A sports breathing intelligent prediction method and a sports health management system

Abstract

The application provides a motion breathing intelligent prediction method and a motion health management system, belongs to the technical field of motion monitoring and management, collects multi-modal time sequence data, detects key motion breathing events from each modal time sequence data respectively, and performs time alignment processing on each modal time sequence data, takes the feature vectors of each modal at different event nodes after time alignment as nodes, takes fuzzy correlation degrees calculated based on event time differences, feature similarities and event importance as edge weights, constructs a space-time heterogeneous graph, performs cross-modal feature fusion on breathing event features of each modal based on the space-time heterogeneous graph, obtains multi-modal fusion features, predicts breathing states in a motion process according to the multi-modal fusion features, can construct a unified key motion breathing event correlation mechanism for multi-modal physiological time sequence data, realizes accurate time alignment and cross-modal deep fusion of different modal breathing events, and improves the prediction accuracy and stability of breathing states in the motion process.

Description

Technical Field

[0001] This application relates to the field of sports monitoring and management technology, and more specifically, to a method for intelligent prediction of exercise respiration and a sports health management system. Background Technology

[0002] With the development of smart wearable devices and sports health management technologies, wearable devices have been widely used in fields such as sports monitoring, heart rate monitoring, and blood oxygen monitoring. Among these, respiratory status is of great significance for sports safety assessment and sports intensity adjustment. Existing wearable devices typically collect physiological data during exercise using smart sensors such as photoplethysmography (PPG) sensors, blood oxygen sensors, and inertial sensors. They then use single physiological parameters such as heart rate changes, blood oxygen changes, or exercise intensity, or employ simple data fusion algorithms to estimate or predict respiratory status such as respiratory rate and rhythm, in order to achieve sports health monitoring and health management.

[0003] However, respiratory regulation is a complex dynamic process involving exercise load, cardiovascular circulation, oxygen transport, and autonomic nervous system regulation. Different physiological signals exhibit significant differences in response time and mode to the same respiratory activity. For example, changes in exercise intensity typically manifest first as changes in motion state acquired by inertial sensors, while physiological parameters such as heart rate, pulse wave, and blood oxygen saturation show varying degrees of response delay due to physiological conduction processes, resulting in asynchrony in the temporal dimension of time-series data across different modalities. Simultaneously, different modalities differ in their sensitivity and representational ability regarding respiratory states. Traditional methods often directly sample or simply stitch together raw multimodal data, failing to establish correspondences between key exercise-respiratory events across different modalities or effectively characterize the temporal evolution of correlations between different physiological modalities. This easily leads to respiratory event mismatches, insufficient cross-modal feature fusion, and low accuracy in predicting respiratory states during exercise. Therefore, constructing a unified key exercise-respiratory event correlation mechanism for multimodal physiological time-series data, achieving precise temporal alignment and deep cross-modal fusion of respiratory events across different modalities, and improving the accuracy and stability of respiratory state prediction during exercise remains a challenging problem for the industry. Summary of the Invention

[0004] This application provides an intelligent prediction method for exercise-induced breathing and an exercise health management system. It can construct a unified key exercise-induced breathing event association mechanism for multimodal physiological time-series data, realize accurate time alignment and deep cross-modal fusion of breathing events of different modalities, and improve the prediction accuracy and stability of breathing status during exercise.

[0005] In a first aspect, this application provides a method for intelligent prediction of respiratory movement, comprising the following steps: By using smart sensors deployed inside wearable devices, multimodal time-series data related to respiratory function can be collected during exercise. Key motor and respiratory events are detected from the time series data of each modality, and time alignment is performed on the time series data of each modality based on the time series of key motor and respiratory events of all modalities. Using the feature vectors of each modality at different event nodes after time alignment as nodes, and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights, a spatiotemporal heterogeneous graph is constructed. Based on the spatiotemporal heterogeneous graph, cross-modal feature fusion is performed on the respiratory event features of each modality to obtain multimodal fusion features characterizing the respiratory state. Based on the multimodal fusion features, the respiratory state during the movement process is predicted.

[0006] In some embodiments, the multimodal time-series data includes photoplethysmography pulse wave time-series data, blood oxygen saturation time-series data, triaxial motion inertial time-series data, and skin temperature time-series data.

[0007] In some embodiments, detecting key motor respiratory events from each modal time series data specifically includes: For each mode, denoising and stabilization processing is performed on the modal time series data to eliminate high-frequency interference and suppress baseline drift, resulting in a continuous modal signal sequence. Extract fundamental temporal features from the modal signal sequence to characterize the respiratory cycle structure; Based on the aforementioned basic temporal characteristics, periodic peaks, troughs, and amplitude variation nodes of the modal signal are identified to generate a candidate respiratory event set; Preset timing and amplitude rules are applied to the candidate respiratory event set to filter out key motion respiratory events of the corresponding modality.

[0008] In some embodiments, time alignment processing of the temporal data of each modality based on the timing of key motor respiratory events across all modalities specifically includes: Acquire key motion-respiratory events and their corresponding event time identifiers obtained from each modality detection; Key motor respiratory events of different modalities are sorted by time according to the event time identifier; For cross-modal key motion-respiratory events that are within a preset time neighborhood and point to the same respiratory cycle stage, perform event merging processing to generate a unified event node; Organize the event node sequence according to the time order of unified event nodes to construct an event reference timeline; Using the event reference timeline as a reference, time alignment processing is performed on the timing data of each modality.

