A method and system utilizing biosensors and hardware processors to generate confidence scores for characteristic physiological event phases, integrating multiple classifiers and temporal modeling techniques to accurately detect seizures and reduce false alarms.
The system achieves a low false alarm rate of 0.05/day in inpatient settings and 0.13/day in outpatient settings, with perfect positive agreement and reasonable latency, enhancing caregiver intervention and reducing the risk of sudden unexpected death in epilepsy (SUDEP).
Smart Images

Figure 2024542431000001_ABST
Abstract
Description
[Technical field]
[0001] [CROSS-REFERENCE TO RELATED APPLICATIONS] This application claims priority to U.S. Patent Application No. 63 / 284,891, filed December 1, 2021, the disclosure of which is incorporated herein by reference in its entirety.
[0002] [Government support] This invention was made with Government support under Grant No. 2018-MII-4922 awarded by the Maryland Innovation Initiative. The Government has certain rights in this invention.
[0003] The present application is directed to methods and systems for physiological detection and alerting. [Background technology]
[0004] Uncontrolled seizures affect up to 56% of epilepsy patients and impose a substantial physical, psychological and economic burden. Accurate information is necessary for better management, but is difficult to obtain in outpatient settings. Patient-reported outcomes (PROs) have been proposed in the form of seizure diaries, but are limited by poor adherence and postictal amnesia. Uncontrolled tonic-clonic seizures (TCS) significantly increase the risk of sudden unexpected death in epilepsy (SUDEP) by up to 27-fold. The risk of SUDEP can be significantly reduced by caregiver intervention, but this requires continuous monitoring and timely warning. Indeed, multiple studies have highlighted the need for wearable devices with reliable seizure monitoring capabilities that alert caregivers and provide an accurate record of seizures, and various such devices have been developed over the past decade.
[0005] To date, non-electroencephalography (non-EEG) based TCS monitoring devices have generally demonstrated acceptable sensitivity (>90%) in inpatient settings, but their false alarm rates (FAR) were too high in both inpatient and outpatient settings. High FARs risk causing alarm fatigue in both caregivers and patients, resulting in low adherence to device monitoring and alerts. This is even more prevalent in outpatient user settings, where routine activities may frequently trigger false alarms. To date, in large-scale video-electroencephalography (vEEG) validated prospective multicenter studies (>4000 hours of recording), the mean FAR ranged from 0.2 / day to 0.83 / day, with significant variability across algorithms, age groups, and activity levels within the EMU. Thus, there remains a need to develop seizure detection algorithms that can significantly reduce the FAR while maintaining or improving current sensitivity standards in both EMU and outpatient settings. Summary of the Invention
[0006] According to an embodiment of the present disclosure, a method for physiological event detection and alerting is disclosed, which includes the steps of obtaining a set of biosensor data from a user from one or more biosensors, generating, by one or more hardware processors, a set of processed biosensor data from the set of biosensor data, generating, by the hardware processors, a set of features from the processed biosensor data, the set of features being associated with one or more characteristic physiological event phases, wherein an association between the set of features and the one or more characteristic physiological event phases is stored in one or more non-transitory storage media, and using the one or more hardware processors to generate a set of features associated with one or more characteristic physiological event phases from the generated set of features, the set of processed biosensor data, or both the generated set of features and the processed biosensor data. determining, for each characteristic physiological event phase, a confidence score indicative of the presence of that phase in the data partition; determining, using one or more hardware processors, a final confidence score suggesting an occurrence of a physiological event based on a relationship between the confidence scores of all of the physiological event phases from a relationship between the confidence scores of each characteristic physiological event phase; determining, using one or more hardware processors, a cumulative confidence score suggesting an occurrence of a particular physiological event from an accumulation of the final confidence scores, wherein the physiological event is composed of one or more characteristic physiological event phases; and providing, by the one or more hardware processors, a warning of a potential physiological event based on the cumulative confidence score.
[0007] According to an embodiment of the present disclosure, a system for physiological event detection and alert is disclosed that includes one or more biosensors that capture, record, or both capture and record biosensor data from a user, one or more hardware processors, one or more non-transitory computer-readable media that store instructions that, when executed by the one or more hardware processors, implement a method for physiological detection and alert, and a user interface that provides alerts of potential physiological events.The physiological detection and alert method includes the steps of obtaining a set of biosensor data from a user from one or more biosensors; generating, by one or more hardware processors, a set of processed biosensor data from the set of biosensor data; generating, by the hardware processors, a set of features from the processed biosensor data, the set of features being associated with one or more characteristic physiological event phases, wherein an association between the set of features and the one or more characteristic physiological event phases is stored in one or more non-transitory storage media; and using the one or more hardware processors to generate a set of features associated with the one or more characteristic physiological event phases from the generated set of features, the set of processed biosensor data, or both the generated set of features and the processed biosensor data. determining, for each characteristic physiological event phase of the physiological event data, a confidence score indicative of the presence of that phase in the data partition; determining, using one or more hardware processors, a final confidence score suggesting an occurrence of a physiological event based on a relationship between the confidence scores of all of the physiological event phases from a relationship between the confidence scores of each characteristic physiological event phase; determining, using one or more hardware processors, a cumulative confidence score suggesting an occurrence of a particular physiological event from an accumulation of the final confidence scores, wherein the physiological event is comprised of one or more characteristic physiological event phases; and providing, by the one or more hardware processors, a warning of a potential physiological event based on the cumulative confidence score.
[0008] The methods and / or systems for physiological detection and alerting may include a variety of additional features, including one or more of the features described below.
[0009] The one or more biosensors include one or more of an accelerometer, a photoelectric (PPG) sensor, a gyroscope, a microphone, a blood oxygen level sensor, a blood pressure sensor, a blood glucose sensor, an eye sensor, an electrodermal activity sensor, an eye gaze sensor or tracker, a pupillometer sensor, or combinations thereof.
[0010] The one or more biometric sensors are incorporated into a wearable device consisting of a watch, glasses, a cuff, a necklace, a bracelet, eyeglasses, a headset, one or more finger rings, or a combination thereof.
[0011] The pre-processed biometric data set includes biometric data that has been filtered for noise reduction and interpolation.
[0012] The method for physiological detection and alerting may further include processing the set of biosensor data to create a set of processed biosensor data; reducing dataset imbalance between physiological events and non-physiological events in the processed biosensor data by iteratively training and using one or more models to identify anomalous classifications in the biosensor data of non-physiological events to create a balanced dataset, the one or more models including one or more anomaly detection methods; and using the balanced dataset to train one or more classifiers for each characteristic physiological event phase that creates a confidence score for each characteristic physiological event phase.
[0013] The one or more anomaly detection methods include one or more of a separation forest, a one-class support vector machine (SVM), a hidden Markov model (HMM), an autoencoder, a variational autoencoder, cluster-based outlier detection, or a combination thereof.
[0014] Each characteristic physiological event phase includes an event that causes a characteristic biosignal pattern associated with all or a portion of the physiological event, where the characteristic biosignal pattern includes one or more of tonic movements and / or associated physiological changes, clonic movements and / or associated physiological changes, postictal movement suppression or impairment and / or associated physiological changes, prodromal movements and / or associated physiological changes, early ictal movements and / or associated physiological changes, late ictal movements and / or associated physiological changes, ictal screaming and / or associated physiological changes, specific automatisms including one or more of hand tremors, tremors, paroxysmal eye blinks or gaze fixations, saccades, fixations, noises, movement arrest, or specific physiological responses including one or more of heart rate changes or blood pressure changes.
[0015] The set of features generated from the processed biosensor data uses techniques including one or more of manual feature extraction, automatic feature extraction, or a combination thereof. Manual feature extraction includes one or more of time domain feature extraction, frequency domain feature extraction, or a combination thereof. Time domain feature extraction includes one or more of line crossing, variance, skewness, kurtosis, or a combination thereof. Frequency domain feature extraction includes one or more of Fan-Chirp transform, Fourier transform, Chirp-Z transform, constant-Q transform, wavelet transform, or a combination thereof. Automatic feature extraction includes one or more of deep learning methods and one or more convolutional neural networks.
