Establishing blood parameters based on motion induced by cardiac activity

Abstract

Disclosed is a computing device for establishing blood parameters. The computing device comprises a processing component configured to process at least one signal representing motion induced by cardiac activity, wherein the at least one signal is processed using a trained model to establish one or more blood parameters. Furthermore, a system and a computer implemented method are described.

Description

[0001] ESTABLISHING BLOOD PARAMETERS BASED ON MOTION INDUCED BY

[0002] CARDIAC ACTIVITY

[0003] TECHNICAL FIELD

[0004] The present invention relates to a computing device for establishing blood parameters. The present invention may also relate to a system for establishing blood parameters, and a corresponding computer-implemented method.

[0005] BACKGROUND

[0006] Blood analysis plays a crucial role in physiological applications. Traditional methods for blood parameter analysis often involve complex laboratory procedures, requiring specialized equipment and trained personnel. These methods can be time-consuming, expensive, and not always accessible, especially in remote or resource-limited settings.

[0007] Recent advancements in medical technology have led to the development of portable and user- friendly devices for blood analysis. However, these technologies often face limitations in terms of accuracy, reliability, and the range of parameters they can measure. Additionally, they may still require specific environmental conditions or external power sources, limiting their applicability.

[0008] Existing methods are typically invasive and often require significant volumes of blood, which can be a limiting factor in repeated testing, particularly for vulnerable populations like infants or patients with anemia. There is also a growing need for real-time or near-real-time data to make prompt decisions, which many current methodologies cannot adequately support.

[0009] Physiological monitoring refers to the continuous or periodic observation, measurement, and recording of physiological parameters, like heart rate, blood pressure, body temperature, respiratory rate, and oxygen saturation, and is frequently conducted in order to assess an individual's health status. Physiological monitoring can be critical in various settings, ranging from self-monitoring applications, such as in fitness apps and personal health tracking tools, to medical and clinical applications, such as in care units, during surgery, or for chronic disease management at home. Physiological monitoring can include the capture of seismocardiography signals. Seismocardiography is a technique used to measure vibrations caused by the heartbeat, typically by placing an accelerometer on the chest. These vibrations provide valuable information about cardiac function and can be used to assess heart health and detect abnormalities. Seismocardiograms capture an aspect of the heart's mechanical activity and are used to complement other monitoring techniques like electrocardiograms.

[0010] Furthermore, methods of machine learning are known. For example, convolutional autoencoders cam be applied in order to compress health data by reducing the complexity or noise in biological signals. In other examples, neural networks can be applied to analyze ECG data or other signals from various sensor types.

[0011] SUMMARY OF THE INVENTION

[0012] The invention focuses on a non-invasive approach for establishing blood parameters without the need for traditional invasive approaches, including blood draws, thereby enhancing patient comfort and compliance. This is solved by establishing blood parameters based on data or signals representing motion induced by cardiac activity.

[0013] The objective of the present invention is to provide a device and a system for establishing blood parameters, and a corresponding method, which improve physiological monitoring of the prior art.

[0014] In particular, the objective of the present invention is to introduce a non-invasive approach for accurately and efficiently determining blood parameters, which eliminates the need for blood draws, thereby reducing discomfort and risk of infection, further facilitate easy, pain-free monitoring, particularly beneficial for patients requiring regular testing. This may further enhance accessibility and convenience, allowing for self-monitoring and potentially reducing the need for frequent clinical visits. Another object of the present invention is to provide rapid and reliable results, enabling better disease management and patient outcomes. A first aspect of the invention provides a computing device for establishing blood parameters, comprising a processing component configured to process at least one signal representing motion induced by cardiac activity, wherein the at least one signal is processed using a trained model to establish one or more blood parameters.

[0015] The device of the first aspect establishes one or more blood parameters, wherein at least one cardiac activity-induced signal is captured. A cardiac activity-induced signal can refer to a signal that is caused by cardiac activity, such as cardiac movements. The at least one signal may include a plurality of signals representing the motion induced by the cardiac activity. This may include signals of different types. The at least one signal may be a continuous signal, a sampled time series of data points, or any other type of a signal or data stream, such as a wave-shaped time curve(s). The signals signal can be defined as a time -varying representation of the motion induced by cardiac activity. Signals may be captured as electrical, mechanical, or chemical fluctuations that correspond to various physiological functions.

[0016] The trained model may be a machine learning (ML) or artificial intelligence (Al) model, which trained to interpret or analyze signals representing motion induced cardiac activity. This may include applying Al techniques to extract meaningful insights, predict outcomes, or make decisions based on complex and often non-linear data patterns, which are defined by the cardiac activity-induced motion signals. In particular, the trained model is trained using the signals in to establish said one or more blood parameters.

[0017] In a first implementation of the device according to the first aspect, the processing component is further configured to filter the at least one signal. The processing component may implement a signal processing pipeline where the incoming physiological signal(s) are subjected to a filtering process. Filtering may be applied to remove unwanted components from the signal(s), such as noise or interference, which could be due to electrical, environmental, or biological sources. The goal is to enhance the quality and clarity of the signal for more accurate analysis and interpretation using the trained model. This may include various types of filters, such as low-pass filters, high-pass filters, or band-pass filters, in any number and combination. Furthermore, notch filters could be used to remove specific frequencies or a narrow band of frequencies. Filtering can be crucial because physiological signals may often be weak and can be easily corrupted by external noise or interference from other bodily signals. According to one example, a suitable filter for carrying out the filtering can be a Butterworth or polynomial filter. If the filtering includes a high-pass filter, a limit frequency of the high-pass filter can be in a range of 5 Hz to 8 Hz, for example, in order to reliably reduce effects of movement artifacts on the at least one signal. If the filtering includes a band-pass filter, a first limit frequency can be in a range of 5 Hz to 8 Hz and a further limit frequency in a range of 30 Hz to 35 Hz, for example, in order to reduce movement artifacts that are outside the range of 8 Hz to 30 Hz, for example. The filtering cam be carried out in particular by Butterworth filters or polynomial filters.

[0018] It is to be understood that ranges could be defined as including the range boundaries or excluding the numerical values of the range boundaries.

[0019] In a further implementation of the device according to the first aspect, the processing component is further configured to provide the at least one signal or the filtered at least one signal as input to the trained model, wherein the one or more blood parameters are establish based on an output of the trained model. When the processing component provides the cardiac activity- induces motion signal (either raw or filtered) as input to the trained model, the output of the model can vary depending on its design and purpose. The trained model can classify the signal into different categories. Additionally or as an alternative, the output could also be a continuous value or a set of values. Additionally or as an alternative, the trained model could output a signal indicating the presence of an anomaly or an unusual pattern. Additionally or as an alternative, trained model might also analyze and output trends over time. This could involve extrapolating from current signal data to predict future states. Additionally or as an alternative, the trained model's output might be a set of features or characteristics extracted from the cardiac activity- induced motion signal, which are then used for further analysis or interpretation. Thus, the output of the trained model can take various forms, from diagnostic classifications related to blood parameters to individual blood parameter values or continuous monitoring metrics indicating blood parameter. Accordingly, the output of the trained model can represent the one or more blood parameters or could be used to determine the one or more blood parameters.

