A smart watch health data monitoring method based on machine learning

Abstract

The application discloses a kind of smart watch health data monitoring methods based on machine learning, including the following steps: collecting health data in smart watch end, generates health data time sequence;Collecting behavior events, generate behavior event sequence;Select health data in pre-window event analysis time, generate event response trajectory set;Extract response start time offset, response duration and response change rate, form historical response hysteresis characteristic sequence;Improved Jamba is constructed, introduce time state evolution constraint mechanism, generate historical hysteresis structure representation;Real-time response hysteresis characteristic sequence is input into improved Jamba, and real-time hysteresis structure representation result is generated;Determine based on real-time hysteresis structure representation result, output health monitoring result.The application adopts improved Jamba two-way time modeling method, realizes health data structural hysteresis determination, with the advantages of accurate determination, response stability and strong adaptability.

Description

Technical Field

[0001] This invention relates to the field of health data monitoring in smartwatches, and more particularly to a method for monitoring health data in smartwatches based on machine learning. Background Technology

[0002] With the popularization of smart wearable devices, more and more health monitoring methods have been proposed to track users' health data in real time. These methods usually rely on a variety of physiological signals collected from the device, such as heart rate, blood oxygen, and sleep status, to achieve health monitoring and early warning through data analysis and modeling. Existing health data monitoring methods mostly focus on single time series prediction or are based solely on the evolution of unidirectional time states, lacking comprehensive analysis of health data. In particular, they fail to effectively model the hysteresis effect in complex health events. Traditional methods cannot fully capture the reverse effects of health status and its time dependence, making it impossible to accurately determine whether the health status can recover to the original stable state.

[0003] This invention proposes a health data monitoring method based on a two-way time propagation mechanism, which solves the problems of inaccurate health status determination and inability to accurately model lag phenomena in traditional technologies. By introducing a two-way time propagation mechanism and regressibility determination, this invention can more accurately determine whether a health status can return to a stable state, thereby improving the accuracy and reliability of health monitoring. This method integrates bidirectional analysis of historical and real-time data, not only considering forward changes in health status, but also backward assessing its regressibility, greatly enhancing the model's ability to capture health lag effects. Summary of the Invention

[0004] One objective of this invention is to propose a health data monitoring method for smartwatches based on machine learning. This invention employs an improved Jamba bidirectional time modeling method to achieve structural hysteresis determination of health data, which has the advantages of accurate determination, stable response, and strong adaptability.

[0005] A method for monitoring health data of a smartwatch based on machine learning according to an embodiment of the present invention includes the following steps: Health data is collected on the smartwatch and preprocessed to generate a time-series health data sequence. Collect behavioral events generated during the use of smartwatches and organize them to generate behavioral event sequences; A unified timeline is constructed based on health data time series and behavioral event series. The behavioral event timestamp is used as the time anchor point. Within the preset event analysis time window, the corresponding health data is selected to generate a set of event response trajectories corresponding to each behavioral event. The event response trajectory in the event response trajectory set is identified by the response start time and determined by the response end time. The response start time offset, response duration and response change rate are extracted to generate response hysteresis feature vectors and form a historical response hysteresis feature sequence in chronological order. Using historical response hysteresis feature sequences as input, an improved Jamba is constructed. A time state evolution constraint mechanism based on behavioral events is introduced to model the historical response hysteresis feature sequences under constrained time evolution rules, generating a historical hysteresis structure representation. For real-time collected health data and behavioral events, a real-time response hysteresis feature sequence is generated. Under the backtracking constraint formed by the historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba, and the constrained time state evolution and backtracking decision processing are performed to generate the real-time hysteresis structure representation result. Based on the real-time hysteresis structure characterization results, a health monitoring result is outputting whether there is structural hysteresis in the health data response process.

[0006] Optionally, the health data specifically includes heart rate data, heart rate variability data, blood oxygen saturation data, body surface temperature data, posture change angular velocity data, and sleep duration data. The preprocessing specifically includes outlier removal, missing data completion, unified time index alignment, and amplitude normalization.

[0007] Optionally, the generation of the behavioral event sequence specifically includes: During the use of the smartwatch, event records corresponding to the user's behavior are collected, including event records triggered by user operations, event records triggered by sensors, and event records generated during operation. Each event record is associated with corresponding timestamp information. The collected event records are filtered for validity, and event records with missing timestamp information, incomplete event type identification, or abnormal occurrence time are removed to form a set of valid event records; Based on the preset behavioral event classification rules, the event records in the valid event record set are merged by event type, and event records with the same event type identifier and occurring consecutively within a preset time interval are merged into the same behavioral event; The time boundaries of the merged behavioral events are determined, the start time and end time of each behavioral event are determined, and the merged behavioral events are mapped to behavioral event items with unique event identifiers. All behavioral event items are sorted according to their start time, forming a sequence of behavioral events arranged in chronological order.

[0008] Optionally, the generation of the event response trajectory set specifically includes: A unified timeline is established based on health data time series and behavioral event series, and the health data time series and behavioral event series are time-aligned. In the unified timeline, the timestamp of each behavioral event in the sequence of behavioral events is used as the time anchor point to determine the anchor position of each behavioral event on the unified timeline. For each behavioral event, the time range is determined at the corresponding time anchor point according to the preset event analysis time window. The preset event analysis time window consists of the analysis time range before the event occurs and the analysis time range after the event occurs. Based on a unified timeline and a preset event analysis time window, health data segments corresponding to the time anchor points of each behavioral event are selected from the health data time series. Health data fragments are organized according to health data type to generate event response trajectories corresponding to single behavioral events; For each behavioral event in the behavioral event sequence, the time anchor point determination, time window limitation, and health data selection and processing are repeatedly executed to generate event response trajectories corresponding to each behavioral event. The event response trajectories corresponding to each behavioral event are collected in the order of the behavioral event sequence to generate a set of event response trajectories corresponding to each behavioral event.