[0016] The confidence score for each characteristic phase is calculated using a classifier, including classical techniques, including one or more linear models, one or more tree-based techniques, one or more clustering techniques, one or more probabilistic graphical models, one or more deep learning models, or a combination thereof.
[0017] The relationship between the confidence scores that determines the final confidence score includes a technique for aggregating confidence scores, including one or more of non-temporal techniques that analyze a single time point, one or more classical temporal techniques that analyze multiple past time points, one or more deep learning techniques, or a combination thereof. The non-temporal techniques include one or more of averages, weighted averages, confidence score formulas, or a combination thereof. The temporal techniques include probabilistic graphical techniques. The probabilistic graphical techniques include one or more hidden Markov models, one or more conditional random fields, or both. The deep learning techniques include one or more of recurrent neural networks, long short-term memory networks, gated recurrent unit networks, temporal convolutional networks, convolutional neural networks, multi-layer perceptrons, or a combination thereof.
[0018] The accumulation of the final confidence scores to generate a cumulative confidence score includes one or more of a low pass filter and a temporal modeling technique, the temporal modeling technique including one or more of a hidden Markov model, a conditional random field, a recurrent neural network, a long short-term memory network, a gated recurrent unit network, a temporal convolutional network, a convolutional neural network, or a combination thereof.
[0019] The potential physiological event alert is provided on a user interface of a wearable device worn by a user. The potential physiological event alert is provided to one or more of a user, a caregiver, a health care provider, or a legal guardian. A physiological event is any event that causes one or more characteristic patterns identifiable through one or more biosensors, including epileptic seizures, syncope, psychogenic non-epileptic seizures, movement disorders, or combinations thereof.
[0020] The one or more hardware processors include a first processor in a first device worn on the head or face and a second processor in a second device worn on the wrist or other part of the body. The first device worn on the head or face is a pair of glasses and the second device is a wristwatch. The physiological event includes a neurological event, a cardiac event, or a combination thereof. The neurological event is a seizure. [Brief description of the drawings]
[0021] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present teachings and, together with the description, serve to explain the principles of the present teachings.
[0022] [Figure 1] 1A and 1B are back and front perspective views, respectively, of a smartwatch according to an example of the present disclosure. [Diagram 2] FIG. 2 is a diagram of smart glasses according to an example of the present disclosure. [Diagram 3] 3A and 3B are diagrams showing the distribution of individual FAR rates for different patients partitioned by EMU and outpatient environments according to an example of the present disclosure, where the circumference of the polar plot designates the time of day in 24-hour format and the radius indicates the number of sessions recorded during that time period. [Figure 4] FIG. 4 illustrates the distribution of individual FAR rates for different patients partitioned by EMU and outpatient settings according to an example of the present disclosure. [Diagram 5] FIG. 5 illustrates a portion of tonic-clonic seizures (TCS) detected during a prospective study centered on the time of detection according to an example of the present disclosure. [Figure 6] FIG. 6 is a diagram illustrating a method for physiological event detection and alerting according to an example of the present disclosure. [Figure 7] FIG. 7 is a diagram illustrating a method for training data according to an example of the present disclosure. [Figure 8] FIG. 8 illustrates data from a generalized tonic-clonic seizure according to an example of the present disclosure. [Figure 9] FIG. 9 shows data from clonic phase only (no TCS) according to an example of the present disclosure. [Figure 10] FIG. 10 illustrates data from an example exercise with predominantly tonic phase (non-TCS), in accordance with an example of the present disclosure. [Figure 11] FIG. 11 illustrates data from a long single phase according to an example of the present disclosure. [Figure 12] FIG. 12 illustrates data from a focal onset bilateral tonic-clonic seizure (FBTCS) according to an example of the present disclosure. [Figure 13] FIG. 13 is a schematic diagram of a computing system according to an example of the present disclosure. [Figure 14] FIG. 14 illustrates an exemplary method for data training and physiological event detection and alerting, according to an example of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0023] Reference will now be made in detail to the embodiments, examples of which are illustrated in the accompanying figures and drawings. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without these specific details. In other instances, well-known methods, procedures, components, circuits, and networks have not been described in detail so as not to unnecessarily obscure aspects of the embodiments.
[0024] As used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural references unless the context clearly dictates otherwise. Thus, for example, reference to "a method" includes one or more methods, and / or steps of the type described herein and / or that will become apparent to those skilled in the art upon reading this disclosure and so forth.
[0025] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only, and is not intended to be limiting. Further, unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs.
[0026] In general, the present disclosure describes a method and system for physiological detection and alerting. The method and system may use one or more biosensors to collect, record, process, pre-process, or any combination thereof, biodata from a user. The one or more biosensors may operate with one or more hardware processors, either co-located with or remote from one or more of the one or more biosensors, and include communicating the biodata over a network environment to process the biodata. The disclosed method and system obtains a set of biosensor data from the one or more biosensors from the user. The disclosed method and system then generates a set of processed biosensor data from the set of biosensor data. The disclosed method and system then generates a set of features associated with one or more characteristic physiological event phases from the processed biosensor data. The disclosed method and system then determines a confidence score for each characteristic physiological event phase of the one or more characteristic physiological event phases from the generated set of features, the processed biosensor data, or both the generated set of features and the processed biosensor data, indicating the presence of that phase in the data partition. The disclosed method and system then determines a final confidence score suggesting the occurrence of a physiological event based on the relationship between the confidence scores of all physiological event phases from the relationship between the confidence scores of each characteristic physiological event phase. The disclosed method and system then determines a cumulative confidence score suggesting the occurrence of a particular physiological event from an accumulation of the final confidence scores, where the physiological event is comprised of one or more characteristic physiological event phases. The disclosed method and system then provides a warning of a potential physiological event based on the cumulative confidence score.
[0027] The following discussion deals primarily with work done in relation to seizures, but this is but one example of a physiological event that can be detected and alerted using the methods and systems of the present disclosure. Other neurological or cardiac events can be detected and alerted based on biosensor data collected from one or more biosensors and analyzed by one or more processors operating the methods of the present disclosure.
[0028] 1A and 1B show rear perspective 102 and front perspective 104 views of a smartwatch 100 according to an example of the present disclosure. The rear of the smartwatch 100, which contacts the user's skin, has one or more biosensors 106, such as one or more PPG sensors. The front of the smartwatch 100 includes a user interface 108 for providing personalized alerts to the user. The smartwatch 100 can include additional sensors, such as, but not limited to, an accelerometer, a gyroscope, a microphone, a blood oxygen level sensor, a blood pressure sensor, a blood glucose sensor, an electrodermal activity sensor, or combinations thereof. FIG. 2 shows smart glasses with sensors 205 according to an example of the present disclosure. The sensors 205 can include, but are not limited to, an eye sensor, a gaze sensor or tracker, a pupillometer sensor, or combinations thereof. The smart glasses can be used alone or in combination with the smartwatch 100 to provide biosensor data that can be used for physiological event detection and alerts.
[0029] For example, monitoring for tonic-clonic seizures (TCS) is important to enhance safety, promote independence, and avoid sudden unexpected death in epilepsy (SUDEP). Any system for TCS monitoring needs to be highly sensitive, exhibit low false alarm rates (FAR), and provide alerts to caregivers with low latency for use in both inpatient and outpatient settings. Ideally, these devices should be non-invasive, multifunctional, and avoid stigma. The pilot study described below illustrates the performance characteristics of a seizure monitoring application implemented in a consumer smartwatch using the disclosed methods and tested in both inpatient and outpatient settings. In this study, TCS encompassed all major tonic-clonic seizure types defined by the ILAE, including generalized tonic-clonic seizures (GTCS), focal onset bilateral tonic-clonic seizures (FBTCS), tonic-clonic seizures of unknown onset (UTCS), and myoclonic-tonic-clonic seizures (MTCS).