[0020] In a further implementation of the device according to the first aspect, the output of the trained model represents a time series of blood parameters. The one or more blood parameters can be established based on the time series. Accordingly, The one or more blood parameters may represent an output signal of the one or more blood parameters. This may enable a continuous monitoring of the one or more blood parameters based on the cardiac activity-induced motion (input) signal(s). In a preferred embodiment, the at least one signal representing motion induced by cardiac activity may form the input variable and the one or more blood parameters may form the output variable of the processing carried out by the trained model. In particular, there may be no extraction of features from the at least one cardiac activity-induced motion signal, which then may form input variables of the processing of the trained model. Accordingly, unprocessed cardiac activity-induced motion signals may form the input variable of the processing of the trained model. The at least one cardiac activity-induced motion signal can also form the only input variable of the processing of the trained model, so that the trained model may rely on cardiac activity-induced motion signals only and does not require any further input variables for establishing the blood parameters.

[0021] In a further implementation of the device according to the first aspect, the at least one signal includes at least one of a seismocardiography signal, a phonocardiography signal, or a ballistocardiography signal. Seismocardiography (SCG), Phonocardiography (PCG), and Ballistocardiography (BCG) are examples of different types of signals that represent motion induced by cardiac activity, each capturing various aspects of the heart's function and motion. SCG may focus on the mechanical movements of the heart and chest wall, PCG on the sounds produced by the heart's activity, and BCG on the body's overall response to the heart's pumping action. The at least one signal may include a plurality of signals of the same type, which may be acquired using different sensors placed at various locations. The plurality of signals may also include at least two signals of a different type, such as an SCG signal and a BCG signal.

[0022] SCG records the vibrations induced by the heartbeat. These signals are generated by the mechanical motion of the heart and the consequent small movements of the chest wall. In terms of signal characteristics, SCG waveforms reflect the timing and intensity of various cardiac events, such as the opening and closing of heart valves and the contraction and relaxation of the heart muscle. These signals can provide information about heart rate, rhythm, and even some aspects of heart function.

[0023] PCG involves recording the acoustic signals produced by the heart, such as heart murmurs, the sounds of heart valves opening and closing, and other heart sounds. PCG signals are audi- tory in nature and can be captured by microphones placed on the chest. The resulting waveform provides information on the timing of the heart sounds and can be used to detect abnormalities in heart valves or other parts of the heart.

[0024] BCG measures the mechanical effect of the blood ejection from the heart and the resultant reaction forces on the body. It captures the body's subtle movements in response to the heartbeat. In terms of signal, BCG waveforms represent the recoil of the body as the heart pumps blood, providing information about the cardiac output and the mechanical efficiency of the heart.

[0025] In a further implementation of the device according to the first aspect, the one or more blood parameters include one or more of blood glucose levels, lipid profiles, and electrolyte levels. Blood glucose levels, lipid profiles, and electrolyte levels represent three examples types of blood parameters, which can be established using embodiments of the present disclosure. This may require a use of a trained model that has been trained for the respective (combination of) types of blood parameters. The one or more blood parameter may include (and the trained model may be trained for) parameters of the same type or parameters of a different type. As an example, blood glucose levels measures the amount of glucose (sugar) in the blood. It is a crucial indicator for diagnosing and monitoring diabetes. Normal ranges vary but typically, fasting blood sugar levels are considered normal when they are between 70 to 99 mg / dL. Levels above this may indicate pre-diabetes or diabetes. A lipid profile is a group of tests that measure different types of fats in the blood, mainly cholesterol and triglycerides. It typically includes one or more of total cholesterol, low-density lipoprotein (LDL), high-density lipoprotein (HDL), and / or triglycerides. Electrolytes are minerals in blood and other body fluids that carry an electric charge. Important electrolytes include sodium, potassium, calcium, bicarbonate, magnesium, chloride, and phosphate. Imbalances in electrolyte levels can lead to a variety of symptoms and can indicate issues such as dehydration, kidney disease, or other metabolic disorders. These blood parameters are often measured through invasive blood tests. The present invention enables a determination of blood parameters based on signal(s) representing motion induced by cardiac activity, which can be determined in a non-invasive manner.

[0026] In a further implementation of the device according to the first aspect, the at least one signal includes a time series of data points representing the motion. The time series of data points representing cardiac activity-induced motion may include a sequence of measurements captured over time. Each of these signals can be broken down in terms of their time series characteristics. For example, an SCG signal may be a time series of data points representing the mechanical vibrations caused by the heart's movement. Each data point may correspond to a specific moment in time and may reflect a magnitude of chest wall motion due to cardiac activity. The time series may capture a rhythm and intensity of the heart's mechanical movements. In PCG, for example, the time series may include data points representing an acoustic signals (sounds) produced by the heart. Each point in the time series may correspond to an acoustic measurement at a specific time interval. This time series can reveal the timing of heart sounds. A BCG signal may be a time series representing the body's movements in response to heartbeats. Each data point in this time series may correspond to the body's position or motion at a given moment, reflecting the recoil and other movements caused by the ejection of blood from the heart.

[0027] In a further implementation of the device according to the first aspect, the at least one signal includes at least one channel, each including a time series of data points representing the motion. Preferably, each channel may be linked to a different sensor. The term “channel” may refer to a single stream of data collected from one sensor or measurement point. For instance, in SCG, a channel might represent the vibrations measured at a specific location on the chest wall. In PCG, for example, a channel could correspond to acoustic data captured from a point on the chest, where different locations might pick up different aspects of heart sounds. BCG channels could represent the data from sensors capturing the body's motion in response to the heart's pumping action, possibly placed at different points on equipment, such as a bed or chair.

[0028] In a further implementation of the device according to the first aspect, the device further comprises a user interface configured to render a representation of the one or more blood parameters. The user interface can be used to continuously monitor the one or more blood parameters. Further, the device can include a graphics processing unit, which may be used to render the representation of the blood parameters on the user interface. The graphics processing unit can also be used to process the trained model. This may include performing calculations related to modules and components of the trained model in a highly efficient manner. For example, the graphics processing unit could be used to calculate convolution operations in a convolutional neural network. In a further implementation of the device according to the first aspect, the motion is induced by cardiac activity in a subject and the one or more blood parameters reflect blood conditions of the subject. Accordingly, the computing device may be configured for establishing blood parameters in the subject, wherein the processing component processes the at least one signal representing motion induced by cardiac activity in the subject, and the at least one signal is processed using the trained model to establish the one or more blood parameters of the subject. However, it is to be understood that the computing device can also operate on pre-recorded signals without directly processing signals measured on a subject.

[0029] In a further implementation of the device according to the first aspect, the at least one signal is detected using one or more sensors, wherein at least one of the one or more sensors is placed on a thorax of a subject or wherein at least one of the one or more sensors is configured to detect motion without contact to a subject. The sensors can be either in direct contact with the subject or placed remotely, which may also be referred to as remote physiological monitoring. It is to be understood that the trained model may require a placement of the sensors in relation to the subject. However, the present disclosure is not limited to a particular placement (and number) of sensors. Rather, for any kind placement and combination of sensors acquiring the signals representing motion induced by cardiac activity, which reflects the cardiac activity in a sufficient manner, the trained model can be adjusted and adapted to process such signals and enable an establishment of the blood parameters.