[0009] In some embodiments, time alignment processing of the modal time series data based on the event reference time axis specifically includes: Map the key motion and breathing events of each modality to the corresponding unified event nodes in the event reference timeline; Calculate the time offset of each modal key motion breathing event relative to the unified event node; Based on the time offset, the time series data of the corresponding mode is time-shifted to generate time-synchronized multimodal time series data.

[0010] In some embodiments, the spatiotemporal heterogeneous graph is constructed using the feature vectors of each modality at different event nodes after time alignment as nodes, and the fuzzy correlation degree calculated based on event time difference, feature similarity, and event importance as edge weights. This specifically includes: In time-aligned multimodal time series data, respiratory event feature vectors of the corresponding modality are extracted at each event node, and the respiratory event feature vectors are used as heterogeneous graph nodes for modeling. Based on the event reference timeline, the time interval parameters between event nodes are determined to characterize the temporal adjacency relationship between nodes; Similarity analysis is performed on the respiratory event feature vectors of different modal event nodes to obtain feature similarity parameters used to characterize the degree of cross-modal correlation. The feature similarity parameters are parameters used to characterize the degree of correlation between different modal event nodes at the feature level. Configure event importance parameters for event nodes based on their structural position in the event sequence; The time interval parameter, feature similarity parameter, and event importance parameter are input into the fuzzy inference model, and the fuzzy correlation degree between nodes is generated according to the preset fuzzy rules. Using the fuzzy correlation degree as the edge weight, a connection relationship is established between event nodes to obtain a spatiotemporal heterogeneous graph that integrates temporal correlation and cross-modal correlation.

[0011] In some embodiments, predicting the respiratory state during motion based on the multimodal fusion features specifically includes: Perform feature scale normalization processing on the multimodal fusion features; The normalized multimodal fusion features are input into a pre-defined prediction network; Based on the prediction network, the multimodal fusion features are processed to perform pattern discrimination, generating corresponding breathing state determination information, and then outputting the breathing state during the movement process.

[0012] Secondly, this application provides a sports health management system, including a sports breathing prediction unit for executing a sports breathing intelligent prediction method, wherein the sports breathing prediction unit includes: The data acquisition module is used to collect multimodal time-series data related to respiratory function during exercise through smart sensors deployed inside the wearable device; The time alignment module is used to detect key motion and respiratory events from the time series data of each modality, and to perform time alignment processing on the time series data of each modality based on the time series of key motion and respiratory events of all modalities. The heterogeneous graph construction module is used to construct a spatiotemporal heterogeneous graph by using the feature vectors of each modality at different event nodes after time alignment as nodes and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights. The breathing prediction module is used to perform cross-modal feature fusion on the breathing event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fusion features characterizing the breathing state, and predict the breathing state during the movement process based on the multimodal fusion features.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent prediction method for motor respiration.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent prediction method for motor respiration.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: This application provides a method for intelligent prediction of respiration during exercise and a sports health management system. The method involves collecting multimodal time-series data related to respiratory function during exercise using intelligent sensors deployed within a wearable device. Key respiration events are detected from each modality's time-series data, and the data is time-aligned based on the temporal sequence of these key respiration events. A spatiotemporal heterogeneous graph is constructed using the feature vectors of each modality at different event nodes as nodes and the fuzzy correlation degree calculated based on event time difference, feature similarity, and event importance as edge weights. Cross-modal feature fusion is performed on the respiratory event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fusion features representing the respiratory state. The respiratory state during exercise is predicted based on these multimodal fusion features.

[0016] Therefore, this application firstly detects key motion-respiratory events from the temporal data of each modality, and performs time alignment processing on the data of each modality based on the temporal relationship of key motion-respiratory events of all modalities. This transforms the asynchronous physiological data, which was originally caused by differences in sensor sampling frequency, delays in physiological response transmission, and different signal change mechanisms, into correlated data with a unified time reference. Then, using the feature vectors of each modality at different event nodes as nodes, a spatiotemporal heterogeneous graph is constructed. This transforms the originally independent multimodal physiological data during exercise into a unified graph structure with temporal evolution relationships and cross-modal correlation relationships, thus breaking through the limitations of traditional methods that rely solely on time sequence. This method overcomes the limitations of data fusion through sequential or simple feature splicing. Finally, it performs cross-modal feature fusion on respiratory event features of each modality based on spatiotemporal heterogeneous graphs, and uses the fused multimodal fusion features to predict respiratory state during exercise. This fully utilizes the complementary information between different physiological modalities in the temporal and physiological functional dimensions, avoiding the information fragmentation and missing correlation problems caused by traditional methods that use single physiological signals or simple feature splicing. Furthermore, it achieves accurate time alignment and deep cross-modal fusion of respiratory events of different modalities based on spatiotemporal heterogeneous graphs, which can improve the completeness, consistency and discriminative ability of respiratory state representation during exercise, and ultimately improve the accuracy and stability of respiratory state prediction.