[0009] Optionally, the generation of the historical response hysteresis feature sequence specifically includes: Based on the event response trajectory set, the event response trajectory corresponding to each behavior event is selected sequentially according to the order of the behavior events in the behavior event sequence; For each behavioral event, the baseline time range before the event occurs is determined within a preset event analysis time window, and the corresponding baseline state is determined based on the event response trajectory within this baseline time range. Using the time anchor point corresponding to the behavioral event as a reference, the event response trajectory is detected moment by moment within the time range after the event occurs, and the moment when the event response trajectory first deviates continuously from the baseline state is identified to determine the response start time corresponding to the behavioral event. In the event response trajectory, based on the response start time, time-by-time detection is continued to identify the moment when the event response trajectory recovers to the baseline state, and the response end time corresponding to the behavior event is determined. Based on the response start time, response end time, and the time anchor point corresponding to the behavioral event, the response start time offset and response duration are extracted. Based on the change trend of the event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration, and response change rate corresponding to the same behavioral event are combined to generate the response hysteresis feature vector corresponding to that behavioral event. The response hysteresis feature vectors corresponding to each behavioral event in the behavioral event sequence are arranged and aggregated according to the time order of the behavioral event sequence to form a historical response hysteresis feature sequence.

[0010] Optionally, the generation of the historical hysteresis structure representation specifically includes: In the improved Jamba, a unified time state space is constructed, and the historical response hysteresis feature sequence is written into a unified time axis. A corresponding time index is assigned to each time position after writing, and a forward time state and a reverse time state are established at each time index to form a bidirectional time state pair. Within the improved Jamba, based on the event analysis time window corresponding to the behavior event, when the time index is within the event analysis time window of the behavior event, the bidirectional time state pair is frozen, blocking the state update of the forward time state and the reverse time state. When the time index exceeds the event analysis time window, the freezing constraint of the bidirectional time state pair is released. In the improved Jamba, based on the time state corresponding to the historical response hysteresis feature sequence that has been written into the unified time state space, forward time propagation is performed on the forward time state in the direction of increasing time index, and reverse time propagation is performed on the reverse time state in the direction of decreasing time index, forming a bidirectional time propagation state sequence that evolves in parallel within the same time state space. During bidirectional time propagation, at each time index, the corresponding forward and reverse time states are analyzed for state consistency in the improved Jamba. Based on the state consistency relationship between the forward and reverse time states, a non-reversibility determination is performed. When the forward and reverse time states meet the consistency condition, the time state at that time index is determined to be a reversible state. When the forward and reverse time states do not meet the consistency condition, the time state at that time index is determined to be a non-reversible state. Write the reversible and non-reversible states generated at each time index into the time state space as time state attributes, and use the non-reversible state as a blocking flag for the evolution of time state. The set of non-reversible time indices corresponding to each behavioral event within the event analysis time window is collected, and then organized in conjunction with the time anchor points corresponding to each behavioral event. A historical lag structure representation is generated according to the time sequence of the behavioral events.

[0011] Optionally, the generation of the real-time hysteresis structure characterization results specifically includes: Based on the real-time collected health data and behavioral events, organize and generate real-time health data time series and real-time behavioral event sequences; Based on real-time health data time series and real-time behavioral event series, a unified time axis is established to generate real-time unified time axis alignment results; Using the real-time unified time axis alignment results, the timestamps of each behavioral event in the real-time behavioral event sequence are used as time anchors. Within the event analysis time window, corresponding health data segments are continuously selected from the real-time health data time sequence. The health data segments are organized according to the health data type to generate real-time event response trajectories corresponding to each behavioral event, and then aggregated to form a set of real-time event response trajectories. For each real-time event response trajectory in the real-time event response trajectory set, a baseline time range before the event occurs is determined based on the time anchor point. Within the baseline time range, the states of the real-time event response trajectories are aggregated to generate the baseline state of the corresponding behavioral event. Using the time anchor point as a reference, the real-time event response trajectory is detected hour by hour within the time range after the event occurs. After the response start time, the moment when the real-time event response trajectory recovers to the baseline state is detected. The moment when the real-time event response trajectory first continuously deviates from the baseline state is identified to determine the response start time, and the moment when the response end time is determined. Based on the response start time, response end time and time anchor point, the response start time offset and response duration are extracted. Based on the change trend of the real-time event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration and response change rate are combined to generate a real-time response hysteresis feature vector. The real-time response hysteresis feature vectors corresponding to each behavioral event in the real-time behavioral event sequence are arranged and aggregated in chronological order to form a real-time response hysteresis feature sequence. Under the constraint of historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba. Using a two-way time propagation mechanism, the time state evolution of the real-time response hysteresis feature sequence is carried out at different time scales. Based on the backtrackable constraint conditions formed by the historical hysteresis structure representation, the backtracking decision processing of the time state is performed to generate the real-time hysteresis structure representation result.

[0012] Optionally, the generation of the health monitoring results specifically includes: Based on the real-time hysteresis structure representation results, a set of historical hysteresis structure representation results corresponding to the real-time hysteresis structure representation results is obtained. The set of historical hysteresis structure representation results consists of hysteresis structure representation results formed during the historical health data response process. The real-time hysteresis structure representation results are compared with the set of historical hysteresis structure representation results in the structure representation space to generate a real-time hysteresis structure deviation representation. Based on the real-time hysteresis structure deviation characterization, the structural change state of the real-time health data response process relative to the historical health data response process is determined. When the structural change state meets the criteria for structural hysteresis, a health monitoring result is generated that characterizes the existence of structural hysteresis in the health data response process. When the structural change state does not meet the criteria for structural hysteresis, a health monitoring result is generated that represents the health data response process without structural hysteresis.