[0030] In this study, we first collected data from 340 patients in four epilepsy monitoring units (EMUs) and 21 outpatient users (13 outpatients with epilepsy and 8 healthy controls without epilepsy). Accelerometer (ACM) and heart rate signals were recorded using an application developed for the smartwatch. Other biometric data can also be collected, including but not limited to electromyography (EMG), photoplethysmography (PPG), and electrodermal activity (EDA). Seizures in the EMUs were verified with video electroencephalography (vEEG), but seizure events in the outpatient users were self-reported and not included in the training set because they were difficult to verify. This resulted in a dataset of 20,388 hours including 79 TCSs (58 from EMU patients) and 5,642 seizure-free hours from the outpatient users (outpatients and healthy controls). This data was used to train a novel classifier in an offline environment and then implemented on the smartwatch as part of the seizure monitoring application. A prospective study was performed on 85 EMU patients and 15 outpatient users (9 outpatients, 6 healthy controls). EMU patients were blinded to seizure detection and seizures were verified by vEEG. Outpatient users were not blinded to seizure detection and seizures were either self-reported or identified retrospectively through an independent biosignal analysis. The study dataset consisted of 4,279 h in EMU with 19 seizures (15 patients) and 6,735 h in outpatient with 10 self-reported seizures (3 patients). The prospective study showed a positive agreement rate (PPA) of 100%, with a FAR of 0.05 per day (positive predictive value (PPV) of 68%) in EMU and 0.13 per day (PPV of 22%) in outpatient users. Eight of the 31 total false alarms were due to one outpatient patient. The FAR for all other outpatient users except this outpatient patient was 0.10 per day. The mean detection latency was 37.38 seconds (standard deviation 13.24 seconds) for EMU users and 32.07 seconds (standard deviation 10.22 seconds) for outpatient users.
[0031] In this study, all biosensor data were collected using smartwatches and paired smartphones. Each smartwatch was equipped with an application called EpiWatch, which enabled data collection from the watch's built-in triaxial accelerometer (ACM; sampling frequency 50 Hz) and PPG (sampling frequency 0.2 Hz) sensors. The watch was worn on the dorsal side of the wrist, ensuring a snug fit between the PPG sensor and the skin to prevent data loss. In all EMU facilities, when motor onset was known to be more pronounced, the watch was worn on the arm with the more pronounced motor onset.
[0032] When the watch was actively monitoring, the EpiWatch application de-identified, encrypted, and stored the watch sensor data in a cloud-based backend for further analysis and algorithm development. In the study, a TCS detection algorithm was implemented in the EpiWatch application that continuously monitored the sensor data and provided TCS detection and alerts. All alerts and events from the monitoring were also stored in the cloud-based backend. For redundancy, all alerts were also stored in the watch and could be retrieved manually or automatically stored in the cloud-based backend in case of temporary connectivity interruption.
[0033] The EMU training dataset, consisting of 20,388 hours of data, was recorded across four sites: Johns Hopkins Hospital Adult EMU (JHA), Le Bonheur Children's Hospital Pediatric EMU (LBH), Ruber International Hospital Adult EMU (RBI), and the University of Maryland Medical Center Adult EMU (UMD). There were a total of 340 users and 79 seizures (by 58 users). Each seizure was verified using vEEG by two board-certified clinical neurologists and classified as TCS or not.
[0034] The outpatient user (AMB) dataset consisted of a total of 5,462 hours recorded across outpatient epilepsy users (OUT) and healthy control users without epilepsy (NC). The outpatient set consisted of PWE testing of the algorithm during normal activities outside of the hospital. Outpatient user seizures were not included in the training dataset due to the inherent limitations of validating seizure type and occurrence without vEEG. These seizure segments were discovered using Patient Reported Outcomes (PROs), automatic motion detection, and visual analysis of time series segments. Despite best efforts, TC seizures may still be present in this dataset.
[0035] The healthy control dataset consists of everyday user activities acquired to estimate the FAR during activities that traditionally cause false alarms in monitoring devices. These activities include brushing teeth, exercising, washing dishes, playing drums, dancing, etc. The breakdown of the training dataset is shown in Table 1.
[0036] Table 1. Data used for monitor training consist of time logged, number of users logged, total number of tonic-clonic seizures (TCS), and number of users who experienced at least one seizure. The outpatient (AMB) environment is divided into outpatients (OUT) and non-control users (NC; outpatient users without epilepsy). The EMU environment is divided into four separate hospital EMUs: Johns Hopkins Adult (JHA), University of Maryland (UMD), Ruber International (RBI), and La Bonheur Children's (LBH).
[0037] [Table 1]
[0038] A real-time out-of-sample performance test of the detection algorithm was performed across four epilepsy monitoring units (EMUs): the adult and pediatric EMUs (JHA 6 beds and JHP 4 beds, respectively) at Johns Hopkins Hospital, the pediatric EMU (LBH 10 beds) at Le Bonheur Hospital, and the adult EMU (RBI 3 beds) at Ruber International Hospital. Patients who participated in the testing phase were different from those who provided data for training the seizure detection algorithm. The algorithm was not modified during the entire testing period. To prevent patient / caregiver bias, all detections were silent and alerts were sent directly to a central coordinator with corresponding timestamps recorded in a cloud-based backend. Seizures in the EMUs were verified using video-EEG by two board-certified clinical neurologists. Video-EEG data were reviewed at the hospital with the EMU facility or at the Johns Hopkins Hospital after transmission via encrypted hard drive. To ensure continuous tracking, nursing staff at the EMU facility were asked to charge the watches twice a day, once at approximately 7 am and once more at approximately 7 pm, or to use multiple watches per patient if watches were available. These times were chosen to reduce the risk of missing a seizure during the charging period. Tracking was automatically discontinued if charging was not performed manually to avoid artificially lengthening the overall recording period.
[0039] The detection algorithm was also tested on outpatients to obtain its performance in a real-life environment. Users either already owned or were provided with a smartwatch paired with their smartphone and were asked to download the application through an application store. The application incorporated an electronic consent feature for subjects and caregivers, and once the subject or their legally authorized representative consented, they were invited to participate in testing the algorithm in an outpatient environment. Subjects and caregivers were warned not to rely on EpiWatch detection as the sole method to obtain seizure help. No changes were made to the detection algorithm for outpatients during the entire test period. As most outpatients used the application in a realistic manner, the alert feature was enabled for these patients, and alerts were sent to selected caregivers.
[0040] Seizures in outpatients were verified primarily through patient / caregiver communication, PROs (see above), and manual biosignal analysis. Some subjects with video monitoring installed at home provided video evidence of seizure behavior. Nurses also interviewed patients' caregivers who witnessed the seizure or arrived shortly after the seizure occurred. To reduce the risk of false negatives due to potentially unreported seizures, follow-up communication with outpatients and / or caregivers was performed in addition to automated motion detection and manual biosignal analysis. Although unlikely, it is possible that false negatives still occurred during a prospective outpatient study.
[0041] In the EMU training and prospective test sets, a set of exclusion criteria was developed to classify specific recording segments as valid TCS for this pilot study. Seizures that met any of the criteria were excluded from the final metrics. E1. No EEG data was available for the time period during which the seizure occurred. E2. The patient was not seen on video during the time the seizure occurred. E3. The Apple Watch with the working detection algorithm was not worn on the patient's wrist and was not actively monitored. E4. At least 2 board-certified epileptologists disagree on the TCS classification after review of the video-EEG according to the ILAE classification system. E5. The motor symptoms of TCS did not occur in the limb monitored by the watch, such as unilateral TCS. E6. Excessive external (caregiver / nurse / doctor) restraint of limb movement. Note that TCS monitoring is not required in this case.
[0042] In addition, during the prospective study, as soon as the device recorded a detection, it was counted in the results as a detection, unless it could be confirmed via video-EEG and collected data that the watch was not worn by and tracking the patient at the time of detection.
[0043] During the study, each vEEG verified seizure and / or algorithm detection was treated as an event classified as either a true positive (TP; segment was TCS and was detected by the algorithm), a false positive (FP; segment was not TCS but was detected by the algorithm, also called a false alarm), or a false negative (FN; segment was TCS but was not detected by the algorithm). To allow comparison with other papers in the field, similar metrics have been used, namely sensitivity / positive agreement (PPA) with 95% confidence intervals (CIs), precision / positive predictive value (PPV) with 95% CIs, false alarm rate (FAR) with 95% CIs, and latency.