[0030] In a further implementation of the device according to the first aspect, the device further comprises at least one of the one or more sensors. For example, the device may be a portable or mobile device, such as a smartphone. Portable or mobile devices can be equipped with various sensors that can be used to detect motion induced by cardiac activity, primarily through the measurement of subtle movements or vibrations of the body. This may include accelerometers. Accelerometers may be placed, for example, in smartphones to measure acceleration forces acting on the smartphone, which can include subtle movements of the chest or body caused by heartbeat. By placing the portable or mobile device on the chest, the accelerometer may detect vibrations generated by the heartbeat. Additionally or as an alternative, gyroscope(s) can be used to determine the orientation of the portable or mobile device, and / or to detect motion. For example, when combined with accelerometer data, a gyroscope can help to differentiate between cardiac-induced motion and other types of movement, providing more accurate readings. Additionally or as an alternative, one or more microphones can be disposed in the portable or mobile device and used to capture a sound of the heartbeat. Portable or mobile devices can detect motion induced by cardiac activity primarily through accelerometers and / or gyroscopes, offering a convenient and accessible way to acquire the at last one signal. The device can be further connected to other sensors and measurement equipment or storage equipment to receive or retrieve captured or stored signals.

[0031] The device or at least one of the sensors may also directly contact a subject or may be arranged in the subject. For example, the device or sensor may be integrated in a pacemaker. Preferably, the pacemaker can comprise an acceleration sensor or any other type of sensor capable of detecting motion induced by cardiac activities. A signal detected by the sensor can be transmitted to the computing device, e. g., wirelessly or using an interconnect via suitable methods for data transmission. Additionally or as an alternative, the pacemaker may comprise a computing unit, wherein the latter then carries out the processing. Accordingly, the computing device can be integrated into the pacemaker in the form of an embedded system. Preferably, the computing device can be implemented as an integrated circuit configured to perform the processing of the computing device. The use of a computing device or sensor which is integrated into a pacemaker advantageously enables the use of sensors which are already present and are arranged close to the heart, which results in good and reliable signal quality of the cardiac activity-induced signals. This in turn improves measurement accuracy and thus also the accuracy of blood parameters established according to the invention.

[0032] Processing can be performed at least partially on the computing device, for example, using a (graphics) processing unit or a dedicated circuit, such as an neural network accelerator. However, in a further implementation of the device according to the first aspect, the device further comprises a communication component configured to transmit the at least one signal to a remote processing component, wherein the at least one signal is processed by the remote processing component. Accordingly, the computing device can be an acquisition device or any other type of portable device, which may have limited processing resources, and the processing can be performed on the remote processing component, such as on a remote computing device quipped with respective processing resources, on a server or in a distributed manner in a cloud. Accordingly, in this example, the computing device can be a terminal device or a client device, and the remote processing component may be provided on a server or in a cloud. In a further implementation of the device according to the first aspect, the device further comprises an interface configured to receive the at least one signal. The at least one signal may be retrieved, via the interface, from a database or another storage. The at least one signal may be acquired by a terminal device and transmitted to the computing device via a network or a similar interconnect. In this example, the computing device can be a remote computing device or a server, which may operate in a cloud-based environment and which may receive the signals from a terminal device or a client device.

[0033] In a further implementation of the device according to the first aspect, the trained model includes at least one trained neural network, including one or more of an autoencoder, a convolutional neural network (CNN), a recurrent neural network (RNN), long short-term memory network (LSTM), or a ID convolutional network (ID CNN). The trained model can be used to extract, analyze, and interpret complex patterns within the data of the at least one signal. Autoencoders, CNNs, RNNs, LSTMs, or ID CNN represent examples of neural networks, which can be used individually or in combination. For example, a hybrid model combining CNNs for feature extraction and LSTMs for sequence analysis might be used for highlighting both spatial and temporal patterns in cardiac signals. Autoencoders can be used for unsupervised learning, primarily for dimensionality reduction or feature learning. Regarding signals representing motion induced by cardiac activity, an autoencoder could be used to reduce the dimensionality of the signal data, making it easier to process, or to learn useful features from the raw data. CNNs can be used to identify patterns within the motion signals that might be indicative of blood parameter conditions. The convolutional layers in the CNN can detect localized patterns in the data, which may be crucial in understanding complex physiological signals and their influence on blood parameters. RNNs are designed to work with sequences of data, making them ideal for time-series analysis of the input motion signals. They can be used to process sequential data from SCG, PCG, or BCG, capturing the temporal dynamics and dependencies in the signal. This is particularly useful for analyzing the rhythmic patterns in cardiac signals and their influence on the blood parameters. LSTMs are a type of RNN. However, LSTMs are capable of capturing long-term dependencies in sequence data. They can be particularly effective in analyzing cardiac signals where understanding long-term patterns and relations (like changes in heart rhythm over time) may be crucial for establishing the blood parameters. As another example, ID CNNs represent a variant of convolutional neural networks that are well-suited for analyzing time-series data. They can efficiently handle large volumes of data, making them ideal for processing continuous and voluminous data generated by SCG, PCG, or BCG devices. Additionally or as an alternative, various mathematical algorithms for machine learning can be used. This may include decision tree-based methods, ensemble methods (e.g. boosting, random forest)-based methods, regression-based methods, Bayesian methods (e.g. Bayesian belief net- works)-based methods, kernel methods (e.g. support vector machines)-based methods, instance (e.g. k-nearest neighbor)-based methods, association rule learning-based methods, Boltzmann machine-based methods, artificial neural networks (e.g. perceptron)-based methods, deep learning (e.g. convolutional neural networks, stacked autoencoders)-based methods, dimensionality reduction-based methods, regularization methods-based methods.

[0034] The model may be formed as an autoencoder. An autoencoder advantageously requires less computational resources for processing. Hence the trained model can be implemented in a simplified form, for example, as an embedded systems or in portable terminals such as e.g. portable or mobile devices, such as smartphones, in a reliable and fast manner.

[0035] Implementing the trained model as a CNN further advantageously provides reduced complexity of the model suitable for devices with low computing power. This relates both to the training phase and to the application to signals (inference phase). The time required for training the model may be shorter for CNNs than, for example, with regard to LSTMs. However, LSTM networks may be well suitabile for the analysis of signals, since their architecture may take into account temporal dependencies.

[0036] In a further implementation of the device according to the first aspect, the trained model is generated by training a model with training data, the training data including recordings of signals, each representing motion induced by cardiac activity, and labels indicating one or more blood parameters associated with the respective signal. This may be further referred to as supervised learning. Other approaches can be used to trained a model for a particular combination of signal(s) and blood parameter(s). When creating a trained model for analyzing signals representing motion induced by cardiac activity (like SCG, PCG, or BCG) and correlating them with blood parameters, the process may typically involve several key steps. The training data set may include recordings of signals, such as SCG, PCG, or BCG and the like, in any combination, each capturing different aspects of cardiac-induced motion. Alongside these recordings, the dataset may also include labels indicating one or more of said blood parameters associated with each signal. The raw signal data and blood parameter labels can be pre-processed. This might involve normalizing the signal data, handling missing values, and converting the data into a format suitable for the chosen model. Preprocessing could also include filtering out noise, segmenting the data into relevant portions, and extracting features that could be informative for the training task. The model including a neural network architecture, like CNN, RNN, or LSTM, to name a few, individually or in any combination, can be trained using the prepaired dataset. During training, the model learns to map the input signals to the corresponding blood parameters. This process typically involves adjusting the model's parameters (weights and biases) using algorithms like backpropagation, guided by a loss function that measures the difference between the model's predictions and the actual labels. Once trained, the trained model can be used to process (and analyze) any cardiac motion signals, such as the at least one signal representing motion induced by cardiac activity, and compute an output that is used to establish the one or more blood parameters. The trained model may also undergo further training or refinement as more data becomes available or as requirements evolve. Accordingly, embodiments may include re-training the trained model using further training data.