[0013] The beneficial effects of this invention are: This invention models health data collected by smartwatches and user behavior events along a unified timeline, introducing an improved Jamba bidirectional time state evolution mechanism to structurally characterize the time lag phenomenon in the health data response process. This effectively overcomes the limitations of existing technologies that rely solely on unidirectional time prediction or simple threshold judgment. By constructing event response trajectories driven by behavioral events and further extracting lag features such as response start time offset, response duration, and response change rate, this invention can accurately reflect the true response process of health data to behavioral events in the time dimension, providing a reliable data foundation for subsequent lag structure analysis.

[0014] This invention models the historical response hysteresis feature sequence through bidirectional time propagation to form a historical hysteresis structure representation. This representation is then used as a fallback constraint in the real-time analysis process to guide the temporal evolution of real-time health data. This approach, which directly involves the historical structure in real-time judgment, enables health monitoring to move beyond instantaneous changes or short-term trends. Instead, it allows for the identification of whether irreversible structural changes occur in health data under the influence of behavioral events, thereby significantly improving the accuracy and stability of the judgment of abnormal health states.

[0015] This invention achieves the determination of the reversibility of time states by analyzing the consistency between forward and reverse time states, thus providing clear structural criteria for health monitoring results. This avoids the misjudgment problems caused by individual differences or short-term fluctuations in existing methods. Overall, this invention achieves refined modeling and structural determination of health data response lag while ensuring engineering feasibility. It not only improves the reliability and interpretability of health monitoring results, but also provides more robust technical support for the application of wearable devices in long-term health assessment and risk warning. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a smartwatch health data monitoring method based on machine learning proposed in this invention; Figure 2This is a schematic diagram of the bidirectional time state propagation and time state freezing and unfreezing mechanism in Jamba, which is an improved method for monitoring health data of smartwatches based on machine learning proposed in this invention. Figure 3 This is a schematic diagram illustrating the generation of historical hysteresis structure representations for a machine learning-based smartwatch health data monitoring method proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A machine learning-based method for monitoring health data on smartwatches includes the following steps: Health data is collected on the smartwatch and preprocessed to generate a time-series health data sequence. Collect behavioral events generated during the use of smartwatches and organize them to generate behavioral event sequences; A unified timeline is constructed based on health data time series and behavioral event series. The behavioral event timestamp is used as the time anchor point. Within the preset event analysis time window, the corresponding health data is selected to generate a set of event response trajectories corresponding to each behavioral event. The event response trajectory in the event response trajectory set is identified by the response start time and determined by the response end time. The response start time offset, response duration and response change rate are extracted to generate response hysteresis feature vectors and form a historical response hysteresis feature sequence in chronological order. Using historical response hysteresis feature sequences as input, an improved Jamba is constructed. A time state evolution constraint mechanism based on behavioral events is introduced to model the historical response hysteresis feature sequences under constrained time evolution rules, generating a historical hysteresis structure representation. For real-time collected health data and behavioral events, a real-time response hysteresis feature sequence is generated. Under the backtracking constraint formed by the historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba, and the constrained time state evolution and backtracking decision processing are performed to generate the real-time hysteresis structure representation result. Based on the real-time hysteresis structure characterization results, a health monitoring result is outputting whether there is structural hysteresis in the health data response process.

[0019] In this embodiment, the health data specifically includes heart rate data, heart rate variability data, blood oxygen saturation data, body surface temperature data, posture change angular velocity data, and sleep duration data. The preprocessing specifically includes outlier removal, missing data completion, unified time index alignment, and amplitude normalization.

[0020] In this embodiment, the generation of the behavioral event sequence specifically includes: During the use of the smartwatch, event records corresponding to the user's behavior are collected, including event records triggered by user operations, event records triggered by sensors, and event records generated during operation. Each event record is associated with corresponding timestamp information. The collected event records are filtered for validity, and event records with missing timestamp information, incomplete event type identification, or abnormal occurrence time are removed to form a set of valid event records; Based on the preset behavioral event classification rules, the event records in the valid event record set are merged by event type, and event records with the same event type identifier and occurring consecutively within a preset time interval are merged into the same behavioral event; The time boundaries of the merged behavioral events are determined, the start time and end time of each behavioral event are determined, and the merged behavioral events are mapped to behavioral event items with unique event identifiers. All behavioral event items are sorted according to their start time, forming a sequence of behavioral events arranged in chronological order.

[0021] In this embodiment, the generation of the event response trajectory set specifically includes: A unified timeline is established based on health data time series and behavioral event series, and the health data time series and behavioral event series are time-aligned. In the unified timeline, the timestamp of each behavioral event in the sequence of behavioral events is used as the time anchor point to determine the anchor position of each behavioral event on the unified timeline. The determination of the anchoring position specifically includes: Based on the timestamp of the corresponding behavioral event, the corresponding time index position is located in the unified timeline; when the behavioral event has a start time and an end time, the timestamp corresponding to the start time of the behavioral event is used as the time anchor point of the behavioral event; when multiple behavioral events overlap in time in the unified timeline, the corresponding time anchor point positions are determined according to the order in which the behavioral events occur; when the timestamp of the behavioral event is not completely aligned with the sampling time point of the health data time series, time index mapping is performed based on the unified timeline to determine the health data sampling time point closest to the timestamp of the behavioral event as the anchor position. For each behavioral event, the time range is determined at the corresponding time anchor point according to the preset event analysis time window. The preset event analysis time window consists of the analysis time range before the event occurs and the analysis time range after the event occurs. The determination of the time range specifically includes: Centered on the time anchor point corresponding to the behavioral event, the analysis time range is determined separately in the directions before and after the event on a unified time axis. For different behavioral events in the sequence, the length of the analysis time range before and after the event is determined according to the event type. When a behavioral event has a start time and an end time, the time interval between the start and end times is included in the analysis time range after the event. When a behavioral event is an instantaneously triggered event, the analysis time range after the event is determined starting from the time anchor point corresponding to the behavioral event. On the unified time axis, the analysis time range before and after the event is concatenated to form the event analysis time range corresponding to the behavioral event. Based on a unified timeline and a preset event analysis time window, health data segments corresponding to the time anchor points of each behavioral event are selected from the health data time series. The selection of the health data segments specifically includes: Under a unified timeline, the time anchor point corresponding to each behavioral event and its corresponding event analysis time range are read. Based on the event analysis time range, a time index interval corresponding to that time range is determined in the health data time series. According to the time index interval, continuous health data records within the event analysis time range are extracted from the health data time series to form corresponding health data segments. When the health data time series contains multiple health data types, data of each health data type within the time index interval are selected separately to form health data segments that are synchronously aligned across multiple health data types. When the sampling time point of the health data time series does not completely coincide with the boundary of the event analysis time range, the sampling time point closest to the boundary of the event analysis time range under the unified timeline is used as the boundary of the health data segment. The health data segment selected for the same behavioral event is taken as the health data segment corresponding to that behavioral event. Health data fragments are organized according to health data type to generate event response trajectories corresponding to single behavioral events; For each behavioral event in the behavioral event sequence, the time anchor point determination, time window limitation, and health data selection and processing are repeatedly executed to generate event response trajectories corresponding to each behavioral event. The event response trajectories corresponding to each behavioral event are collected in the order of the behavioral event sequence to generate a set of event response trajectories corresponding to each behavioral event.