[0044] PPA indicates how effectively a detector can detect TCS. Because of the potential adverse effects of false negatives, a PPA value would ideally be 100%, but in practice this is not possible without producing an unacceptably high false alarm rate. PPV indicates how likely a detection is to be a true positive. Detectors that produce many false positives (false alarms) will have low PPV values (closer to 0). Because PPA and PPV are both measures of binomial proportions (and therefore between 0 and 1), we used the Wilson score with continuity correction to calculate 95% CIs. The Wilson score has been shown to be the most accurate and robust of the known methods for calculating binomial proportion confidence intervals. We used the continuity correction because it provides slightly lower CI estimates compared to uncorrected cases.
[0045] FAR is defined as the number of false alarms (FA, or false positives FP) occurring per 24 hours. Since there is no universal standard on how the data should be split to calculate FAR, we decided to use the micro-FAR and macro-FAR metrics. Micro-FAR is calculated as the total number of false alarms across all users divided by the total number of recorded hours across all users. Macro-FAR is calculated as the average of the FARs calculated individually for each user. Micro-FAR can be thought of as a weighted version of macro-FAR, weighted by the proportional number of recorded hours for a particular user. Generally, most papers report micro-FAR in their metrics.
[0046] Because the FAR is not a binomial proportion, we used a nonparametric bootstrap method to approximate the 95% confidence intervals. Sampling with replacement was performed at the patient level, drawing 10,000 individual samples of the FAR to account for within-patient variability and to be large enough for our chosen alpha value of 0.05. The resulting distribution is approximately normal. To obtain the confidence intervals, the 2.5th and 97.5th percentiles of the samples were selected from the nearest sample values. The method differed slightly for the micro- and macro-FARs, which were calculated separately for each of the 10,000 samples. However, we performed bootstrap only once, and the same sample data were used to estimate the CIs for the micro- and macro-FARs.
[0047] Latency is defined as the time from seizure onset until detection. Setting seizure onset time is somewhat subjective, making it a difficult metric to standardize. We selected the beginning of motor expression of the tonic-clonic phase as evident in the video and recorded ACM data as the onset time. All seizure onset times in this study were determined by at least one board-certified epileptologist. Mean and standard deviation values of latency were measured for inpatients and outpatients.
[0048] 3A and 3B are diagrams showing the distribution of individual FAR rates for different patients partitioned by EMU and outpatient environments according to an example of the present disclosure, where the circumference of the polar plot designates the time of day in 24-hour format and the radius indicates the number of sessions recorded during that time period.
[0049] Over a 206-day period (4,279.88 h), 85 new out-of-sample (OOS; not previously included in the training data) patients diagnosed with epilepsy were enrolled for vEEG monitoring at four EMU sites: JHA (29 patients), LBH (17 patients), RBI (37 patients), and Johns Hopkins Hospital Pediatric EMU (JHP; 2 patients). These patients were monitored using smartwatches running the EpiWatch research app, which provides real-time monitoring of TCS. There were 28 detections across the four EMUs, and 19 TCS from 15 patients were confirmed by vEEG. The breakdown of EMU session coverage in Figure 3A and Figure 3B shows that, on average, the smartwatches were charged approximately twice per day, once at 7 am and once at 7 pm.
[0050] A total of 15 users (6 out-of-sample users) from two groups, outpatient users (9, 5 out-of-sample) and healthy control users (6, 1 out-of-sample), were tested over 6,735.03 hours of outpatient user testing. The outpatient users were asked to use the algorithm as frequently as possible, with the goal of covering 24 hours, although there was some variation in timing due to users charging their devices at different times of the day. The breakdown of session coverage for the outpatient users is shown in Figures 3A and 3B.
[0051] Table 2. Total recording time during prospective testing, and the number of total and out-of-sample (OOS) users. Data are divided by outpatient (AMB) and EMU settings and their respective facilities. The outpatient setting is divided into outpatients (OUT) and non-control users (NC; outpatient users without epilepsy). The EMU setting is divided into four separate hospital EMUs: Johns Hopkins Adult (JHA), Johns Hopkins Pediatric (JHP), Ruber International (RBI), and La Bonheur Children's (LBH).
[0052] [Table 2]
[0053] For the performance evaluation of continuous monitoring, we determined whether the performance metrics of the detection algorithm were not biased by the amount of data collected at any given time of day. To this end, polar plots of the cumulative recording duration are shown in Figure 3A and Figure 3B to confirm that time periods throughout the day were monitored approximately uniformly. In particular, the decrease in recording duration during the time periods 06:00–08:00 and 19:00–21:00 for EMU users and 11:00–13:00 and 19:00–21:00 for outpatients indicates the time periods when the Apple Watch was charging. Outpatient users tended to have more tracking sessions at night (especially 01:00–09:00), while EMU users tended to have more tracking sessions during the day (especially 13:00–17:00).
[0054] Figures 3A and 3B show the distribution of time slots recorded during tracking sessions over time. Figure 3A shows the tracking distribution for EMUs, and Figure 3B shows the tracking distribution for ambulatory users. The circumference of the polar plots specifies the time of day in 24-hour format. The radius indicates the number of sessions recorded during that time slot.
[0055] A total of 29 seizures occurred during 10,990.81 hours of monitoring across the EMU (19 TCS) and outpatients (10 TCS). Notably, in the EMU, the distribution of seizures was relatively even, with 15 patients experiencing seizures. In the outpatients, only three patients experienced seizures, with one patient experiencing eight TCS. The number of recorded hours for adult and pediatric patients was relatively similar, with 2,483 hours in the adult EMU and 1,796 hours in the pediatric EMU.
[0056] In both outpatient and EMU patients, all seizures were detected by the monitoring application, resulting in a PPA of 100% (1.0). Because the outpatient dataset had only 10 seizures and used the more conservative continuity-corrected Wilson method, the 95% CI was wide, with a lower limit of 0.66. The 95% CI for EMU PPA was much narrower, due to the increased sample size, with a lower limit of 0.79. A recent FDA clearance for this class of device specifies a minimum required PPA CI lower limit of 0.7 [5] (although it is unclear what method was used to calculate the CI).
[0057] At the operating point chosen for this study, there were only nine false alarms in 4,280 hours of follow-up across all four EMUs. This translated to a micro-FAR of 0.05 / day, with a 95% CI of [0.02, 0.08]. The micro- and macro-FARs seemed to be in good agreement across all EMUs, meaning that no single patient significantly influenced the FAR calculation. The exception was RBI, where the macro-FAR was 0.22, suggesting that a small number of patients had high FARs over short recording periods. This was indeed the case, with one RBI patient having one FA over a total recording period of 7.9 hours, giving that patient an FAR of 3.05 / day. Only 19 seizures occurred over a total recording time of 4,280 hours in the EMUs, with a PPV of 0.68 [0.48, 0.83], meaning approximately two TPs per FA. Interestingly, when comparing the macroFAR of 0.03 / day in pediatric patients (LBH and JHP) with that of 0.07 / day in adult patients (JHA and RBI), patients in the adult EMU appeared to have more FA than those in the pediatric EMU.
[0058] In the outpatient user dataset, there were 36 false alarms in 6,735 hours of recording, resulting in a micro-FAR of 0.13 / day [0.08, 0.24]. As expected, outpatient users had higher FARs than inpatient users. In fact, one outpatient user had 8 false alarms in 24 hours of recording, which resulted in a request to stop using the device. This suggests that there were activities that the algorithm had not yet adequately identified, likely because the distribution of all possible outpatient behaviors was not well represented in the training set. Because this user's FAR was so high, we have included in Table 3 the adjusted metrics for classifying this user as an outlier (metrics in italics). While the micro-FAR only decreased by 0.03 / day (0.14 / day → 0.11 / day), the macro-FAR decreased from 0.97 to 0.09 / day, confirming the large impact on FAR from one beta tester. The PPV for the outpatient dataset was 0.22 [0.11, 0.37], and the detector performed approximately one TP for every three FAs.