[0017] In summary, the technical solution adopted in this application can construct a unified key motion respiratory event association mechanism for multimodal physiological time series data, realize accurate time alignment and cross-modal deep fusion of respiratory events of different modalities, and improve the accuracy and stability of respiratory state prediction during exercise. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an exemplary flowchart of an intelligent prediction method for exercise-induced breathing according to some embodiments of this application; Figure 2 These are schematic diagrams of the structure of a wearable device according to some embodiments of this application; Figure 3 This is an exemplary flowchart illustrating the detection of key motor respiratory events according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a motor breathing prediction unit according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device for implementing an intelligent prediction method for motor respiration, according to some embodiments of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] To better understand the above technical solutions, a detailed description of the technical solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. (Refer to...) Figure 1 This figure is an exemplary flowchart of an intelligent prediction method for exercise-induced breathing according to some embodiments of this application. The figure mainly includes the following steps: In step S101, multimodal time-series data related to respiratory function during exercise are collected by smart sensors deployed inside the wearable device.

[0022] It should be noted that, in this application, the multimodal time-series data includes photoplethysmography pulse wave time-series data, blood oxygen saturation time-series data, triaxial motion inertial time-series data, and skin temperature time-series data.

[0023] like Figure 2 As shown, the wearable device integrates a photoplethysmography (PPG) sensor, a blood oxygen saturation sensor, a microelectromechanical inertial sensor, and a skin temperature sensor.

[0024] The photoplethysmography (PPG) sensor is positioned on the back of the wristband where it contacts the skin of the wrist. It emits detection light of a specific wavelength into the skin tissue through a light-emitting unit and collects the light intensity change signal after reflection by the skin tissue using a photoelectric receiving unit. Since the blood volume changes with the cardiac cycle during heartbeat, and respiratory activity affects heart rate fluctuations and pulse wave morphology through changes in thoracic pressure, autonomic nervous system regulation, and venous return, the collected PPG time series data can characterize the cardiopulmonary coupling changes during exercise.

[0025] The blood oxygen saturation detection sensor uses light sources of different wavelengths to alternately irradiate wrist tissue. By detecting the difference in light intensity attenuation after the light of different wavelengths is absorbed by blood tissue, the ratio of oxyhemoglobin to deoxyhemoglobin is calculated, thereby obtaining time-series data of blood oxygen saturation during exercise.

[0026] The triaxial motion inertial sensor includes a triaxial acceleration detection unit and a triaxial angular velocity detection unit. It acquires triaxial motion inertial time-series data by collecting the linear acceleration changes and rotational motion changes during the user's wrist movement in real time.

[0027] The skin temperature sensor continuously collects data on skin temperature changes during exercise by attaching to a temperature detection area on the skin's surface. Since increased muscle activity during exercise leads to increased metabolic heat production, and sympathetic nerve excitation causes peripheral vasoconstriction or dilation, thereby altering the heat exchange process on the skin surface, changes in skin temperature can indirectly reflect the body's exercise load and recovery status.

[0028] Finally, the time-series data of photoplethysmography (PPG), blood oxygen saturation, triaxial motion inertia, and skin temperature collected by the different sensors were timestamped, and a multimodal synchronous acquisition mechanism was established based on the sampling frequency, signal response delay, and physiological conduction path differences of each sensor. Specifically, motion inertia time-series data was used as a reference for motion stimulation, PPG time-series data as a reference for circulatory response, blood oxygen saturation time-series data as a reference for oxygen exchange response, and skin temperature time-series data as a reference for metabolic regulation. This resulted in a multimodal respiratory-related time-series data set with characteristics of exercise induction, cardiopulmonary response, oxygen metabolism feedback, and thermoregulation.

[0029] In step S102, key motion and breathing events are detected from the time series data of each modality, and time alignment processing is performed on the time series data of each modality based on the time series of key motion and breathing events of all modalities.

[0030] Preferably, in some embodiments, reference is made to Figure 3 As shown, this figure is an exemplary flowchart of detecting key motor and respiratory events according to some embodiments of this application. In this embodiment, the detection of key motor and respiratory events from the time-series data of each modality can be achieved by the following steps: In step S1021, for each mode, denoising and stabilization processing are performed on the modal time series data to eliminate high-frequency interference and suppress baseline drift, thereby obtaining a continuous modal signal sequence. In step S1022, basic temporal features for characterizing the respiratory cycle structure are extracted from the modal signal sequence; In step S1023, the periodic peaks, troughs, and amplitude change nodes of the modal signal are identified based on the basic time-series characteristics to generate a candidate respiratory event set; In step S1024, preset timing rules and amplitude rules are applied to the candidate respiratory event set to filter out key motion respiratory events of the corresponding modality.

[0031] In practice, firstly, for each mode, denoising and stabilization processing can be performed on the modal time series data to eliminate high-frequency interference and suppress baseline drift. Since wrist swing, muscle vibration, and environmental interference during movement will introduce different forms of noise into different modal signals, an adaptive filtering model is first constructed based on the spectral distribution characteristics of the corresponding modal signal. For photoplethysmography pulse wave time series data, motion artifacts caused by wrist swing are suppressed; for blood oxygen saturation time series data, random sampling fluctuations and optical interference are suppressed; for triaxial motion inertial time series data, high-frequency vibration components caused by instantaneous impact are filtered out; and for skin temperature time series data, short-term interference caused by rapid fluctuations in ambient temperature is filtered out. Subsequently, the filtered signal is used as input to stabilize the time series data of each mode. By establishing a local sliding reference window, the local mean and local variation trend of the signal within each time window are calculated, and the local trend term is used to compensate for the original signal to eliminate baseline drift caused by long-term sensor drift, changes in wearing tightness, and changes in movement posture, so that each modal signal fluctuates around a stable reference, resulting in a continuous and stable modal signal sequence.