[0022] In this embodiment, the generation of the historical response hysteresis feature sequence specifically includes: Based on the event response trajectory set, the event response trajectory corresponding to each behavior event is selected sequentially according to the order of the behavior events in the behavior event sequence; For each behavioral event, the baseline time range before the event occurs is determined within a preset event analysis time window, and the corresponding baseline state is determined based on the event response trajectory within this baseline time range. The generation of the baseline state specifically includes: Within a preset event analysis time window, the time interval preceding the time anchor point corresponding to the behavioral event is determined as the baseline time range for that behavioral event. Based on the baseline time range, event response trajectory segments within the baseline time range are selected from the event response trajectory corresponding to the behavioral event. The health data corresponding to each time point in the event response trajectory segments are collected to form the baseline response data set corresponding to the behavioral event. Based on the baseline response data set, a baseline state representing the health data status before the occurrence of the behavioral event is generated. Using the time anchor point corresponding to the behavioral event as a reference, the event response trajectory is detected moment by moment within the time range after the event occurs, and the moment when the event response trajectory first deviates continuously from the baseline state is identified to determine the response start time corresponding to the behavioral event. The generation of the response start time specifically includes: Starting from the time anchor point corresponding to the behavioral event, the time range after the event is determined within the event analysis time window. Based on the time range after the event, the event response trajectory corresponding to the behavioral event is detected moment by moment according to the time order of a unified time axis. During the moment-by-moment detection process, the event response trajectory state corresponding to the current time point is compared with the baseline state corresponding to the behavioral event. The time point in the event response trajectory that first appears and deviates from the baseline state at multiple consecutive adjacent time points is identified. The time point that first meets the continuous deviation condition is determined as the response start time corresponding to the behavioral event. In the event response trajectory, based on the response start time, time-by-time detection is continued to identify the moment when the event response trajectory recovers to the baseline state, and the response end time corresponding to the behavior event is determined. The generation of the response end time specifically includes: Starting from the response start time corresponding to the behavioral event, the time range after the response start time is determined within the event analysis time window. Based on the time range, the event response trajectory corresponding to the behavioral event is further detected moment by moment according to the time order of a unified time axis. During the moment-by-moment detection process, the event response trajectory state corresponding to the current time point is compared with the baseline state corresponding to the behavioral event. The time point in the event response trajectory that first appears and remains consistent with the baseline state at multiple consecutive adjacent time points is identified. The time point that first meets the condition of continuous consistency is determined as the response end time corresponding to the behavioral event. Based on the response start time, response end time, and the time anchor point corresponding to the behavioral event, the response start time offset and response duration are extracted. Based on the change trend of the event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration, and response change rate corresponding to the same behavioral event are combined to generate the response hysteresis feature vector corresponding to that behavioral event. The response hysteresis feature vectors corresponding to each behavioral event in the behavioral event sequence are arranged and aggregated according to the time order of the behavioral event sequence to form a historical response hysteresis feature sequence.