[0037] In a further implementation of the device according to the first aspect, the training data is split into a training subset, a validation subset, and a testing subset, wherein the model is trained using the training subset, a performance of the model during training is assessed using the validation subset, and the performance of the trained model is evaluated using the testing subset. The model's performance may be evaluated using the separate validation subset not seen during training. This may be used to tune the train model and may be helpful in avoiding overfitting. A final test with another distinct testing subset can provide an unbiased evaluation of the model's performance. The bulk of the data is usually allocated to the training subset. This is the data the model is exposed to during the training phase. The model learns to map signals representing cardiac activity-induced motion to the associated blood parameters using the training subset. The learning process involves adjusting the model's internal parameters to minimize the difference between the model's predictions and the actual labels. The validation subset may be a portion of the data that the model does not train on. Instead, it's used to assess the model's performance and generalizability during the training process. By evaluating the model on the validation subset, potential issues like overfitting can be monitored, where the model performs well on the training data but poorly on unseen data. This subset is crucial for tuning the model's hyperparameters, like learning rate, number of layers, and the like. The testing subset can be another portion of data that the model hasn't seen during the training phase. It's used to evaluate the final performance of the trained model. This subset provides an unbiased assessment of how well the model generalizes to new, unseen data. The performance on the testing subset is often considered a good indicator of how the model will perform in real -world applications. Using subsets of training data for training, validation, and testing ensures that the model is not only accurate on the data it was trained on, including the training subset data, but can also generalize well to new, unseen data.

[0038] In a further implementation of the device according to the first aspect, the device is further configured to calculate an error function indicating a deviation between established one or more blood parameters and reference blood parameters to generate the trained model, wherein the established one or more blood parameters and / or the reference blood parameters and / or the deviation are weighted in the error function. This allows for a nuanced evaluation of the model's predictions, especially when dealing with motion signals and their correlation with blood parameters. The error function, which may also be referred to as a loss function, may quantify the deviation between the model's predictions (established blood parameters or data allowing to establish the blood parameters) and the actual values (reference blood parameters or respective data). This function may guide the optimization process. The model's parameters may be adjusted to minimize the error function. Weighting can involve giving different importance to various aspects of the error function. Weighting can be applied to the established blood parameters, reference blood parameters, or the deviation itself. For example, if certain blood parameters are more sensitive or critical for patient outcomes, the error associated with these parameters can be weighted more heavily in the loss function. Similarly, if the cost of overestimation is different from the cost of underestimation for a certain parameter, these deviations can be differently weighted. Incorporating the weighted error function into the model's training process ensures that the model not only becomes accurate in a general sense but also becomes sensitive and specific in relevant areas.

[0039] Training of the model, such as a neural network, requires a large amount of training data in order to ensure a desired quality of correlation between the input signal and the blood parameters. The amount of training data cam be dependent on factors such as the complexity of the underlying problem, the required accuracy and the desired adaptability of the network to be trained. The field of application, i.e., the domain in which the model is to be used, is often the most important element in the determination of these factors and thus the determination of the amount of training data. With corresponding previous knowledge about the domain, it is possible to prepare the training data in such a manner that they lead to a faster convergence for the optimal solution, or directly converge and thus require less training data.

[0040] The approach according to embodiments of the present disclosure could be used in medical environments, which may require a high accuracy of correlation and results. In addition, the complexity is high since blood parameters (values or respective signals) and cardiac activity- induced motion signals relate to sensors modalities. This typically requires an increased amount of data for training the model. One possible step for reducing the required amount of data consists in filtering the training data, in particular the input data and / or the output data. For example, input and output data of a training dataset can be generated by simultaneously generating cardiac activity- induced motion signals and blood parameter data (or signals) and then filtering them before training. The memory requirement and also the required resources, i.e., computing time and power, for building the trained model can be reduced. It is thus possible to filter the training data with a filter, in particular a bandpass filter, e.g. a Butterworth filter, in order to attenuate high-frequency and low-frequency components in the training data. For example, a first, lower limit frequency of a bandpass filter can be 0.5 Hz and a further, upper limit frequency can be 200 Hz. It is also conceivable to use high- and / or low-pass filters or other filters (e.g. polynomial filters) which filter out corresponding unwanted frequencies from the training data.

[0041] A second aspect of the present disclosure refers to a system, comprising at least one server; a network; and at least one client device connectable to the at least one server via the network, the at least one client device according to embodiments of the computing device of the first aspect. In particular, the client device can comprise a processing component configured to process at least one signal representing motion induced by cardiac activity, wherein the at least one signal is processed using a trained model to establish one or more blood parameters.

[0042] In a further implementation of the system according to the second aspect, the at least one server is configured to process the at least one signal using the trained model to establish the one or more blood parameters.

[0043] In a further implementation of the system according to the second aspect, the at least one server is configured to train the trained model. A third aspect of the invention provides a method for establishing blood parameters, comprising processing at least one signal representing motion induced by cardiac activity, wherein the at least one signal is processed using a trained model to establish one or more blood parameters.

[0044] The method may be a computer-implemented method. The method may be implemented in a computing device or on a system with one or more computing devices, including one or more processors or processing units (PUs) that may be coupled to each other or communicate with each other, and that may be configured to perform individual method steps. The one or more processors or PUs may implement hardware and / or software modules that execute the individual method steps.

[0045] In a first implementation of the method according to the third aspect, the method further comprises filtering the at least one signal.

[0046] In a further implementation of the method according to the third aspect, the at least one signal or the filtered at least one signal is provided as input to the trained model and the one or more blood parameters are establish based on an output of the trained model. Preferably, the output of the trained model may represent a time series of blood parameters, wherein the one or more blood parameters are established based on the time series.

[0047] In a further implementation of the method according to the third aspect, the at least one signal includes at least one of a seismocardiography signal, a phonocardiography signal, or a ballistocardiography signal.

[0048] In a further implementation of the method according to the third aspect, the one or more blood parameters include one or more of blood glucose levels, lipid profiles, and electrolyte levels.

[0049] In a further implementation of the method according to the third aspect, the at least one signal includes a time series of data points representing the motion. Preferably, the at least one signal includes at least one channel, each including a time series of data points representing the motion. In a further implementation of the method according to the third aspect, the method further comprises rendering a representation of the one or more blood parameters on a user interface.

[0050] In a further implementation of the method according to the third aspect, the method further comprises retrieving an identification of the subject and using the identification to adapt said processing to variations in the subject.

[0051] In a further implementation of the method according to the third aspect, the method further comprises transmitting the at least one signal to a computing device, wherein the at least one signal is processed by the computing device.

[0052] In a further implementation of the method according to the third aspect, the trained model is generated by training a model with training data, the training data including recordings of signals, each representing motion induced by cardiac activity, and labels indicating one or more blood parameters associated with the respective signal. Accordingly, the method may further include training the model with said training data.

[0053] In a further implementation of the method according to the third aspect, the method further comprises splitting the training data into a training subset, a validation subset, and a testing subset; training the model using the training subset; assessing a performance of the model during training using the validation subset; and evaluating the performance of the trained model using the testing subset.

[0054] In a further implementation of the method according to the third aspect, the method further comprises evaluating an error function indicating a deviation between established one or more blood parameters and reference blood parameters to generate the trained model, wherein said established one or more blood parameters and / or the reference blood parameters and / or the deviation are weighted using weighting factors during said evaluating of the error function.