[0059] Table 3. Performance characteristics of the seizure monitoring application partitioned by outpatient and EMU environments and their respective facilities. The outpatient environment is partitioned into outpatients (OUT) and non-control users (NC; outpatient users without epilepsy). The EMU environment is partitioned into four separate hospital EMUs: Johns Hopkins Adult (JHA), Johns Hopkins Pediatric (JHP), Ruber International (RBI), and La Bonheur Children's (LBH).
[0060] [Table 3]
[0061] FIG. 4 shows the distribution of individual FAR rates for different patients, partitioned by EMU and outpatient settings, according to an example of the present disclosure.
[0062] Latency tests were performed on all captured seizure data by determining the difference between the seizure behavior onset time and the detection time. These latencies were recorded in real time whenever the algorithm generated a detection during testing, using the timestamp of the detection captured directly in the backend. The resulting mean and standard deviation of the latencies was 37.38 seconds (13.24 seconds) for outpatients and 32.07 seconds (10.22 seconds) for EMU patients. The range of latencies was [22 seconds-67 seconds] for outpatients and [20 seconds-57 seconds] for EMU patients. In Figure 4, a portion of the ACC signal during a seizure is shown from the detection time to disappearance to illustrate the latency of seizure detection relative to seizure onset.
[0063] FIG. 5 illustrates a portion of tonic-clonic seizures (TCS) detected during a prospective study centered on the time of detection according to an example of the present disclosure.
[0064] To facilitate the prevention of SUDEP and increase the overall control of patients' epilepsy, TCS monitoring should have a PPA so high and latency so small that caregivers can administer assistance in time and prevent potentially life-threatening consequences. To promote consistent and continuous use of the monitor in the daily life of epilepsy patients, it is equally important that the monitoring does not give frequent false alarms and create alarm fatigue for users and caregivers. Surveys among epilepsy patients and their caregivers have shown interest in standalone, non-EEG-based, multifunctional devices that can be worn without the risk of stigma. With these in mind, we aimed to evaluate the performance of TCS monitoring in EMU and outpatient settings using a custom application (EpiWatch) developed for a smartwatch.
[0065] Common metrics used to evaluate seizure detection devices are PPA, FAR, PPV, and latency. These are also metrics commonly evaluated by the FDA when determining if a particular device will be cleared for seizure detection. The algorithm demonstrated a perfect PPA of 1.0 in both outpatient (10 TCS) and inpatient (19 TCS) studies. This is generally similar to the performance of other commercially available seizure detection devices, but many of the larger studies have been able to test the algorithm against seizures with a larger sample size. Due to the nature of SUDEP, it is important not only to detect a seizure, but to do so quickly, as SUDEP is preventable if assistance is provided within less than a minute of the seizure ending. Assuming most TCS are around 70 seconds in length, if it is detected within 50 seconds of seizure onset, caregivers should be alerted and have enough time to provide assistance. The mean latency of the current algorithm is 32 seconds (SD = 10.22 seconds) for EMU seizures and 37 seconds (SD = 13.24 seconds) for outpatient seizures, which is similar to the performance of currently commercially available FDA-cleared seizure detectors and is fast enough to alert caregivers and prevent SUDEP. Some seizures still have longer latencies, with the slowest detected seizures being 67 seconds for EMU seizures and 57 seconds for outpatient seizures. It may be possible to fine-tune the operating characteristics of the algorithm to further reduce latency, but this may come at the cost of higher FARs and reduced adherence.
[0066] In general use, it is important that both caregivers and patients respond appropriately to alarms. A seizure detector with too many false alarms will cause alarm fatigue, leading to a loss of confidence in the device's effectiveness and reduced adherence to monitoring. The algorithm tested in this study showed a good FAR of 0.05 / day in EMUs and 0.13 / day in outpatients, translating to approximately 1 false alarm per 20 days and 1 false alarm per 7.5 days, respectively. The FAR rate was higher for outpatients as expected, with most false alarms reportedly caused by activities such as mowing the lawn, starting a motor outdoors, and playing drums. The macro-FAR for EMUs also differed between adults (0.07 / day) and children (0.03 / day). This is likely due to the training distribution being well tuned to the behavior of pediatric patients, despite them being more active than adult patients.
[0067] FIG. 6 illustrates a method 600 of physiological event detection and alert according to an example of the present disclosure. The method 600 includes obtaining a set of biosensor data from a user from one or more biosensors, as at 605. In some examples, the one or more biosensors comprise one or more of an accelerometer, a photoelectric (PPG) sensor, a gyroscope, a microphone, a blood oxygen level sensor, a blood pressure sensor, a blood glucose sensor, an eye sensor, an electrodermal activity sensor, an eye gaze sensor or tracker, a pupillometer sensor, or a combination thereof. In some examples, the one or more biosensors are incorporated into a wearable device consisting of a watch, a cuff, a necklace, a bracelet, glasses, a headset, one or more finger rings, or a combination thereof.
[0068] The method 600 further includes generating, by one or more hardware processors, a set of processed biosensor data from the set of biosensor data, as at 610. In some examples, the set of features generated from the processed biosensor data uses techniques including one or more of manual feature extraction, automatic feature extraction, or a combination thereof. In some examples, the manual feature extraction includes one or more of time domain feature extraction, frequency domain feature extraction, or a combination thereof. In some examples, the time domain feature extraction includes one or more of line crossing, variance, skewness, kurtosis, or a combination thereof. In some examples, the frequency domain feature extraction includes one or more of Fan-Chirp transform, Fourier transform, Chirp-Z transform, constant-Q transform, wavelet transform, or a combination thereof. In some examples, the automatic feature extraction includes one or more of one or more deep learning methods and one or more convolutional neural networks. In some examples, the one or more hardware processors comprise a first processor in a first device worn on the head or face and a second processor in a second device worn on the wrist or other part of the body, In some examples, the first head- or face-worn device is a pair of glasses and the second head- or face-worn device is a watch.
[0069] In some examples, the pre-processed biodata set includes biodata filtered for noise reduction and interpolation. In some examples, the processed biosensor data set may be processed as shown in FIG. 7 as follows. In some examples, the method 700 may further include processing the biosensor data set to create a processed biosensor data set, as at 705. The method 700 may also include reducing a data set imbalance between physiological events and non-physiological events in the processed biosensor data by iteratively training and using one or more models to identify abnormal segments in the biosensor data of non-physiological events to create a balanced data set, as at 710, the one or more models including one or more anomaly detection methods. The method 700 may also include training one or more classifiers for each characteristic physiological event phase, as at 715, using the balanced data set to create a confidence score for each characteristic physiological event phase.
[0070] In some examples, the one or more anomaly detection methods include one or more of a separation forest, a one-class support vector machine (SVM), a hidden Markov model (HMM), an autoencoder, a variational autoencoder, cluster-based outlier detection, or a combination thereof.
[0071] Returning to FIG. 6 , method 600 further includes generating, by a hardware processor, a set of features from the processed biosensor data, as at 615, associated with one or more characteristic physiological event phases. In some examples, each characteristic physiological event phase includes an event that causes a characteristic biosignal pattern associated with all or a portion of the physiological event, where the characteristic biosignal pattern includes one or more of tonic movements and / or associated physiological changes, clonic movements and / or associated physiological changes, postictal movement suppression or impairment and / or associated physiological changes, prodromal movements and / or associated physiological changes, early ictal movements and / or associated physiological changes, late ictal movements and / or associated physiological changes, ictal screaming and / or associated physiological changes, specific automatisms including one or more of hand tremors, tremors, paroxysmal eye blinks or gaze, saccades, fixations, noises, movement arrest, or specific physiological responses including one or more of heart rate changes or blood pressure changes.
[0072] Method 600 further includes determining, using one or more hardware processors, from the set of generated features, the set of processed biosensor data, or both the set of generated features and the processed biosensor data, for each of the one or more characteristic physiological event phases, a confidence score indicative of the presence of that phase in the data partition, as at 620. In some examples, the confidence score for each characteristic phase is calculated using a classifier, including classical techniques, including one or more linear models, one or more tree-based techniques, one or more clustering techniques, one or more probabilistic graphical models, one or more deep learning models, or a combination thereof.