[0023] In this embodiment, the generation of the historical hysteresis structure representation specifically includes: In the improved Jamba, a unified time state space is constructed, and the historical response hysteresis feature sequence is written into a unified time axis. A corresponding time index is assigned to each time position after writing, and a forward time state and a reverse time state are established at each time index to form a bidirectional time state pair. The improvements to Jamba specifically include: The structural improvement involves simultaneously constructing a forward time state and a reverse time state for each time index within a unified time state space, binding them into a bidirectional time state pair, thus expanding the time state structure from a unidirectional state structure to a bidirectional coupled state structure. The rule improvement introduces a time state freezing and unfreezing mechanism based on behavioral event time windows, pausing bidirectional time state updates within the event analysis time window and resuming bidirectional time propagation outside the time window, thereby constraining the evolution of time states to behavioral events. The judgment improvement performs time rollback determination based on the consistency relationship between the forward and reverse time states, using bidirectional state consistency as the basis for time evolution legality constraints, and using non-rollback states as the basis for identifying hysteresis structures. Within the improved Jamba, based on the event analysis time window corresponding to the behavior event, when the time index is within the event analysis time window of the behavior event, the bidirectional time state pair is frozen, blocking the state update of the forward time state and the reverse time state. The state update blocking specifically includes: Within the time index range corresponding to the event analysis time window, stop the time recursion update of the forward time state in the bidirectional time state pair, and keep the state value of the forward time state unchanged when entering the event analysis time window; stop the time backtracking update of the reverse time state in the bidirectional time state pair, and keep the state value of the reverse time state unchanged when entering the event analysis time window; during the state update blocking period, prohibit the forward time state and the reverse time state from participating in the state transfer and state synthesis of subsequent time index positions, so that the time state within the event analysis time window does not participate in the time evolution process, and complete the state update blocking process; When the time index exceeds the event analysis time window, the freezing constraint of the bidirectional time state pair is released. In the improved Jamba, based on the time state corresponding to the historical response hysteresis feature sequence that has been written into the unified time state space, forward time propagation is performed on the forward time state in the direction of increasing time index, and reverse time propagation is performed on the reverse time state in the direction of decreasing time index, forming a bidirectional time propagation state sequence that evolves in parallel within the same time state space. During bidirectional time propagation, at each time index, the corresponding forward and reverse time states are analyzed for state consistency in the improved Jamba. Based on the state consistency relationship between the forward and reverse time states, a non-reversibility determination is performed. When the forward and reverse time states meet the consistency condition, the time state at that time index is determined to be a reversible state. When the forward and reverse time states do not meet the consistency condition, the time state at that time index is determined to be a non-reversible state. The generation of the irreversibility determination specifically includes: During bidirectional time propagation, the corresponding forward and reverse time states are obtained at each time index, and these two are treated as a pair of time states under the same time index for joint processing. Based on the correspondence between the forward and reverse time states in the state space, the state difference results between them in the state representation dimension are calculated. The state difference results at adjacent time indices are continuously tracked along the time index direction. When the state difference results show a convergence trend during forward and reverse time propagation, it is determined that the forward and reverse time states have the ability to backtrack to each other at that time index, and the time state at that time index is determined to be a reversible state. When the state difference results continue to remain or amplify during forward and reverse time propagation, and there is no trend of convergence to the same state region, it is determined that the forward and reverse time states cannot backtrack to each other at that time index, and the time state at that time index is determined to be a non-reversible state. Write the reversible and non-reversible states generated at each time index into the time state space as time state attributes, and use the non-reversible state as a blocking flag for the evolution of time state. The set of non-reversible time indices corresponding to each behavioral event within the event analysis time window is collected, and then organized in conjunction with the time anchor points corresponding to each behavioral event. A historical hysteresis structure representation is generated according to the time sequence of the behavioral events. The generation of the historical hysteresis structure representation specifically includes: In the improved Jamba, for each behavioral event, its corresponding event analysis time window is read. Time indices within that time window and deemed irreversible are extracted from the unified time state space, generating a set of irreversible time indices for that behavioral event. This set is then time-aligned with the time anchor point of the behavioral event. The time offset of each irreversible time index relative to the time anchor point is calculated and sorted in ascending order of time offset, generating a hysteresis time offset sequence for that behavioral event. For each hysteresis time offset sequence, its start offset position, end offset position, and the distribution range of consecutive irreversible time indices are recorded, generating an event-level hysteresis description result that describes the hysteresis duration and distribution pattern of the behavioral event. The event-level hysteresis description results for each behavioral event are arranged and aggregated according to the chronological order of occurrence of the behavioral events in the behavioral event sequence, forming a hysteresis description sequence organized chronologically. Finally, the hysteresis description sequence is output as a whole to generate a historical hysteresis structure representation.

[0024] In this embodiment, the generation of the real-time hysteresis structure characterization results specifically includes: Based on the real-time collected health data and behavioral events, organize and generate real-time health data time series and real-time behavioral event sequences; Based on real-time health data time series and real-time behavioral event series, a unified time axis is established to generate real-time unified time axis alignment results; Using the real-time unified time axis alignment results, the timestamps of each behavioral event in the real-time behavioral event sequence are used as time anchors. Within the event analysis time window, corresponding health data segments are continuously selected from the real-time health data time sequence. The health data segments are organized according to the health data type to generate real-time event response trajectories corresponding to each behavioral event, and then aggregated to form a set of real-time event response trajectories. For each real-time event response trajectory in the real-time event response trajectory set, a baseline time range before the event occurs is determined based on the time anchor point. Within the baseline time range, the states of the real-time event response trajectories are aggregated to generate the baseline state of the corresponding behavioral event. Using the time anchor point as a reference, the real-time event response trajectory is detected hour by hour within the time range after the event occurs. After the response start time, the moment when the real-time event response trajectory recovers to the baseline state is detected. The moment when the real-time event response trajectory first continuously deviates from the baseline state is identified to determine the response start time, and the moment when the response end time is determined. Based on the response start time, response end time and time anchor point, the response start time offset and response duration are extracted. Based on the change trend of the real-time event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration and response change rate are combined to generate a real-time response hysteresis feature vector. The real-time response hysteresis feature vectors corresponding to each behavioral event in the real-time behavioral event sequence are arranged and aggregated in chronological order to form a real-time response hysteresis feature sequence. Under the constraint of historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba. Using the bidirectional time propagation mechanism, the time state evolution of the real-time response hysteresis feature sequence is carried out at different time scales. Based on the backtrackable constraint conditions formed by historical hysteresis structure representation, the backtracking judgment processing of time state is performed to generate the real-time hysteresis structure representation result. The generation of the real-time hysteresis structure characterization results specifically includes: Under the reversible constraint of the historical hysteresis structure representation, the real-time response hysteresis feature sequence is written into the improved Jamba time state space in time index order. At each time index, corresponding forward and reverse time states are simultaneously established. Forward time propagation is performed on the forward time state in the increasing direction of the time index, and reverse time propagation is performed on the reverse time state in the decreasing direction of the time index, enabling the real-time response hysteresis feature sequence to form a bidirectional time propagation time state evolution trajectory within the same time state space. During the bidirectional time propagation process, the forward and reverse time states are compared for state consistency at each time index. Based on the reversible constraint of the corresponding behavioral events in the historical hysteresis structure representation, the current time index is... The feasibility of time state rollback is determined. When the forward and reverse time states are consistent under the constraints of the historical hysteresis structure representation, the time state at that time index is determined to be a rollback state, and the time state is allowed to continue to evolve along the bidirectional propagation direction. When the forward and reverse time states are inconsistent under the constraints of the historical hysteresis structure representation, the time state at that time index is determined to be a non-rollback state, and a blocking mark for time state evolution is formed at that time index. The rollback and non-rollback states generated at each time index are recorded in the order of the time index. Combined with the time anchor points of the corresponding behavioral events, the time state evolution results formed by the real-time response hysteresis feature sequence within the event analysis time window are organized to generate real-time hysteresis structure representation results.