[0055] A further aspect of the invention refers to one or more computer-readable media storing instructions thereon that, when executed by one or more computing devices, configure the one or more computing devices to perform a method according the third aspect or one of the implementations of the third aspect. It is to be understood that embodiments according to the first and second aspect may include logic, processor(s), or functional device(s), which may be configured according to features of an embodiment of the third aspect of the disclosure or any implementations of the third aspect, in any combination. Preferably, the computing device and the system may be configured according to embodiments of method of the third aspect, in any combination. Likewise, the method according to embodiments of the third aspect may include processing steps reflecting functionality of structural features of embodiments of the computing device or the system of the first and second aspect, in any combination.

[0056] BRIEF DESCRIPTION OF THE DRAWINGS

[0057] To illustrate the technical features of embodiments of the present invention more clearly, the accompanying drawings provided for describing the embodiments are introduced briefly in the following. The accompanying drawings in the following description are merely some embodiments of the present invention, modifications on these embodiments are possible without departing from the scope of the present invention as defined in the claims.

[0058] FIG. 1 is a schematic diagram of a device according to an embodiment of the present disclosure;

[0059] FIG. 2 is a schematic diagram of a system according to an embodiment of the present disclosure;

[0060] FIG. 3 is a flow chart of a method in accordance with a further embodiment of the present disclosure;

[0061] FIG. 4 illustrates training and validation results of a model applicable in an embodiment of the present disclosure;

[0062] FIG. 5 provides performance results of a trained model applicable in an embodiment of the present disclosure; and

[0063] FIG. 6 provides performance results of a trained model applicable in an embodiment of the present disclosure. DETAILED DESCRIPTION OF EMBODIMENTS

[0064] In the following description, reference is made to drawings which show by way of illustration various embodiments. Also, various embodiments will be described below by referring to several examples. It is to be understood that the embodiments may include changes in design and structure without departing from the scope of the claimed subject matter.

[0065] The techniques described herein may be implemented in various computing systems, examples of which Eire described in greater detail below. Such systems generally involve the use of suitably-configured computing devices implementing a number of modules, each providing one or more operations needed to complete execution of such techniques. Each module may be implemented in its own way; all need not be implemented the same way. As used herein, a module may be a structural component of a system which performs an operational role, which may be a portion of or an entire software element (e.g., a function of a process, a discrete process, or any other suitable embodiment). A module may comprise computer-executable instructions and may be encoded on a computer storage medium. Modules may be executed in parallel or serially, as appropriate, and may pass information between one another using a shared memory on the computer on which they are executing, using a message passing protocol or in any other suitable way. Exemplary modules are described below carrying out one or more tasks, though it should be appreciated that the modules and division of tasks described may be merely illustrative of the type of modules that may implement the exemplary techniques described herein, and that the invention is not limited to being implemented in any specific number, division, or type of modules. In some implementations, all functionalities may be implemented in a single module. Further, the modules are discussed below as all executing on a single computing device for clarity, though it should be appreciated that, in some implementations, the modules may be implemented on separate computing devices adapted to communicate with one another.

[0066] Fig. 1 is a schematic diagram of a device according to an embodiment of the present disclosure.

[0067] The device 100 may be configured to establish blood parameters based on at least one motion signal. The device may include a processing component 102, which may receive at least one signal 104. The signal 104 may represent motion induced by cardiac activity. Examples may include a seismocardiography signal, a phonocardiography signal, or a ballistocardiography signal, and the like, in any combination. The processing component 102 may process the at least one signal 104. This may include filtering and / or pre-processing of the at least one signal 104 in the processing component 102.

[0068] The at least one signal 104 is processed using a trained model to establish one or more blood parameters 106. The one or more blood parameters 106 may include one or more of blood glucose levels, lipid profiles, electrolyte levels, and the like in any combination.

[0069] The trained model may be implemented on the device 100 and the processing component 102 may execute the trained model. The at least one signal 104 may be pre-processed and / or filtered and provided, either directly or in the pre-processed and / or filtered form, as input to the trained model. The output of the trained model may be data, which either directly indicate the one or more blood parameters 106, or which can be further processed to establish the one or more blood parameters 106.

[0070] The trained model may also be implemented on a remote device (not shown). Accordingly, the processing component 102 may include a communication component (not shown), which may transmit the at least one signal 104 to the remote device for processing, and receive the output of the trained model, to establish the one or more blood parameters 106.

[0071] The trained model may include at least one trained neural network, preferably including one or more of an autoencoder, a convolutional neural network (CNN), a recurrent neural network (RNN), long short-term memory network (LSTM), or a ID convolutional network (ID CNN).

[0072] In one example, the trained model may be a sequential neural network model, specifically structured for processing the at least one signal 104, which may be a time series of motion data induced by the cardiac activity.

[0073] The neural network may include a plurality of layers, including one or more of Conv, Max- Pooling, GlobalAveragePooling, and Dense layers, in any combination. Conv layers and, particularly Conv ID layers, are multi- or one-dimensional convolutional layers, which may be used for extracting features from sequential data. A model may have at least one, at least two, at least three, at least four, or more Conv layers. MaxPooling layers and, particularly MaxPool- inglD layers, may follow convolutional layers. These layers may reduce dimensionality of the data, in order to condense the information and reduce computation. GlobalAveragePooling layers and, particularly GlobalAveragePooling ID layers, may be used to reduce the data to a single vector, which averages out the features across the time steps. Dense layers are fully connected neural network layers, used towards the end of the model to make predictions based on extracted features.

[0074] An example model may be defined as follows:

[0075] Layer (type) Output Shape Param # convld 12 (ConvlD) (1, 129, 32) 288 global_average_poolingld_3 (1, 64) 0

[0076] (Global ver agePool inglD) dense_9 (Dense) (1, 128) 8320 dense lO (Dense) (1, 32) 4128 dense_ll (Dense) (1, 1) 33

[0077] Total parameters: 62, 081 Trainable parameters: 62, 081 Non-trainable parameters: 0

[0078] The output shape mentioned next to each layer, such as, (1, 129, 32) for the first ConvlD layer may represent an output dimensions of that layer. The first number may be the batch size, the second number represents a length of the sequence after the layer's operation, and the third number may be the number of features or channels at that layer.

[0079] The number of parameters (Param #) indicates the number of trainable parameters in each layer. Convolutional layers typically have more parameters due to filters applied to the input data. The total parameters, for example, 62,081 in this example model, indicate the complexity and the learning capacity of the model. The example model described above may start with a ConvlD layer and alternate between convolutional and max-pooling layers. This pattern is suitable for models processing time-series or sequence data, to extract features at various scales and reduce dimensionality. After the series of convolutional and pooling layers, the model may use a global average pooling layer to condense the features into a single vector. This layer may be followed by three dense layers, progressively decreasing in size, with the final layer outputting a single value, which may correlate to the one or more blood parameters 106.

[0080] The example model enables processing of the at least one signal 104 to extract temporal features correlating with the sequential input and preferably making predictions based on them, which results in the one or more blood parameters 106.

[0081] It is to be understood that the example model is an example only and that the trained model may have a different form, layout or structure. For example, the number of Conv and MaxPool- ing layers and of Dense layers may be different. Furthermore, the number of parameters may differ. For example, another example model may include a number of Conv and MaxPooing layers, followed by at least one GlobalAveragePooling layer, followed by a number of Dense layers. The present disclosure is not limited to a particular neural network layout. Other examples may include autoencoders, RNNs, LSTM, and the like, in any combination.

[0082] The trained model may be trained using a training dataset with signals reflecting motion induced by cardiac activity that may be labeled with one or more blood parameters. Thus, the trained model may correlate data of the training dataset, which enables the trained model to calculate said one or more blood parameters 106 for the at least one signal 104.