[0032] Furthermore, fundamental temporal features for characterizing the respiratory cycle structure can be extracted from the modal signal sequence. That is, using a continuous and stable modal signal sequence as input, fundamental temporal features for characterizing the respiratory cycle structure are extracted within each continuous time window. These fundamental temporal features include local waveform amplitude, rate of change of adjacent sampling points, local curvature change, energy distribution density, and periodic repetition consistency. The local waveform amplitude is used to describe the current respiratory modulation intensity, the rate of change of adjacent sampling points is used to reflect the speed of change of respiratory state, the local curvature change is used to identify waveform inflection points, the energy distribution density is used to measure the activity level of local respiratory activity, and the periodic repetition consistency is used to evaluate the stability between multiple consecutive respiratory cycles.

[0033] Then, based on the basic temporal characteristics, the periodic peaks, troughs, and amplitude change nodes of the modal signal can be identified. That is, the above-mentioned basic temporal characteristics are jointly analyzed, and local feature evolution curves are constructed on the continuous time axis. Periodic peaks are identified according to the position where the first-order change trend of amplitude changes from positive to negative, and periodic troughs are identified according to the position where the first-order change trend of amplitude changes from negative to positive. The amplitude change nodes are determined by combining the positions of local energy mutation, curvature extrema, and rapid amplitude changes. This generates a set of candidate respiratory events for the corresponding modality. Each candidate respiratory event records the event occurrence time, event duration, amplitude difference between peaks and troughs, local energy change, and corresponding period number.

[0034] Finally, since the candidate respiratory events still include some spurious events caused by motion impact, transient noise, or non-respiratory physiological activities, preset temporal and amplitude rules can be applied to the candidate respiratory event set. First, temporal rules are applied, establishing minimum and maximum time interval constraints between adjacent events based on the normal respiratory rhythm during human motion. When the time interval between two adjacent candidate events is significantly shorter than the normal human respiratory cycle, events with small amplitude changes or low local stability are identified as noise events and eliminated. When the interval between adjacent events is too large, the consistency of waveforms in adjacent cycles is further considered to determine if there are any missed events, and weak events within the local time window are re-searched through backtracking. Next, amplitude rules are applied, adaptively calculating the event amplitude threshold based on the respiratory modulation intensity corresponding to the current motion stage. When the peak-to-valley amplitude difference of a candidate event is lower than a certain proportion of the local average amplitude, or the event energy change is insufficient to characterize real respiratory activity, it is identified as an invalid event. For events that meet the amplitude requirements, a comprehensive evaluation is conducted by combining the periodic consistency and waveform continuity of multiple preceding and following respiratory cycles, retaining only candidate events that simultaneously meet temporal continuity, amplitude stability, and periodic consistency. To further improve the accuracy of identifying critical motor respiratory events, an event credibility index is calculated for the retained events. This credibility index comprehensively considers the stability of the event amplitude, the consistency of local periods, the frequency of recurrence across windows, and the consistency with changes in adjacent modal signals. Events with higher credibility are assigned greater weight, while events with lower credibility are weighted less, ultimately filtering out the critical motor respiratory events for the corresponding modality. Each critical motor respiratory event includes information such as event time, event type, event amplitude, event duration, event period number, and event credibility.

[0035] In some embodiments, time alignment processing of the temporal data of each modality based on the timing of key motor respiratory events of all modalities can be performed in the following manner: Acquire key motion-respiratory events and their corresponding event time identifiers obtained from each modality detection; Key motor respiratory events of different modalities are sorted by time according to the event time identifier; For cross-modal key motion-respiratory events that are within a preset time neighborhood and point to the same respiratory cycle stage, perform event merging processing to generate a unified event node; Organize the event node sequence according to the time order of unified event nodes to construct an event reference timeline; Using the event reference timeline as a reference, time alignment processing is performed on the timing data of each modality.

[0036] In practice, after obtaining the key motor respiratory events corresponding to each modality, the event time, event type, event duration, event amplitude, event confidence level, and the respiratory cycle number to which each key motor respiratory event belongs are used as basic information. First, an independent event time index sequence is established for each modality, where each key motor respiratory event is assigned a unique event time identifier to characterize the occurrence position of the event in the continuous time series data of the corresponding modality. Then, using the event time identifier as a unified sorting criterion, all key motor respiratory events of different modalities are jointly sorted according to time sequence to form a global event candidate sequence. The time interval, event type correspondence, and event evolution direction between any two cross-modal events are calculated. The event evolution direction is used to describe whether the signal change trend before and after the event is consistent, for example, both showing enhancement during the inspiratory phase, weakening during the expiratory phase, or stabilization during the recovery phase.