[0073] The method 600 further includes, as at 625, using one or more hardware processors to determine a final confidence score suggesting the occurrence of a physiological event based on the relationship between the confidence scores of all the physiological event phases from the relationship between the confidence scores of each characteristic physiological event phase. In some examples, the relationship between the confidence scores to determine the final confidence score includes a technique for aggregating confidence scores, including one or more of: one or more non-temporal techniques analyzing a single time point, one or more classical temporal techniques analyzing multiple past time points, one or more deep learning techniques, or a combination thereof. In some examples, the non-temporal technique includes one or more of an average, a weighted average, a confidence score formula, or a combination thereof. In some examples, the temporal technique includes a probabilistic graphical technique. In some examples, the probabilistic graphical technique includes one or more hidden Markov models, one or more conditional random fields, or both. In some examples, the deep learning techniques include one or more of a recurrent neural network, a long short-term memory network, a gated recurrent unit network, a temporal convolutional network, a convolutional neural network, a multi-layer perceptron, or a combination thereof.
[0074] The method 600 further includes, as at 630, determining, using one or more hardware processors, from an accumulation of the final confidence scores, a cumulative confidence score suggestive of an occurrence of a particular physiological event, the physiological event being comprised of one or more characteristic physiological event phases. In some examples, the accumulation of the final confidence scores to generate the cumulative confidence score includes one or more of a low pass filter and a temporal modeling technique. In some examples, the temporal modeling technique includes one or more of a hidden Markov model, a conditional random field, a recurrent neural network, a long short-term memory network, a gated recurrent unit network, a temporal convolutional network, a convolutional neural network, or a combination thereof.
[0075] The method 600 further includes providing, by one or more hardware processors, a warning of the potential physiological event based on the cumulative confidence score, as at 635. In some examples, the warning of the potential physiological event is provided on a user interface of a wearable device worn by a user. The warning of the potential physiological event is provided to one or more of a user, a caregiver, a health care provider, or a legal guardian. A physiological event is any event that causes one or more characteristic patterns identifiable through one or more biosensors, including epileptic seizures, syncope, psychogenic non-epileptic seizures, movement disorders, or combinations thereof. In some examples, the physiological event includes a neurological event, a cardiac event, or combinations thereof. In some examples, the neurological event is a seizure.
[0076] 6 and 7 show example blocks of processes 600 and 700, although in some implementations processes 600 and 700 may include additional blocks, fewer blocks, different blocks, or blocks in a different arrangement relative to the blocks depicted in Figures 6 and 7, respectively. Additionally or alternatively, two or more of the blocks of processes 600 and 700 may be performed in parallel.
[0077] FIG. 8 shows data from a tonic-clonic seizure according to an example of the present disclosure. The top diagram is the X, Y, Z accelerometer traces, and the heart rate (BPM) calculated through PPG. The middle diagram is the model output of the tonic and clonic classifier. Note the overlap and the presence of multiple clonic phases in the TCS. Also note the method of training the classifier, the length of the window of data used for classification (20 seconds), and the fact that the classifier has a lag because the classification is done causally. Dense dot textured regions indicate tonic phases, and sparse dot textured regions indicate clonic phases. The bottom diagram is a cumulative filter with a preset warning threshold. The threshold can be set flexibly, and a lower threshold may result in more false positive warnings (meaning that a seizure is not occurring, but the method has classified the user's activity as a seizure). For example, the threshold can be set to about 0.5 to about 0.7.
[0078] FIG. 9 shows clonic-only (no TCS) data according to an example of the present disclosure. The top panel is the X, Y, Z accelerometer traces and heart rate (BPM) calculated through PPG. The middle panel is the model output of the tonic and clonic classifiers. The bottom panel is the cumulative filter. Note the overlap and that there are multiple clonic phases within the TCS. Also note the way the classifiers are trained, the length of the window of data used for classification (20 seconds), and that the classification is done causally, so there is a lag in the classifier. Note that there is no activation by the tonic phase classifier.
[0079] FIG. 10 shows data from an example exercise with predominantly tonic phases (not TCS) according to an example of the present disclosure. The top panel shows the X, Y, Z accelerometer traces and the heart rate (BPM) calculated through PPG. The middle panel shows the model output of the tonic and clonic classifiers. The bottom panel shows the accumulation filter. Note the overlap and that there are multiple clonic phases within the TCS. Also note how the classifier is trained, the length of the window of data used for classification (20 seconds), and that the classification is done causally, so there is a lag in the classifier. There is both tonic and clonic activation, but it does not occur in the right way (accumulation never reaches the threshold) to detect this interval as a seizure.
[0080] FIG. 11 shows data from a long single phase according to an example of the present disclosure. The top panel is the X, Y, Z accelerometer traces and the heart rate (BPM) calculated through PPG. The middle panel is the model output of the tonic and clonic classifiers. The bottom panel is the accumulation filter. Note that in this case, the accumulation filter only asymptotically approaches 0.5 even though the phase is activated for more than a minute. It never meets the threshold.
[0081] 12 shows data from a focal onset bilateral tonic-clonic seizure (FBTCS) according to an example of the present disclosure. The figure shows X, Y, Z accelerometer traces and heart rate (BPM) calculated through PPG. As shown in the example, the focal seizure lasts from 02:02:00 to 02:03:50, after which the actual TCS begins.
[0082] In some examples, in routine use, the disclosed method may use a refractory period of 10 minutes after detection during which no other detections occur. This refractory period exists to ensure that a single event does not trigger multiple alerts. If a patient's seizure is detected and a caregiver comes to provide assistance, there is a minimal reduction in safety since if a second seizure occurs within 10 minutes, the caregiver will be present at that time. The detector may be trained using a causality window that looks at a certain number of time points in the past. There is often a lag, especially with regard to the end of the clonic phase. This is due to how the model is trained, the causality of the window, and the length of the window. This combines and manifests as a disappearance in the detector output. This is clearly seen in Figures 8 and 9. In some examples, Figure 10 shows an example that illustrates the need for both phases for the detection of a tonic-clonic seizure. Note that only the tonic phase is activated and the seizure likely only approaches but does not cross the selected threshold.
[0083] In some embodiments, any of the disclosed methods may be performed by a computing system. FIG. 13 shows an example of such a computing system 1300, according to some embodiments. The computing system 1300 may include a computer or computer system 1301A, which may be an individual computer system 1301A or an array of distributed computer systems. The computer system 1301A includes one or more analysis modules 1302 configured to perform various tasks according to some embodiments, such as one or more methods disclosed herein. To perform these various tasks, the analysis modules 1302 execute independently or in cooperation with one or more processors 1304 connected to one or more storage media 1306. The processor 1304 is also connected to a network interface 1307 to enable the computer system 1301A to communicate with one or more additional computer systems and / or computing systems, e.g., 1301B, 1301C, and / or 1301D, via a data network 1309. (Note that the computer systems 1301B, 1301C, and / or 1301D may or may not share the same architecture as the computer system 1301A and may be located in different physical locations. For example, the computer systems 1301A and 1301B may be located in a processing facility while communicating with one or more computer systems, such as 1301C and / or 1301D, located in one or more data centers and / or located in various countries on different continents.)
[0084] A processor may include a microprocessor, microcontroller, a processor module or subsystem, a programmable integrated circuit, a programmable gate array, or other control or computing device.
[0085] Storage medium 1306 can be implemented as one or more computer-readable or machine-readable storage media. Storage medium 1306 can be connected or coupled to physiological interpretation machine learning module 1308. Note that while in the exemplary embodiment of FIG. 13 storage medium 1306 is depicted as being within computer system 1301A, in some embodiments storage medium 1306 can be distributed within and / or across multiple internal and / or external enclosures of computing system 1301A and / or additional computing systems. The storage medium 1306 may include one or more different forms of memory, including dynamic or static random access memory (DRAM or SRAM), semiconductor memory devices such as erasable and programmable read-only memory (EPROM), electrically erasable and programmable read-only memory (EEPROM) and flash memory, magnetic disks such as fixed disks, floppy disks, and removable disks, other magnetic media including tape, optical media such as compact disks (CDs) or digital video disks (DVDs), BLURAY® disks, or other types of optical storage devices, or other types of storage devices. It should be noted that the instructions described above may be provided on one computer-readable or machine-readable storage medium, or may be provided on multiple computer-readable or machine-readable storage media distributed within a larger system, possibly having multiple nodes. Such computer-readable or machine-readable storage media are considered to be part of an article (or article of manufacture). An article or article of manufacture refers to a single component or multiple components that are manufactured. The storage medium may be located within the machine that executes the machine-readable instructions, or may be located at a remote location where the machine-readable instructions can be downloaded over a network for execution.