[0083] The at least one signal 104 may be acquired using one or more sensors, which may be placed in relation to a subject. A position and / or orientation of at least one of the one or more sensors may be determined prior to processing of the at least one signal using the trained model. The at least one signal may be processed only if the position and / or orientation of the at least one sensor is determined to be within a threshold distance / orientation of a reference position / orien- tation. The position / orientation may be determined by acquiring the current position / orientation using further sensors, and / or by estimating a quality of the signal, and / or by using suitable machine learning models. Training data for such machine learning models can be generated by assigning a position / ori- entation to a cardiac activity-induced motion signal, preferably automatically or semi-automat- ically. The signal may form the input data and the position / orientation may form the output data of the training dataset.

[0084] To improve placement of the at least one sensor, a user can be informed about the current position / orientation and its suitability. This can be done, for example, by using an acoustic or graphic output device, which may provide indications that guide the user to adjust a current position / orientation of the sensor, if required. This may ensure a high quality of signals and improve correlation of the signals to blood parameters.

[0085] The signal quality and / or the arrangement of the sensors relative to a subject can also be evaluated using further physiological signals and parameters, such as a respiratory signal, and determining, depending on the physiological signals or parameters, whether the signal quality exceeds a predetermined threshold and / or whether the arrangement corresponds to a predetermined position / orientation or deviates from it. For this purpose, e.g. a predetermined assignment between the physiological signal or parameter and the signal quality and / or the arrangement can be determined. This may ensure that the sensor(s) are worn on the body of a subject in a suitable position in order to detect relevant information for estimation of the blood parameters. For example, a (raw) data signal, which represents a property of the respiration, can be transformed from a time domain into a frequency domain, for example with a fast Fourier transformation. Then, e.g. the signal energy E can be calculated using, for example, E = wherein E represents the signal energy of the respiration domain and A the amplitude of the respective frequencies in the respiration domain (e.g. 0.1 Hz to 0.6 Hz). Here, the lower respiration frequency limit is referred to as uR, the upper respiration frequency limit is referred to as oR.

[0086] A lack of respiration can be detected if the signal energy is less than a predetermined threshold. If the signal energy is greater than or equal to the predetermined threshold, a presence of the respiration can be detected. In the case of a lack of respiration, it can be assumed that the system is not worn on the body or arranged in a suitable position / orientation. In this case, no processing of the at least one signal is performed.

[0087] The signal characteristics I parameters for detecting if the system is worn correctly, for example on the body, if the system is placed correctly and the current signal quality at the particular position can furthermore be trained into a neural network. The resulting model can then be used to determine the current placement and signal quality index.

[0088] Fig. 2 is a schematic diagram of a system according to an embodiment of the present disclosure.

[0089] The system 200 may include at least one server 202, which may be reachable via a network 204. One or more client devices 206 may connect to the server 202. At least one of the one or more client devices 206 may correspond to the device 100 of Fig. 1.

[0090] The server 202 may implement a trained model, which may receive, as input, signals reflecting motion induced by cardiac activity, and which may process the signals to derive blood parameters that correlate with the input signals.

[0091] For example, the client device 206 may receive a signal 208 reflecting motion induced by cardiac activity. The client device 206 may pre-process or filter the signal 208. The client device 206 may establish a communication link 210 to the network 204 in order to connect to the server 202. The client device 206 may use the communication link 210 to transmit at least parts of the signal 208 or of the pre-processed and / or filtered signal to the server 202 for processing. The server 202 may receive the at least a part of the (pre-processed or filtered) signal 208 and input the received signal into the trained model. The output of the trained model may be submitted via the network 204 back to the client device 206. Either the server 202 or the client device 206 may establish one or more blood parameters 212, which correlate with the input signal. The client device 206 may provide the one or more blood parameters 212, for example, on a display.

[0092] Embodiments of the present disclosure can be used for monitoring humans, including infants and adults. Further use cases include monitoring of blood parameters of animals, such as dogs, cows, horses, camels, and the like. Sensors can be placed in suitable gear, such as belts, collars, chest belts, holsters, and the like. Irrespective of the monitored subject, monitoring can be performed remotely, i.e. without any contact, or by placing sensors directly on the subject. Remote monitoring may include placing sensors on incubators, in a mattress, on beds, on seats (such as in vehicles) or in rooms, or remotely monitoring animals. Monitoring with contact can include placing sensors or a mobile device, such as a smartphone or a smart watch, on the body of a human or animal, or by integrating sensors within medical devices, such as a pacemaker. Fig. 3 is a flow chart of a method in accordance with a further embodiment of the present disclosure.

[0093] The method 300 may start in item 302. The method 300 may be a computer-implemented method, which may be executed by one or more computing devices, such as the device 100 of Fig. 1, the client device 206 or the server 200 of Fig. 2.

[0094] The method 300 may process at least one signal 304 representing motion induced by cardiac activity, in item 306. The at least one signal 304 may correspond to the signal 104 in Fig. 1 or signal 208 in Fig. 2. The at least one signal 304 may represent a time series of data points representing the motion, including, for example, data points of seismocardiography, phonocardiography, or ballistocardiography signals, in any combination. The at least one signal 304 may include a single or multiple channels, each including a time series of data points representing the motion.

[0095] Preferably, a signal quality of the at least one signal 304 may be determined. If the signal quality is below a threshold, the signal 304 may be discarded and the method 300 may proceed with a further signal. The signal quality can be, for example, a signal-to-noise ratio or a variable representing this ratio. If the signal-to noise ratio is greater than a predetermined amount, the signal is processed using the trained model. Further techniques for determining signal quality may include template comparison, where a reference signal is compared to the current signal. The signal quality can also be determined using neural networks. Training data for the neural networks can be generated by assigning a quality measure representing the signal quality to a cardiac activity-induced motion signal.

[0096] Said processing 306 may include an optional pre-processing or filtering of the at least one signal 304, in item 308. In item 310, the at least one signal 304 or the pre-processed and filtered at least one signal may be provided as input to a trained model 312.

[0097] The trained model 312 may be implemented on the same computing device executing item 306 or on another computing device, such as a server or a remote processing component in a cloud environment, which may be connected via a network or any other data connection suitable for forwarding the at least one signal 304 or the pre-processed and filtered at least one signal. The trained model 312 may process, transform, or correlate the provided input signal(s) and generate an output, which can be used to establish one or more blood parameters 314. The output may be the one or more blood parameters 314 or an intermediate value, which can be transformed into the one or more blood parameters 314.

[0098] The one or more blood parameters 314 may represent individual values of the blood parameters 314. The one or more blood parameters 314 may also represent a time series of blood parameters. In this case, the input signal of the trained model 312 may be transformed into an output signal representing the time series of blood parameters. As an example, the one or more blood parameters may include one or more of blood glucose levels, lipid profiles, and electrolyte levels.

[0099] The method 300 may proceed with rendering the one or more blood parameters 314 on a user interface, in item 316. The one or more blood parameters 314 may be provided on a user interface of the device performing step 306. However, item 316 may be executed on yet another computing device.

[0100] Processing of the method 300 may continue with any new signal 304, sequentially or in parallel. The method 300 may process a further part of the signal 304 or any new signal 304 to establish corresponding new blood parameters 314.

[0101] The method may end in item 318.

[0102] Fig. 4 illustrates training and validation results of a model applicable in one embodiment of the present disclosure. Particularly, the training results may correspond to the example model as disclosed with regard to Fig. 1. However, it is to be understood that other models according to embodiments of the present disclosure may be trained in a similar manner.