[0037] Furthermore, event merging processing can be performed on cross-modal critical motion respiratory events that are within a preset time neighborhood and point to the same respiratory cycle stage to generate a unified event node. That is, when the time interval between any two or more cross-modal critical motion respiratory events is within the preset time neighborhood, and the evolution direction of the corresponding critical motion respiratory events is consistent, and they all point to the same respiratory cycle stage according to their respective respiratory cycle numbers, then these critical motion respiratory events are considered to originate from the same real respiratory regulation process. Further, event merging processing is performed on cross-modal critical motion respiratory events that meet the above conditions. In the event merging process, a certain modal critical motion respiratory event is not simply selected as a representative. Instead, the event fusion weight is calculated based on the event credibility, amplitude stability, and duration evaluation value of each modal critical motion respiratory event, and the unified event occurrence time is determined according to a weighted average method.

[0038] in, Indicates the time when a unified event occurs. This represents the event time of the critical motor-respiratory event in the i-th mode. The event fusion weight represents the weight of the corresponding key motor respiratory event, where n represents the total number of modalities. The event fusion weight comprehensively reflects the event's credibility, local signal stability, and amplitude contribution. Simultaneously, attributes such as event type, duration, peak-to-valley amplitude variation, and local energy variation are synchronously fused to generate a unified event node that fully describes a respiratory regulation process. Each unified event node not only retains the fused event time but also records the number of participating modalities, the response order of each modality, and the event consistency evaluation index. The event consistency evaluation index measures the consistency of responses from different modalities to the same respiratory activity. When a key motor respiratory event fails to find other modal events satisfying the consistency constraints within a preset time neighborhood, the local search window is further expanded, and a secondary matching is performed based on the time interval between adjacent unified event nodes and the continuity of the respiratory cycle. If a correspondence still cannot be formed, the event is marked as an isolated event, and its credibility determines whether to retain it, thus avoiding erroneous nodes in the unified event sequence due to a single modal anomaly. After merging all cross-modal events, the event nodes are arranged sequentially according to their fusion time to construct a sequence covering the entire motion process. An event reference timeline is then established based on the time intervals between adjacent unified event nodes. This timeline no longer uses a fixed sampling time as a benchmark; instead, it uses key motor respiratory events during actual respiratory activity as time anchors, ensuring the timeline accurately reflects the dynamic evolution of human respiratory regulation. Finally, time alignment is performed on the temporal data of each modality based on the event reference timeline.

[0039] In the above embodiments, for each modality's critical motor respiratory event, firstly, the ratio of the absolute value of the difference between the peak-to-trough amplitude of the critical motor respiratory event and the average peak-to-trough amplitude within the local time window to the standard deviation of the local peak-to-trough amplitude can be used as the event confidence of the critical motor respiratory event. This event confidence represents the degree of credibility of the corresponding critical motor respiratory event. Then, the ratio of the standard deviation of the peak-to-trough amplitude of multiple respiratory cycles in the current event's neighborhood to the neighborhood's average peak-to-trough amplitude can be used as the amplitude stability of the critical motor respiratory event. This amplitude stability is used to evaluate the stability of the respiratory amplitude before and after the critical motor respiratory event. Secondly, the duration evaluation value of the critical motor respiratory event can be determined by the duration of the critical motor respiratory event and the local average event duration.

[0040] in, This represents the duration evaluation value of the critical motor respiratory event in the i-th mode. Let L represent the duration of the critical motor respiratory event in the i-th modality, and let L represent the local average event duration. This duration evaluation value is used to indicate how close the current duration of the critical motor respiratory event is to the normal duration. Finally, the event fusion weight of the critical motor respiratory event can be calculated based on the event credibility, event amplitude stability, and duration evaluation value. This involves normalizing the event credibility, event amplitude stability, and duration evaluation value, and then performing a weighted sum. The corresponding weights can be set based on historical experiments or data analysis, which will not be elaborated here. The calculation result can then be used as the event fusion weight of the critical motor respiratory event in the corresponding modality.

[0041] In some embodiments, the time alignment processing of the modal timing data based on the event reference timeline is specifically carried out in the following manner: Map the key motion and breathing events of each modality to the corresponding unified event nodes in the event reference timeline; Calculate the time offset of each modal key respiratory event relative to the unified event node; Based on the time offset, the time series data of the corresponding mode is time-shifted to generate time-synchronized multimodal time series data.

[0042] In specific implementation, after completing the construction of the unified event reference timeline, the event reference timeline is first used as the unified time benchmark for key motion and breathing events in each modality. Specifically, for a key motion and breathing event in any modality, the event time identifier, event type, event amplitude, and event credibility corresponding to the key motion and breathing event are obtained. The event time identifier is then matched with the time position of each unified event node in the event reference timeline. By calculating the time distance between the event time identifier and the time of each reference event node, the unified event node with the smallest time distance and the highest consistency of event type is selected as the mapping target.

[0043] Then, the time offset of each modal key motion-respiratory event relative to the corresponding unified event node can be calculated. That is, for each modal key motion-respiratory event, its actual occurrence time and the corresponding node time in the event reference time axis are obtained, and the time difference between the two is calculated as the time offset. When the time offset is greater than zero, it indicates that the modal event is lagging behind the reference event axis; when the time offset is less than zero, it indicates that the modal event occurs earlier than the reference event node.