[0025] In this embodiment, the generation of the health monitoring results specifically includes: Based on the real-time hysteresis structure representation results, a set of historical hysteresis structure representation results corresponding to the real-time hysteresis structure representation results is obtained. The set of historical hysteresis structure representation results consists of hysteresis structure representation results formed during the historical health data response process. The real-time hysteresis structure representation results are compared with the set of historical hysteresis structure representation results in the structure representation space to generate a real-time hysteresis structure deviation representation. The generation of the real-time hysteresis structure deviation representation specifically includes: The real-time hysteresis structure representation results are mapped to a structure representation space consistent with the set of historical hysteresis structure representation results, forming a real-time structure representation vector. Each historical hysteresis structure representation result is sequentially selected from the set of historical hysteresis structure representation results, and these results are uniformly represented as a set of historical structure representation vectors in the structure representation space. In the structure representation space, using the real-time structure representation vector as a reference, the structural difference relationship between the real-time structure representation vector and each historical structure representation vector is calculated, generating a corresponding set of structural difference vectors. The set of structural difference vectors is then organized and aggregated in chronological order to form a structural difference sequence. Based on this sequence, the offset direction, offset amplitude, and offset distribution characteristics of the real-time structure representation vector in the structure representation space are jointly represented to generate a real-time hysteresis structure deviation representation. Based on the real-time hysteresis structure deviation characterization, the structural change state of the real-time health data response process relative to the historical health data response process is determined. When the structural change state meets the criteria for structural hysteresis, a health monitoring result is generated that characterizes the existence of structural hysteresis in the health data response process. When the structural change state does not meet the criteria for structural hysteresis, a health monitoring result is generated that represents the health data response process without structural hysteresis.

[0026] Example 1: To verify the feasibility of the present invention in practice, it was applied to a health monitoring task. In this task, our goal is to collect users' health data and behavioral events through a smartwatch, thereby accurately monitoring the users' health status, especially to promptly judge changes in the users' health status when faced with changes in the situation. Specifically, our goal is to capture the lag characteristics of users' health data and determine whether there is structural lag in the health data, thereby providing a more accurate basis for health risk warning.

[0027] In a specific application scenario, a user wears a smartwatch and monitors their health during daily activities. The smartwatch continuously collects the user's health data, such as heart rate, blood oxygen, body surface temperature, angular velocity of posture changes, and sleep duration. After preprocessing, this data generates a time-series sequence of health data. Furthermore, it triggers behavioral event recordings through sensor interactions and user behavior, including user activities such as walking, running, sitting, standing, and certain special actions such as high-intensity exercise or eating. Each event is timestamped and associated with the corresponding health data segment.

[0028] By combining these health data with behavioral events to generate a unified timeline, the smartwatch's algorithm can select corresponding health data segments from the health data time series within the event analysis time window, thereby generating an event response trajectory related to each behavioral event. This process is crucial for subsequent health status analysis and lag feature extraction because it helps us establish a precise correlation between each health data and behavioral event, providing a foundation for subsequent health monitoring and lag judgment.

[0029] Hysteresis features such as response start time offset, response duration, and response change rate were extracted between each health data segment and the event response trajectory, thereby generating a historical response hysteresis feature sequence. Since changes in health data usually have a certain lag effect, traditional one-way time models often cannot accurately capture this hysteresis phenomenon, thus affecting the early warning effect of health risks. In this invention, we input the historical response hysteresis feature sequence into an improved Jamba model and use its two-way time propagation mechanism to model the forward and reverse time states respectively, thereby achieving a fine capture of the hysteresis phenomenon in health data. In this way, two-way time propagation not only considers the past state of the data, but also incorporates future influencing factors in reverse, enabling the judgment of the changing trend of health status on a larger scale.

[0030] Through the two-way time propagation mechanism, we can determine whether the health status at each time index is a reversible state. If the current health status meets the reversal rules of historical data, it is considered to be in a reversible state, and the model continues to advance the prediction of the health status along the time direction. If the current health status cannot be reversed to the historical state, the model will make a non-reversible judgment on it, and further remind us of possible health risks or anomalies.

[0031] In practical applications, we have observed that, compared with traditional health monitoring methods, the model of this invention can more accurately capture changes in health status and make more timely and accurate judgments on abnormal health states. For example, during a day of monitoring, a user engages in high-intensity exercise. Before the exercise, the user's heart rate and body surface temperature data are normal, but after the exercise, the heart rate and body temperature rise abnormally. Traditional health monitoring methods would treat these data as instantaneous fluctuations and would not be able to determine whether there is a health problem. However, through the bidirectional time propagation mechanism of this invention, our model can identify the delay effect that occurs during exercise, capture the lag changes in heart rate and body surface temperature, and, based on the backtracking constraints of historical data, promptly detect that the health state cannot regress to a normal state, generating an early warning of health abnormalities.

[0032] The application of this invention in this scenario solves the problem that traditional methods cannot effectively identify hysteresis phenomena in health data and whether health status is reversible. By introducing a two-way time propagation mechanism and a reversibility determination, this invention is superior to existing technologies in terms of accuracy and responsiveness. Its superiority is particularly prominent when it involves the determination of dynamic changes in health status and hysteresis effects.