[0086] It should be understood that computing system 1300 is but one example of a computing system and that computing system 1300 may have more or fewer components than shown, may combine additional components not depicted in the exemplary embodiment of Figure 13, and / or computing system 1300 may have a different configuration or arrangement of the components depicted in Figure 13. The various components depicted in Figure 13 may be implemented in hardware, software, or a combination of both hardware and software, including one or more signal processing and / or application specific integrated circuits.
[0087] Furthermore, the steps of the processing methods described herein may be implemented by operating one or more functional modules in an information processing device, such as a general-purpose processor or an application-specific chip, such as an ASIC, an FPGA, a PLD, or other suitable device. These modules, combinations of these modules, and / or combinations with general hardware are all within the scope of protection of the present invention.
[0088] An exemplary embodiment of a method for physiological event detection and alerting is shown in Figure 14, where a training and inference method 1400 is shown, where the physiological event is an epileptic seizure. The method includes sub-methods for automatically detecting an epileptic seizure and alerting a caregiver.
[0089] The imbalance reduction sub-method 1410 can identify seizure-like categories and reduce imbalance in the dataset. Datasets containing information related to seizure events are prone to imbalance due to the rare and unexpected nature of seizure events, resulting in a low rate of seizure events relative to the rate of non-seizure events. In one embodiment, the imbalance reduction sub-method 1410 can include iteratively training an unsupervised anomaly detection classifier and performing inference on the dataset to identify difficult-to-classify non-seizure categories. In one embodiment, the anomaly detection method can include a one-class support vector machine (OCSVM) technique. In another embodiment, the anomaly detection method can include a support vector data description (SVDD) technique. In another embodiment, the anomaly detection method can include an extended separation forest (IF) technique. In an exemplary embodiment, the anomaly detection method can include a separation forest (IF) technique. In another embodiment, the anomaly detection method can include a combination of OCSVM, SVDD, extended IF, and / or IF. The one or more anomaly detection classifiers can significantly reduce the imbalance in the dataset and enable the use of a supervised classifier. The imbalance reduction sub-method 1410 may include an iterative sub-method. The imbalance reduction sub-method 1410 may output a balanced data set.
[0090] The time domain sub-method 1420 may use the balanced data set that is the output of the imbalance reduction sub-method 1410 to explicitly or implicitly generate time domain features that are useful for identifying features of the input bio-signals of characteristic phases of epileptic seizures. This time domain sub-method 1420 may include manual feature extraction methods or machine learning or deep learning methods that implicitly identify time domain features. There are characteristic features that occur in epileptic seizures, but they are not limited to epileptic seizures and may also occur in non-seizure segments. In one embodiment, the time domain sub-method 1420 may include deep learning feature extraction methods to identify characteristic implicit features of the tonic phase of epileptic seizures from input bio-signals such as movement (from data obtained from an accelerometer) and heart rate (from data obtained from a PPG). However, the tonic phase may also occur within non-seizure segments. The time domain sub-method 1420 may include an iterative sub-method.
[0091] The spectral domain sub-method 1430 may use the balanced data set that is the output of the imbalance reduction sub-method 1410 to explicitly or implicitly identify spectral domain features that are useful for identifying features of the input biosignal that are characteristic of a phase of an epileptic seizure. This spectral domain sub-method 1430 may include manual feature extraction methods or machine learning or deep learning methods that implicitly identify the spectral domain features. There are characteristic features that occur in epileptic seizures, but they are not limited to epileptic seizures and may also occur in non-seizure segments. In one embodiment, the spectral domain sub-method 1430 may include a Fan-Chirp transform to identify a descending chirp from the input biosignal. This descending chirp may be characteristic of the clonic phase of an epileptic seizure. However, descending chirps may also occur within non-seizure segments. The spectral domain sub-method 1430 may include an iterative sub-method.
[0092] The characteristic phase sub-method 1440 may include a characteristic phase classifier that may be trained based on the output of the time domain sub-method 1420 and the spectral domain sub-method 1430 to identify characteristic phases of an epileptic seizure. There is no limit to the number of characteristic phases and therefore characteristic phase classifiers that may be present in this characteristic phase sub-method 1440. In one embodiment, each characteristic phase classifier may output a confidence value corresponding to whether the characteristic phase is present in a segment of data. The output confidence value may determine whether a given segment is a seizure or not, but not necessarily. For example, a non-seizure segment may have a high confidence value in one or more characteristic phases. The characteristic phase sub-method 1440 may describe a classifier that is implemented as any classification method (i.e., supervised, unsupervised, or rule-based). In one embodiment, there may be two characteristic phases identified in an epileptic seizure, a tonic phase and a clonic phase. The classifiers for both these phases may be implemented as deep learning models. For a tonic phase classifier, the characteristic phase sub-method 1440 may be trained end-to-end with the time domain sub-method 1420 and the spectral domain sub-method 1430. For a clonic phase classifier, the characteristic phase sub-method 1440 may be trained independently from the output from the time domain sub-method 1420 and the spectral domain sub-method 1430. The characteristic phase sub-method 1440 may include an iterative sub-method.
[0093] The aggregation sub-method 1450 may include aggregating the output from each characteristic phase classifier listed in the characteristic phase sub-method 1440 in the form of an aggregate confidence score for each characteristic phase classifier. The multiple characteristic phases in the characteristic phase sub-method 1440 may ensure that the detector captures a time segment that includes characteristics specific to epileptic seizures as opposed to other movements. In one embodiment, the aggregation sub-method 1450 may be implemented as an average of the confidence value outputs from the characteristic phase sub-method 1440. In another embodiment, the aggregation sub-method 1450 may be implemented as a weighted sum of the confidence value outputs from the characteristic phase sub-method 1440. The weights may be calculated by performing a grid search over all possibilities (e.g., in 10% increments) and simulating the guesswork on the time series data (including seizures and non-seizures). The weights that result in the best performance are selected. In one non-limiting example, the weights that result in the best performance are 50% for the tonic phase and 50% for the clonic phase. In another embodiment, the aggregation sub-method 1450 can be implemented as an arithmetic expression of the confidence values output from the characteristic phase sub-method 1440. In another embodiment, the aggregation sub-method 1450 can be implemented as a probabilistic graphical model, such as a hidden Markov model, a conditional random field, or both. In another embodiment, the aggregation sub-method 1450 can be implemented as one or more of an average, a weighted sum, an arithmetic expression, or a probabilistic graphical model. For example, in a probabilistic graphical method, similar to a deep learning method, the confidence output over time from the phase detector can be used as an input to a PGM or DL model, which identifies the temporal characteristics associated with tonic-clonic seizures and the phase confidence over time to make a decision.
[0094] An accumulation sub-method 1460 may accumulate aggregate confidence scores from the characteristic phase classifiers so that highly confident transient segments do not trigger detection prematurely. In one embodiment, this accumulation sub-method 1460 may include a first-order infinite impulse response (IIR) filter. For example, an IIR filter may have the following formula: y(n)=a * y(n-1)+b * x(n) is a low pass filter that can be implemented in a real-time recursive manner by using a=0.95, b=0.05 values are selected here. In another embodiment, the accumulation sub-method 1460 may include any accumulation method, including deep learning techniques such as recurrent neural networks, long short-term memory networks, gated recurrent unit networks, temporal convolutional networks, convolutional neural networks, multi-layer perceptrons, or combinations thereof. For example, the deep learning techniques can be configured to be similar to the scenario where PGM for confidence output is used. The input to the DL (or PGM) model can be the confidence output of the previous stage (aggregation), and the output value over a certain time window can be the input to the model. The model creates a temporal association between the aggregated time and seizure detection.