[0044] Finally, time-shifting processing can be performed on the time-series data of the corresponding modality based on the time offset. This involves adjusting the time coordinates of the original time-series signal of the modality as a whole, bringing the positions of all key motion-respiratory events in the modality closer to the corresponding nodes on the event reference time axis. Specifically, for data points whose sampling time points completely correspond to the event reference time axis nodes, the time coordinates are directly corrected according to the time offset. For non-event sampling points located between two key motion-respiratory events, since the overall shift may cause differences between the sampling time interval and the reference time axis, an adaptive interpolation method based on adjacent event node constraints is used to reconstruct the data at the missing time positions. Specifically, the change ratio of intermediate sampling points is calculated based on the time interval between two adjacent unified event nodes and the corresponding modality signal change trend, ensuring that the time-adjusted signal still maintains the continuity of the original waveform and the local variation patterns, avoiding respiratory waveform distortion caused by simple time shifting. Meanwhile, for modal data with different sampling frequencies, resampling is performed according to the time resolution of the event reference time axis to convert each modal data to a unified time granularity, so that respiratory-related information collected by different sensors can be expressed on the same time scale, that is, to generate multimodal time-series data after time synchronization.

[0045] It should be noted that by detecting key motion-respiratory events from each modal time-series data and performing time alignment processing on the data of each modality based on the temporal relationship of key motion-respiratory events of all modalities, asynchronous physiological data that was originally caused by differences in sensor sampling frequency, delay in physiological response transmission, and different signal change mechanisms can be converted into correlated data with a unified time reference. This allows information reflecting the same motion-respiratory regulation process in different modal signals such as photoplethysmography pulse wave, blood oxygen saturation, motion inertia, and skin temperature to be accurately matched at the same event node.

[0046] In step S103, a spatiotemporal heterogeneous graph is constructed using the feature vectors of each modality at different event nodes after time alignment as nodes and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights.

[0047] Preferably, in some embodiments, the spatiotemporal heterogeneous graph is constructed using the feature vectors of each modality at different event nodes after time alignment as nodes, and the fuzzy correlation degree calculated based on event time difference, feature similarity, and event importance as edge weights. Specifically, this can be achieved by: In time-aligned multimodal time series data, respiratory event feature vectors of the corresponding modality are extracted at each event node, and the respiratory event feature vectors are used as nodes of a heterogeneous graph for modeling. Based on the event reference timeline, the time interval parameters between event nodes are determined to characterize the temporal adjacency relationship between nodes; Similarity analysis is performed on the respiratory event feature vectors of different modal event nodes to obtain feature similarity parameters used to characterize the degree of cross-modal correlation. The feature similarity parameters are parameters used to characterize the degree of correlation between different modal event nodes at the feature level. Configure event importance parameters for event nodes based on their structural position in the event sequence; The time interval parameter, feature similarity parameter, and event importance parameter are input into the fuzzy inference model, and the fuzzy correlation degree between nodes is generated according to the preset fuzzy rules. Using the fuzzy correlation degree as the edge weight, a connection relationship is established between event nodes to obtain a spatiotemporal heterogeneous graph that integrates temporal correlation and cross-modal correlation.

[0048] In practical implementation, firstly, in time-aligned multimodal time-series data, respiratory event feature vectors corresponding to each modality can be extracted at each event node. That is, based on the corresponding position of each event node in each modal time-series data, a local signal window within a preset time range before and after the event node is extracted, and corresponding respiratory event feature vectors are extracted according to the physiological characteristics of different modal signals. For the photoplethysmography (PPG) mode, features reflecting changes in cardiopulmonary coupling are extracted, including changes in peak pulse wave value, trends in heart interval changes, respiratory modulation amplitude, and local waveform energy distribution. For the blood oxygen saturation mode, features reflecting the oxygen exchange process are extracted, including the rate of change of blood oxygen, the magnitude of blood oxygen decrease, recovery time, and the stability of oxygen supply. For the triaxial motion inertial mode, features reflecting the effects of motion induction are extracted, including motion intensity, motion cycle, rate of change of acceleration, and motion state switching characteristics. For the skin temperature mode, features reflecting the metabolic regulation process are extracted, including temperature change gradient, thermal recovery rate, and the degree of local temperature fluctuation. By using the above method, continuous time-series signals of different modalities are converted into event-level feature vectors with clear physiological meanings. The respiratory event feature vector of the corresponding modality at each event node is used as node information in the spatiotemporal heterogeneous graph for modeling, so that the graph structure can simultaneously retain the independent expressive ability of different physiological modalities and the correlation between events.

[0049] Then, the time interval parameter between event nodes can be determined based on the event reference timeline. That is, for any two event nodes, the time interval parameter between the event nodes is calculated according to their time positions in the event reference timeline, which is the absolute value of the time difference between the time positions of the two event nodes in the event reference timeline.

[0050] Secondly, similarity analysis can be performed on the respiratory event feature vectors of different modal event nodes. That is, for any two different modal event nodes, their corresponding respiratory event feature vectors are mapped to a feature space, so that physiological features with different dimensions and different physical meanings are transformed into a unified feature space. Then, the feature similarity between the two event nodes is calculated. A similarity calculation method based on feature distance can be used to calculate the feature similarity parameters used to characterize the degree of cross-modal association. The feature similarity parameters are used to characterize the degree of association between different modal event nodes at the feature level. Event importance parameters can be configured for event nodes based on their structural position in the event sequence. These parameters represent the relative weight of the event node within the respiratory cycle structure. Specifically, the local change magnitude, duration, reliability, and impact on the preceding and following respiratory state changes of the event node can be analyzed to construct an event importance evaluation index. For example, the importance parameter can be increased for event nodes that cause simultaneous changes in multiple modalities, are located in the motion state transition phase, or exhibit significant respiratory regulation characteristics; conversely, the importance parameter can be decreased for event nodes with small change magnitudes, existing only in a single modality, or exhibiting low stability. This approach allows key motion-respiratory events in the graph structure to receive higher connection weights, thereby highlighting information that significantly contributes to respiratory state prediction.