[0103] As am example, to train a model for detection of blood glucose as a blood parameter, the training data to build a trained model may be recorded by a continuous glucose monitoring device (CGM; for training data / ground truth). In parallel to CGM data, an acceleration sensor of a portable device, such as a smartphone, may be used to record seismocardiograms, including an acceleration signal, by laying the portable device on a participants chest. During the recording process the participant may bne laying down, e.g., on a couch. In a first processing step, three axes of the acceleration signal may be computed into a vector norm as follows:

[0104] This may be conducted for each acceleration triplet within the input arrays, for example, at a sampling rate of between 50 Hz and 150 Hz, between 80 Hz and 120 Hz, and preferably at 100 Hz. Any other suitable sampling rate can be used. Subsequently, the normed signal may be filtered with a Butterworth bandpass with a passband between 8Hz to 30Hz at a filter order of 4, for example. Subsequently, the normed and filtered signal may be used to perform a Welch transformation with a window size of 256 samples, or any other suitable sample size, such as 128 or 512 samples. The values within the resulting Welch spectra may then be normalized as follows:

[0105] The resulting normalized Welch spectra may be used as input for the model, such as the example model as disclosed in relation to Fig. 1. Accordingly, the model architecture may include four ID-convolutional layers each followed by a max -pooling layer, a global average pooling layer, and three consecutive dense layers. A 129-input shape (non-mirrored normalized Welch spectrum) may be applied as an input and mapped to a single value output (regression).

[0106] Fig. 4 shows the training and validation performance of the model as training loss and validation over training epochs.

[0107] The training process indicates a quick convergence. As indicated, after 18 epochs the model may start to slightly overfit. Thus, early stopping may be performed to ensure a ideally fitted model. The applied loss function is mean squared error. The results for each epoch are shown below:

[0108] In this example, the computed R2value for the test user may be given as R2= 0.7622282269874261. The R value > 0.7 shows a strong correlation between the estimated glucose and the ground truth recording via CGM.

[0109] In another example, the model may be trained to determine potassium data as blood parameters. In the first step the potassium level of each subject within the recording set may be determined by a blood gas analysis. In parallel to taking a blood sample, a seismocardiogram may be recorded with an acceleration sensor on the subject’s precordium. The result of the potassium level provided by the blood gas analysis system may then be added as a ground truth to the recorded three-dimensional acceleration signal of the seismocardiogram.

[0110] The preprocessing steps may be analogous to the example directed at determination of blood glucose levels described above with regard to Fig. 4. The layout of the model may be analogous to the example model as describe above with regard to Fig. 1.

[0111] The computed R2value for a test user may be given as R2= 0.853181110274271.

[0112] It is to be understood that other examples may train a model to other blood parameters or a combination of blood parameters, such as the aforementioned blood glucose levels and the potassium parameters. Any other blood parameters are encompassed by the present disclosure. Blood parameters may include any kind of parameter that may be generated in blood analysis, a fundamental component of diagnostic medicine. Blood analysis may encompass various segments that provide insights into different aspects of physiological health.

[0113] The models, such as those discussed with regard to Figures 1 to 3 may be trained on vast datasets that correlate the recorded at least one signal representing motion induced by cardiac activity with traditional blood markers, enabling accurate estimation of blood parameters. Examples include blood glucose levels and potassium levels, as discussed above. Other examples may include other blood parameters.

[0114] Blood markers and / or blood parameters may include Complete Blood Count (eg Red Blood Cells, Hemoglobin, White Blood Cells, Platelets), Basic Metabolic Panel (eg Glucose, Electrolytes ((eg Sodium, Potassium, Chloride)), Kidney Function (eg BUN, Creatinine), Lipid Panel (eg Total Cholesterol, LDL, HDL), Liver Function Tests (eg ALT, AST, ALP, Bilirubin), Inflammatory Markers (eg CRP, ESR), Thyroid Function Tests (eg TSH, FT4, FT3), Hormone Levels (eg Testosterone, Estrogen, Progesterone, Cortisol), Tumor Markers (eg AFP, CEA, PSA), Immunoglobulin Panel (eg IgA, IgG, IgM), Vitamin and Mineral Levels (eg Vitamin D, Vitamin Bl 2, Iron) and others.

[0115] Fig. 5 provides performance results of a trained model applicable in one embodiment of the present disclosure. For example, Fig 5 may show performance results of the trained model as discussed with regard to Fig. 4 above. Fig. 5 shows a Parks error grid for a test user. The Parkes error grid depicts Freestyle Libre 3 references (mg / dL) against glucose results (mg / dL).

[0116] The Parkes error grid of Fig. 5 shows that all values from the test user (unseen data to the model) are located within the zones A (clinically accurate) and B (clinically acceptable). This indicates a very good usability of the results.

[0117] Fig. 6 provides performance results of a trained model applicable in one embodiment of the present disclosure. For example, Fig 5 may show performance results of the trained model as discussed with regard to Fig. 4 and Fig. 5 above. Fig. 5 shows a Clarke error grid for a test user. The Clarke error grid depicts reference glucose concentration values (mmol / L) over predicted glucose concentration values (mmol / L). The Clarke error grid of Fig. 6 shows all values of the test user (unseen data to the Al model) within the zones A and B (clinically accurate and clinically acceptable, respectively). This result indicate a high accuracy / certainty of the model.

[0118] Embodiments of the present disclosure reflect am innovative solution that combines advanced technologies such as seismocardiograms or other signal data reflecting motion induced by cardiac activities, and machine learning Al to revolutionize blood testing. Traditional blood sampling methods involve inherent risks of infections and complications, as well as limiting continuous measurements that increase the risk of iatrogenic anemia.

[0119] Embodiments of the present disclosure overcome these disadvantages by providing a non-in- vasive and continuous approach to blood testing for a comprehensive range of blood parameters.

[0120] The field of medical diagnostics has made remarkable advances in recent years, aimed at improving patient experience and optimizing testing methods. While traditional blood sampling is effective, it has significant disadvantages, including the risk of infections and complications, as well as limiting continuous monitoring. Embodiments of the present disclosure address these limitations by utilizing machine learning and Al technologies to provide safe, convenient and continuous blood testing for a variety of blood parameters.

[0121] Embodiments of the present disclosure use signals that represent motion induced by cardiac activities. Such signals can be obtained, for example, by seismocardiogram technology. For example, by simply placing a smartphone or sensors on a user's chest, the smartphone or a connected device may detect fine vibrations generated by cardiac activity. These vibrations contain valuable information about blood flow, heart rate and cardiovascular health. Through advanced signal processing and Al algorithms, embodiments of the present disclosure may interpret these vibrations to determine various blood parameters without requiring invasive blood sampling.

[0122] Embodiments of the present disclosure integrate artificial intelligence and machine learning techniques to interpret the motion data and signals. Machine learning based on extensive data sets correlates the recorded vibrations with traditional blood markers and enables accurate estimation of blood glucose levels, lipid profiles, blood picture, coagulation, electrolyte levels, hormones, tumor markers and more, in any combination. This Al-controlled approach not only ensures accuracy, but also adapts to individual variations, which is a personalized and reliable solution for blood testing.

[0123] A prominent example of the capabilities of embodiments of the present disclosure is non-inva- sive and continuous blood glucose measurement. Other examples include lipid profiles, and electrolyte levels, to name a few.