[0095] The alert sub-method 1470 may include an alert system that is triggered when the output of the accumulation sub-method 1460 reaches a particular value, also known as a trigger value or threshold. In one embodiment, the alert system may provide an alert of a potential physiological event on a user interface of a wearable device worn by the user. In another embodiment, the potential physiological alert may be provided to the user, a caregiver, a health care provider, a legal guardian, or a combination thereof.
[0096] Physiological interpretations, models, and / or other interpretation aids may be improved in an iterative manner. This concept is applicable to the method embodiments discussed herein. This may include the use of feedback loops, implemented algorithmically, such as in a computing device (e.g., computing system 1300, FIG. 13), and / or through manual control by a user who may make a decision as to whether a given set of steps, actions, templates, models, or curves has become sufficiently accurate for the evaluation of the signal under consideration.
[0097] The foregoing description has been described with reference to specific embodiments for purposes of illustration. However, the illustrative discussion above is not intended to be exhaustive or to limit the invention to the precise form disclosed. Many modifications and variations are possible in light of the above teachings. Furthermore, the order in which the elements of the method are illustrated and described may be rearranged and / or two or more elements may occur simultaneously. The embodiments have been chosen and described in order to best explain the principles of the invention and its practical application, so as to enable those skilled in the art to best utilize the invention and various embodiments with various modifications as appropriate for the particular use contemplated.
Claims
1. 1. A method for physiological event detection and alerting, comprising: obtaining a set of biosensor data from a user from one or more biosensors; generating, by one or more hardware processors, a set of processed biosensor data from the set of biosensor data; generating, by the one or more hardware processors, a set of features from the processed biosensor data, the set of features being associated with one or more characteristic physiological event phases, wherein data based on the association between the set of features and the one or more characteristic physiological event phases is stored in one or more non-transitory storage media; determining, using the one or more hardware processors, for each characteristic physiological event phase of the one or more characteristic physiological event phases from the generated set of features, the processed set of biosensor data, or both the generated set of features and the processed biosensor data, a confidence score indicative of a presence of that phase within a data partition; determining, using the one or more hardware processors, from a relationship between the confidence scores of each characteristic physiological event phase, a final confidence score indicative of an occurrence of a physiological event based on a relationship between the confidence scores of one or more of the physiological event phases; determining, using the one or more hardware processors, from an accumulation of final confidence scores, a cumulative confidence score indicative of an occurrence of a particular physiological event, the physiological event including one or more characteristic physiological event phases; and providing, by the one or more hardware processors, a warning of a potential physiological event based on the cumulative confidence score.
2. The method of claim 1 , wherein the one or more biosensors include an accelerometer and a photoelectric (PPG) sensor.
3. The method of claim 1 , wherein the one or more biosensors are incorporated into a wearable device consisting of a watch or glasses.
4. 10. The method of claim 1 further comprising: processing the set of biosensor data to create the set of processed biosensor data; reducing a dataset imbalance between physiological and non-physiological events in the processed biosensor data by iteratively training and using one or more models to identify anomalous segments in the biosensor data of non-physiological events to create a balanced dataset, the one or more models including one or more anomaly detection methods; and training one or more classifiers for each characteristic physiological event phase using the balanced data set to generate the confidence score for each characteristic physiological event phase.
5. 10. The method of claim 1, wherein the set of features generated from the processed biosensor data uses techniques including one or more of time domain feature extraction or frequency domain feature extraction.
6. 2. The method of claim 1, wherein the relationship between the confidence scores that determines the final confidence score includes a technique for aggregating the confidence scores, including an average of the confidence scores, a weighted sum of the confidence scores, a formula for the confidence scores, a probabilistic graphical model, or a combination thereof.
7. The method of claim 1 , wherein the accumulation of the final confidence scores to generate the cumulative confidence score comprises a first-order infinite impulse response (IIR) filter.
8. The method of claim 1 , wherein the warning of the potential physiological event is provided on a user interface of a wearable device worn by the user.
9. The method of claim 1 , wherein the alert of the potential physiological event is provided to one or more of the user, a caregiver, a health care provider, or a legal guardian.
10. 2. The method of claim 1, wherein a physiological event is any event that causes one or more characteristic patterns discernible through one or more of the biosensors, the events including epileptic seizures, syncope, psychogenic non-epileptic seizures, movement disorders, or combinations thereof.
11. The method of claim 1 , wherein the physiological event comprises a neurological event, a cardiac event, or a combination thereof.
12. 12. The method of claim 11, wherein the neurological event is a seizure.
13. 1. A system for physiological event detection and alerting, comprising: one or more biometric sensors that capture, record, or both capture and record biometric sensor data from a user; one or more hardware processors; and one or more non-transitory computer-readable media storing instructions that, when executed by the one or more hardware processors, implement a physiological detection and alert method; The physiological detection and alert method includes: obtaining a set of biosensor data from a user from one or more biosensors; generating, by one or more hardware processors, a set of processed biosensor data from the set of biosensor data; generating, by the one or more hardware processors, a set of features from the processed biosensor data, the set of features being associated with one or more characteristic physiological event phases, wherein an association between the set of features and the one or more characteristic physiological event phases is stored in one or more non-transitory storage media; determining, using the one or more hardware processors, for each characteristic physiological event phase of the one or more characteristic physiological event phases from the generated set of features, the processed set of biosensor data, or both the generated set of features and the processed biosensor data, a confidence score indicative of a presence of that phase within a data partition; determining, using the one or more hardware processors, from a relationship between the confidence scores of each characteristic physiological event phase, a final confidence score indicative of an occurrence of a physiological event based on a relationship between the confidence scores of one or more of the physiological event phases; determining, using the one or more hardware processors, from an accumulation of final confidence scores, a cumulative confidence score indicative of an occurrence of a particular physiological event, the physiological event being comprised of one or more characteristic physiological event phases; providing, by the one or more hardware processors, a warning of a potential physiological event based on the cumulative confidence score; and providing a user interface that provides an alert of the potential physiological event.
14. The system of claim 13 , wherein the one or more biosensors include an accelerometer and a photoelectric (PPG) sensor.
15. The system of claim 13 , wherein the one or more biometric sensors are incorporated into a wearable device consisting of a watch or glasses.
16. 14. The system of claim 13, the method further comprising: processing the set of biosensor data to create the set of processed biosensor data; reducing a dataset imbalance between physiological and non-physiological events in the processed biosensor data by iteratively training and using one or more models to identify anomalous segments in the biosensor data of non-physiological events to create a balanced dataset, the one or more models including one or more anomaly detection methods; and training one or more classifiers for each characteristic physiological event phase using the balanced data set to generate the confidence score for each characteristic physiological event phase.
17. 14. The system of claim 13, wherein the set of features generated from the processed biosensor data uses techniques including one or more of time domain feature extraction or frequency domain feature extraction.
18. 14. The system of claim 13, wherein the relationship between the confidence scores to determine the final confidence score includes a technique for aggregating the confidence scores including an average of the confidence scores, a weighted sum of the confidence scores, a formula for the confidence scores, a probabilistic graphical model, or a combination thereof.
19. The system of claim 13 , wherein the accumulation of the final confidence scores to generate the cumulative confidence score comprises a first-order infinite impulse response (IIR) filter.
20. The system of claim 13 , wherein the warning of the potential physiological event is provided on a user interface of a wearable device worn by the user.
21. The system of claim 13 , wherein the alert of the potential physiological event is provided to one or more of the user, a caregiver, a health care provider, or a legal guardian.
22. 14. The system of claim 13, wherein a physiological event is any event that causes one or more characteristic patterns discernible through one or more of the biosensors, these events including epileptic seizures, syncope, psychogenic non-epileptic seizures, movement disorders, or combinations thereof.
23. The system of claim 13 , wherein the physiological event comprises a neurological event, a cardiac event, or a combination thereof.
24. 24. The system of claim 23, wherein the physiological event is a seizure.