[0051] Furthermore, time interval parameters, feature similarity parameters, and event importance parameters can be input into the fuzzy inference model to generate fuzzy correlation degrees between nodes based on preset fuzzy rules. Since the physiological response during human movement and respiration has continuous changes and uncertainties, the correlation between events is usually difficult to accurately describe through fixed thresholds. Therefore, the fuzzy inference model is used to comprehensively evaluate the correlation between multiple factors. Specifically, the time interval parameter, feature similarity parameter, and event importance parameter are first fuzzified and converted into fuzzy variables of different levels (low, medium, and high). Then, inference is performed according to preset fuzzy rules. For example, when two event nodes are close in time, have high feature similarity, and high event importance, they are determined to have a strong correlation. When they are far apart in time, have significant feature differences, and low event importance, they are determined to have a weak correlation. After fuzzy inference, the fuzzy correlation degree between nodes is output. This fuzzy correlation degree is a weight value used to comprehensively characterize the correlation strength of event nodes at the time, feature, and structural levels. It reflects the complex, nonlinear, and uncertain correlation between cross-modal events. Thus, the fuzzy inference mechanism achieves dynamic adaptive coupling of multi-dimensional parameters, breaking through the limitations of traditional fixed weights or single metrics in characterizing the inherent semantic correlation of asynchronous heterogeneous breathing events.

[0052] Finally, fuzzy correlation degree can be used as edge weight to establish connection relationships between event nodes, resulting in a spatiotemporal heterogeneous graph that integrates temporal correlation and cross-modal correlation. This spatiotemporal heterogeneous graph is a graph model used to describe the correlation relationship between multimodal breathing events in the time dimension and feature dimension.

[0053] It should be noted that by using the feature vectors of each modality at different event nodes after time alignment as nodes, and using the fuzzy correlation degree calculated by combining event time difference, feature similarity and event importance as edge weights to construct a spatiotemporal heterogeneous graph, it is possible to transform the originally independent multimodal physiological data during the motion process into a unified graph structure with temporal evolution relationship and cross-modal correlation relationship, thereby breaking through the limitations of traditional methods that rely solely on time sequence or simple feature splicing for data fusion.

[0054] In step S104, cross-modal feature fusion is performed on the respiratory event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fusion features characterizing the respiratory state. Based on the multimodal fusion features, the respiratory state during the movement process is predicted.

[0055] In some embodiments, cross-modal feature fusion of respiratory event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fused features characterizing the respiratory state can be performed in the following manner: The adjacency relationships of each node in the spatiotemporal heterogeneous graph are traversed to obtain the set of cross-modal nodes associated with the target node; Based on the cross-modal node set, the node features are weighted and aggregated to generate an event-level fusion feature representation; All event-level fusion features are integrated across event nodes to form multimodal fusion features that characterize respiratory states.

[0056] In specific implementation, firstly, the adjacency relationships of each node in the spatiotemporal heterogeneous graph can be traversed to obtain the set of cross-modal nodes associated with the target node. That is, taking the target node as the center, based on the edge connections already established in the spatiotemporal heterogeneous graph, all its adjacent nodes are traversed and retrieved to obtain the set of nodes that are associated with the target node. These adjacent nodes include both same-modal event nodes located before and after the target node on the event reference timeline and having a temporally continuous relationship with the target node, and heteromodal event nodes connected through fuzzy correlation degrees that can reflect different physiological responses of the same respiratory regulation process. Subsequently, the fuzzy correlation degrees between the target node and each adjacent node are obtained, and the relationships are then established based on the fuzzy correlation degrees. The association strength screening involves calculating the statistical mean of all edge weights and using this statistical mean as the basic association threshold. Only adjacent nodes with a fuzzy association degree greater than the basic association threshold are retained to avoid weakly associated nodes introducing redundant information. Furthermore, to take into account the overall topological characteristics of the graph structure, the topological importance of each adjacent node in the entire spatiotemporal heterogeneous graph is also calculated. If an adjacent node has a fuzzy association degree slightly lower than the basic association threshold, but it maintains stable connections with multiple event nodes and can undertake the information transmission between different modalities, then the node is retained to participate in subsequent fusion calculations. This avoids the accidental deletion of key bridging nodes due to relying solely on the association degree threshold, ultimately forming a cross-modal node set that takes into account both local association strength and global topological structure.