[0124] In general, taking blood without informed consent of the person concerned is considered to be a violation of their rights worldwide. Due to the significant risks, talking blood without consent could be a bodily injury and may only be carried out (except in emergencies or due to a court order) after appropriate declaration and consent of the patient. Informed consent usually means that the person concerned must be adequately informed of the purpose, course, potential risks and benefits of blood sampling before giving consent. In most international legal systems, taking blood is considered a medical intervention to be carried out by qualified medical professionals such as physicians, nurses or certified medical -technical personnel. These professionals are trained to safely carry out the intervention and thereby minimize the risk of complications.

[0125] Some patients require frequent or continuous blood testing. For example, people with diabetes have to carry out finger prick tests several times a day. Using embodiments of the present disclosure, users can effortlessly monitor their blood glucose levels by placing their smartphone or sensors non-invasively on their chest. While modem continuous glucose monitors (CGM) can only measure blood glucose levels at the same time, embodiments of the present disclosure can capture continuous real-time data around the clock.

[0126] Seismocardiogram technology captures fluctuations in heart activity that are indicative of blood parameter levels, such as glucose levels, and processing the signals using the trained model may provide real-time values of the blood parameters, such as glucose values. This not only improves patient comfort, but also provides a more comprehensive understanding of blood parameter levels throughout the day, which is valuable in health management environments without the risk of infections or complications. Embodiments of the present disclosure provide various benefits. Unlike traditional blood sampling involving needles and punctures, the non-invasive approach of embodiments of the present disclosure eliminates the risk of infections and related complications, which provides safety to patients and medical professionals alike. One of the greatest benefits of embodiments of the present disclosure is the capability for continuous blood testing. Users cam gain real-time insights into various blood parameters, enabling timely interventions and treatment plan adjustments. The non-invasive nature of motion signals, such as seismocardiograms, improves user comfort, especially for those who fear needles or experience fear when taking blood samples.

[0127] Furthermore, placing a smartphone or sensor on the chest is am uncomplicated process that makes the solution easily accessible for self-monitoring at home or in clinical settings. With over 6.5 billion smartphones worldwide, embodiments of the present disclosure provide access to health diagnostics almost anywhere and at any time. Embodiments of the present disclosure provide a wide range of blood tests and parameters and thus provides a holistic diagnostic approach.

[0128] In a preferred embodiment, the disclosure may encompass one or more processing devices with a plurality of means that may be configured to perform functionality of embodiments of the present invention. In particular, at least one device may be configured for establishing blood parameters, and may comprise means for processing at least one signal representing motion induced by cardiac activity, wherein the at least one signal is processed using a trained model to establish one or more blood parameters. The device or another one of the one or more processing devices may further comprise means for filtering the at least one signal. The device or another one of the one or more processing devices may further comprise means for providing the at least one signal or the filtered at least one signal as input to the trained model, wherein the one or more blood parameters are establish based on an output of the trained model. The device or another one of the one or more processing devices may further comprise means for rendering a representation of the one or more blood parameters on a user interface. The device or another one of the one or more processing devices may further comprise means for retrieving an identification of the subject and using the identification to adapt said processing to variations in the subject. The device or another one of the one or more processing devices may further comprise means for detecting the at least one signal using one or more sensors, wherein at least one of the one or more sensors is placed on a thorax of a subject or wherein at least one of the one or more sensors is configured to detect motion without contact to a subject. The device or another one of the one or more processing devices may further comprise means for transmitting the at least one signal to a computing device, wherein the at least one signal is processed by the computing device. The device or another one of the one or more processing devices may further comprise means for training a model with training data, the training data including recordings of signals, each representing motion induced by cardiac activity, and labels indicating one or more blood parameters associated with the respective signal. The device or another one of the one or more processing devices may further comprise means for splitting the training data into a training subset, a validation subset, and a testing subset; training the model using the training subset; assessing a performance of the model during training using the validation subset; and evaluating the performance of the trained model using the testing subset. The device or another one of the one or more processing devices may further comprise means for evaluating an error function indicating a deviation between established one or more blood parameters and reference blood parameters to generate the trained model, wherein said established one or more blood parameters and / or the reference blood parameters and / or the deviation are weighted using weighting factors during said evaluating of the error function.

[0129] Preferred embodiments of the one or more processing devices may include further means for implementing details of aspects, implementations thereof, and further embodiments as disclosed with regard to Fig. 1 to 3 and corresponding description, in any combination.

[0130] It is to be understood that the implementational details as provided in Fig. 1 to 3 represent preferred examples. Other implementations using different components, modules, blocks, units, circuitry, connections, and links can be used, and the present disclosure is not restricted by a particular implementation in silicon.

[0131] While some embodiments have been described in detail, it is to be understood that aspects of the disclosure can take many forms. In particular, the claimed subject matter may be practiced or implemented differently from the examples described, and the described features and characteristics may be practiced or implemented in any combination. The embodiments shown herein are intended to illustrate rather than to limit the invention as defined by the claims.

Claims

AMENDED CLAIMS received by the International Bureau on 11 April 2025 (11.04.2025)1. A computing device for establishing blood parameters, comprising: a processing component configured to process at least one seismocardiography signal representing motion induced by cardiac activity, wherein the at least one seismocardiography signal is processed using a trained model to establish one or more blood parameters.

2. The computing device of claim 1, wherein the processing component is further configured to filter the at least one seismocardiography signal.

3. The computing device of claim 1 or 2, wherein the processing component is further configured to provide the at least one seismocardiography signal or the filtered at least one seismocardiography signal as input to the trained model, wherein the one or more blood parameters are establish based on an output of the trained model.

4. The computing device according to any one of the preceding claims, wherein the output of the trained model represents a time series of blood parameters.

5. The computing device according to any one of the preceding claims, wherein the one or more blood parameters include one or more of blood glucose levels, lipid profiles, and electrolyte levels.

6. The computing device according to any one of the preceding claims, wherein the at least one seismocardiography signal includes at least one channel, each including a time series of data points representing the motion.

7. The computing device according to any one of the preceding claims, further comprising a user interface configured to render a representation of the one or more blood parameters.

8. The computing device according to any one of the preceding claims, wherein the at least one signal is detected using one or more sensors, wherein at least one of the one or more sensors is placed on a thorax of a subject or wherein at least one of the one ormore sensors is configured to detect motion induced by cardiac activity without contact to a subject.

9. The computing device according to any one of the preceding claims, further comprising a communication component configured to transmit the at least one seismocardiography signal to a remote processing component, wherein the at least one seismocardiography signal is processed by the remote processing component.

10. The computing device according to any one of the preceding claims, further comprising an interface configured to receive the at least one signal.

11. The computing device according to any one of the preceding claims, wherein the trained model includes at least one trained neural network, including one or more of an autoencoder, a convolutional neural network (CNN), a recurrent neural network (RNN), long short-term memory network (LSTM), or a ID convolutional network (ID CNN).

12. The computing device according to any one of the preceding claims, wherein the trained model is generated by training a model with training data, the training data including recordings of signals, each representing motion induced by cardiac activity, and labels indicating one or more blood parameters associated with the respective signal, wherein preferably the training data is split into a training subset, a validation subset, and a testing subset, wherein the model is trained using the training subset, a performance of the model during training is assessed using the validation subset, and the performance of the trained model is evaluated using the testing subset.

13. A system, comprising at least one server; a network; and at least one client device connectable to the at least one server via the network, the at least one client device according to any one of the preceding claims.

14. A computer-implemented method for establishing blood parameters, comprising: processing at least one signal representing motion induced by cardiac activity,wherein the at least one signal is processed using a trained model to establish one or more blood parameters.

Images