[0033] Table 1. Performance Comparison of Different Health Data Monitoring Methods in Structural Hysteresis Detection

[0034] As can be seen from Table 1, traditional threshold-based health monitoring methods are feasible for instantaneous judgment of single physiological indicators, but they have significant shortcomings in health data response modeling after behavioral events are triggered. For example, although the accuracy of response judgment for indicators such as heart rate and blood oxygen can reach around 0.80, the accuracy of identifying the start and end times of the response after a behavioral event decreases significantly, reflecting its limited ability to characterize time lag characteristics. This method cannot distinguish between short-term fluctuations and structural changes, which leads to significant fluctuations in judgment results under continuous behavioral events or long-term wear scenarios.

[0035] In comparison, traditional unidirectional time-series modeling methods improve the identification of response start and end points compared to threshold methods by introducing time series modeling, with its accuracy increasing to the range of 0.79–0.80 and the volatility of judgment under continuous behavioral event conditions decreasing. However, these methods still rely solely on forward time evolution and lack clear criteria for whether health data can regress to a historical stable state. Therefore, the improvement in the consistency rate of structural hysteresis judgment and stability under long-term wear conditions is limited.

[0036] The method of this invention shows a more balanced and stable improvement in all the above indicators. In particular, in terms of the consistency rate of structural hysteresis judgment, the method of this invention reaches 0.81, which is about 0.11 and 0.06 higher than the traditional threshold method and the one-way time series model method, respectively. This difference does not come from the simple sum of the accuracy of a single indicator, but from the fact that this invention uses the improved Jamba two-way time propagation mechanism to constrain the reversibility of health data in the time dimension, so that the model can identify whether the health response has undergone irreversible changes from the time structure level.

[0037] As can be seen from the volatility and long-term drift amplitude, the method of the present invention can still maintain low result volatility under continuous behavioral events and long-term wear conditions. This indicates that the historical hysteresis structure representation as a constraint effectively suppresses random drift in the real-time judgment process. This characteristic is particularly important for wearable health monitoring scenarios, because behavioral events are frequent and physiological states fluctuate naturally during actual use. Without structural constraints, the monitoring results are very likely to be unstable.

[0038] In summary, the method of this invention does not merely improve a single numerical indicator, but rather achieves synergistic improvements in multiple key aspects such as response initiation identification, hysteresis feature stability, and consistency of structural hysteresis determination. This elevates health data monitoring from "instantaneous judgment" to "structural judgment," significantly enhancing the reliability and continuity of monitoring results in practical application scenarios.

[0039] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for monitoring health data in a smartwatch based on machine learning, characterized in that, Includes the following steps: Health data is collected on the smartwatch and preprocessed to generate a time-series health data sequence. Collect behavioral events generated during the use of smartwatches and organize them to generate behavioral event sequences; A unified timeline is constructed based on health data time series and behavioral event series. The behavioral event timestamp is used as the time anchor point. Within the preset event analysis time window, the corresponding health data is selected to generate a set of event response trajectories corresponding to each behavioral event. The event response trajectory in the event response trajectory set is identified by the response start time and determined by the response end time. The response start time offset, response duration and response change rate are extracted to generate response hysteresis feature vectors and form a historical response hysteresis feature sequence in chronological order. Using historical response hysteresis feature sequences as input, an improved Jamba is constructed. A time state evolution constraint mechanism based on behavioral events is introduced to model the historical response hysteresis feature sequences under constrained time evolution rules, generating a historical hysteresis structure representation. For real-time collected health data and behavioral events, a real-time response hysteresis feature sequence is generated. Under the backtracking constraint formed by the historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba, and the constrained time state evolution and backtracking decision processing are performed to generate the real-time hysteresis structure representation result. Based on the real-time hysteresis structure characterization results, a health monitoring result is outputting whether there is structural hysteresis in the health data response process.

2. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The health data specifically includes heart rate data, heart rate variability data, blood oxygen saturation data, body surface temperature data, posture change angular velocity data, and sleep duration data. The preprocessing specifically includes outlier removal, missing data completion, unified time index alignment, and amplitude normalization.

3. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the behavioral event sequence specifically includes: During the use of the smartwatch, event records corresponding to the user's behavior are collected, including event records triggered by user operations, event records triggered by sensors, and event records generated during operation. Each event record is associated with corresponding timestamp information. The collected event records are filtered for validity, and event records with missing timestamp information, incomplete event type identification, or abnormal occurrence time are removed to form a set of valid event records; Based on the preset behavioral event classification rules, the event records in the valid event record set are merged by event type, and event records with the same event type identifier and occurring consecutively within a preset time interval are merged into the same behavioral event; The time boundaries of the merged behavioral events are determined, the start time and end time of each behavioral event are determined, and the merged behavioral events are mapped to behavioral event items with unique event identifiers. All behavioral event items are sorted according to their start time, forming a sequence of behavioral events arranged in chronological order.

4. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the event response trajectory set specifically includes: A unified timeline is established based on health data time series and behavioral event series, and the health data time series and behavioral event series are time-aligned. In the unified timeline, the timestamp of each behavioral event in the sequence of behavioral events is used as the time anchor point to determine the anchor position of each behavioral event on the unified timeline. For each behavioral event, the time range is determined at the corresponding time anchor point according to the preset event analysis time window. The preset event analysis time window consists of the analysis time range before the event occurs and the analysis time range after the event occurs. Based on a unified timeline and a preset event analysis time window, health data segments corresponding to the time anchor points of each behavioral event are selected from the health data time series. Health data fragments are organized according to health data type to generate event response trajectories corresponding to single behavioral events; For each behavioral event in the behavioral event sequence, the time anchor point determination, time window limitation, and health data selection and processing are repeatedly executed to generate event response trajectories corresponding to each behavioral event. The event response trajectories corresponding to each behavioral event are collected in the order of the behavioral event sequence to generate a set of event response trajectories corresponding to each behavioral event.

5. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the historical response hysteresis feature sequence specifically includes: Based on the event response trajectory set, the event response trajectory corresponding to each behavior event is selected sequentially according to the order of the behavior events in the behavior event sequence; For each behavioral event, the baseline time range before the event occurs is determined within a preset event analysis time window, and the corresponding baseline state is determined based on the event response trajectory within this baseline time range. Using the time anchor point corresponding to the behavioral event as a reference, the event response trajectory is detected moment by moment within the time range after the event occurs, and the moment when the event response trajectory first deviates continuously from the baseline state is identified to determine the response start time corresponding to the behavioral event. In the event response trajectory, based on the response start time, time-by-time detection is continued to identify the moment when the event response trajectory recovers to the baseline state, and the response end time corresponding to the behavior event is determined. Based on the response start time, response end time, and the time anchor point corresponding to the behavioral event, the response start time offset and response duration are extracted. Based on the change trend of the event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration, and response change rate corresponding to the same behavioral event are combined to generate the response hysteresis feature vector corresponding to that behavioral event. The response hysteresis feature vectors corresponding to each behavioral event in the behavioral event sequence are arranged and aggregated according to the time order of the behavioral event sequence to form a historical response hysteresis feature sequence.

6. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the historical hysteresis structure representation specifically includes: In the improved Jamba, a unified time state space is constructed, and the historical response hysteresis feature sequence is written into a unified time axis. A corresponding time index is assigned to each time position after writing, and a forward time state and a reverse time state are established at each time index to form a bidirectional time state pair. Within the improved Jamba, based on the event analysis time window corresponding to the behavior event, when the time index is within the event analysis time window of the behavior event, the bidirectional time state pair is frozen, blocking the state update of the forward time state and the reverse time state. When the time index exceeds the event analysis time window, the freezing constraint of the bidirectional time state pair is released. In the improved Jamba, based on the time state corresponding to the historical response hysteresis feature sequence that has been written into the unified time state space, forward time propagation is performed on the forward time state in the direction of increasing time index, and reverse time propagation is performed on the reverse time state in the direction of decreasing time index, forming a bidirectional time propagation state sequence that evolves in parallel within the same time state space. During bidirectional time propagation, at each time index, the corresponding forward and reverse time states are analyzed for state consistency in the improved Jamba. Based on the state consistency relationship between the forward and reverse time states, a non-reversibility determination is performed. When the forward and reverse time states meet the consistency condition, the time state at that time index is determined to be a reversible state. When the forward and reverse time states do not meet the consistency condition, the time state at that time index is determined to be a non-reversible state. Write the reversible and non-reversible states generated at each time index into the time state space as time state attributes, and use the non-reversible state as a blocking flag for the evolution of time state. The set of non-reversible time indices corresponding to each behavioral event within the event analysis time window is collected, and then organized in conjunction with the time anchor points corresponding to each behavioral event. A historical lag structure representation is generated according to the time sequence of the behavioral events.

7. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the real-time hysteresis structure characterization results specifically includes: Based on the real-time collected health data and behavioral events, organize and generate real-time health data time series and real-time behavioral event sequences; Based on real-time health data time series and real-time behavioral event series, a unified time axis is established to generate real-time unified time axis alignment results; Using the real-time unified time axis alignment results, the timestamps of each behavioral event in the real-time behavioral event sequence are used as time anchors. Within the event analysis time window, corresponding health data segments are continuously selected from the real-time health data time sequence. The health data segments are organized according to the health data type to generate real-time event response trajectories corresponding to each behavioral event, and then aggregated to form a set of real-time event response trajectories. For each real-time event response trajectory in the real-time event response trajectory set, a baseline time range before the event occurs is determined based on the time anchor point. Within the baseline time range, the states of the real-time event response trajectories are aggregated to generate the baseline state of the corresponding behavioral event. Using the time anchor point as a reference, the real-time event response trajectory is detected hour by hour within the time range after the event occurs. After the response start time, the moment when the real-time event response trajectory recovers to the baseline state is detected. The moment when the real-time event response trajectory first continuously deviates from the baseline state is identified to determine the response start time, and the moment when the response end time is determined. Based on the response start time, response end time and time anchor point, the response start time offset and response duration are extracted. Based on the change trend of the real-time event response trajectory between the response start time and response end time, the response change rate is extracted. The response start time offset, response duration and response change rate are combined to generate a real-time response hysteresis feature vector. The real-time response hysteresis feature vectors corresponding to each behavioral event in the real-time behavioral event sequence are arranged and aggregated in chronological order to form a real-time response hysteresis feature sequence. Under the constraint of historical hysteresis structure representation, the real-time response hysteresis feature sequence is input into the improved Jamba. Using a two-way time propagation mechanism, the time state evolution of the real-time response hysteresis feature sequence is carried out at different time scales. Based on the backtrackable constraint conditions formed by the historical hysteresis structure representation, the backtracking decision processing of the time state is performed to generate the real-time hysteresis structure representation result.

8. The method for monitoring health data of a smartwatch based on machine learning according to claim 1, characterized in that, The generation of the health monitoring results specifically includes: Based on the real-time hysteresis structure representation results, a set of historical hysteresis structure representation results corresponding to the real-time hysteresis structure representation results is obtained. The set of historical hysteresis structure representation results consists of hysteresis structure representation results formed during the historical health data response process. The real-time hysteresis structure representation results are compared with the set of historical hysteresis structure representation results in the structure representation space to generate a real-time hysteresis structure deviation representation. Based on the real-time hysteresis structure deviation characterization, the structural change state of the real-time health data response process relative to the historical health data response process is determined. When the structural change state meets the criteria for structural hysteresis, a health monitoring result is generated that characterizes the existence of structural hysteresis in the health data response process. When the structural change state does not meet the criteria for structural hysteresis, a health monitoring result is generated that represents the health data response process without structural hysteresis.

Images