[0057] Then, weighted aggregation of node features can be performed based on the cross-modal node set. This involves obtaining the respiratory event feature vector corresponding to each event node in the cross-modal node set, and simultaneously reading the fuzzy correlation degree between the target node and the corresponding event node. This fuzzy correlation degree is used as the node contribution weight, and the weights of each respiratory event feature vector are modulated so that nodes with higher correlation degrees contribute a larger proportion during the aggregation process, while nodes with lower correlation degrees only provide auxiliary information. This fully reflects the differences in the impact of different nodes on the respiratory state of the target node. Subsequently, a vector-level merging operation is performed on the multiple respiratory event feature vectors after weight modulation. This vector-level merging operation can be performed by weighted superposition or dimension-by-dimensional fusion, uniformly calculating the dimensions corresponding to features from different modalities. This allows respiratory event features from different sources to form a unified comprehensive feature representation while maintaining their respective physiological meanings, resulting in an event-level fused feature representation. Since the fuzzy correlation strength between nodes has been comprehensively considered during the aggregation process, the generated event-level fusion feature representation not only retains the local respiratory information contained in the target node itself, but also integrates a variety of physiological information such as motion induction, cardiopulmonary circulation response, oxygen exchange changes and thermoregulation changes, so as to achieve information complementarity between different modalities.

[0058] Finally, all event-level fusion features can be integrated across event nodes to form multimodal fusion features representing respiratory states. This involves summarizing all event-level fusion features based on their temporal distribution throughout the respiratory event sequence, and analyzing the continuity of respiratory state changes, the stability of respiratory rhythm, and the trend of exercise load changes between adjacent events. Subsequently, each event-level fusion feature is integrated according to its contribution to the entire respiratory regulation process, ensuring that different respiratory states, such as the start of exercise, the stable exercise phase, and the recovery phase, are reasonably represented in the final feature representation, while fully preserving the dynamic evolutionary relationships between each phase. After overall integration and computation, a multimodal fusion feature is finally generated that comprehensively reflects multi-source physiological information such as exercise-induced response, cardiopulmonary circulation regulation, oxygen metabolism capacity, and thermoregulation state throughout the entire exercise monitoring cycle. This multimodal fusion feature not only retains the local features of a single event node but also integrates the temporal dependencies and cross-modal correlations between different events, providing a high-quality feature representation with global spatiotemporal correlation characteristics for subsequent respiratory state prediction during exercise, thereby improving the accuracy, continuity, and stability of respiratory state prediction results.

[0059] In some embodiments, predicting the respiratory state during movement based on the multimodal fusion features can be achieved in the following ways: Perform feature scale normalization processing on the multimodal fusion features; The normalized multimodal fusion features are input into a pre-defined prediction network; Based on the prediction network, the multimodal fusion features are processed to perform pattern discrimination, generating corresponding breathing state determination information, and then outputting the breathing state during the movement process.

[0060] In practical implementation, firstly, to avoid the impact of differences in dimensions between different feature dimensions and differences in individual physiological baselines on the prediction results, feature scale normalization processing can be performed on the multimodal fusion features. This involves using the multimodal fusion feature sequence generated during continuous monitoring as input, statistically analyzing the distribution range, average level, and dispersion of each feature dimension within the current movement phase, and combining this with the individual physiological baseline established under the user's resting state to perform adaptive scaling transformation on each feature dimension, mapping fusion features from different sources to a unified feature space. Specifically, for features with large fluctuations, their numerical variation range is compressed to reduce the impact of outliers; for features with small fluctuations but sensitive to respiratory status, their local variation trend is preserved to enhance their identification ability in subsequent prediction processes. Simultaneously, the gradient of change between adjacent event-level fusion features is further analyzed. Abnormal features that change abruptly in a short period but are not persistent are smoothed and corrected, while changes that continuously reflect the evolution trend of respiratory status are retained. This ensures that the normalized multimodal fusion features have a unified data scale while fully preserving the dynamic changes in respiratory regulation during movement, thereby generating a normalized fusion feature sequence suitable for prediction network input.

[0061] Then, the normalized multimodal fusion features can be input into a pre-defined prediction network to perform in-depth modeling of the temporal evolution of respiratory state during exercise. The prediction network employs a dual-path prediction structure that integrates temporal modeling and state correlation modeling. The first path extracts the temporal dependency of respiratory state changes between consecutive event nodes, while the second path analyzes the synergistic changes between different physiological characteristics. Specifically, the normalized multimodal fusion features are first input into the prediction network sequentially according to the event reference timeline, enabling the network to learn the continuous changes in respiratory state over time during the start, stable, and recovery phases of exercise. Subsequently, a hierarchical correlation is established within the network between short-term respiratory change features and long-term respiratory regulation trends. This allows the network to identify both instantaneous increases in respiratory rate and fluctuations in blood oxygenation during exercise, and to comprehensively analyze information from multiple historical event nodes to determine whether the respiratory state shows a trend of continuous deterioration or gradual recovery. Meanwhile, to further enhance the predictive network's adaptability to complex motion scenarios, an adaptive feature selection mechanism is introduced within the network. Based on the changes in the contribution of each feature dimension to the respiratory state prediction at different motion stages, the participation ratio of different fusion features is dynamically adjusted. This allows cardiopulmonary circulation-related features, oxygen exchange-related features, exercise load-related features, and thermoregulation-related features to automatically participate in the prediction calculation according to the current respiratory state, thereby improving the predictive model's adaptability to different exercise intensities, different respiratory rhythms, and different individual physiological differences.

[0062] Furthermore, pattern discrimination processing can be performed on the normalized multimodal fusion features based on the prediction network. First, the prediction network extracts a high-dimensional state representation corresponding to the current respiratory state based on the input features, and calculates the matching degree between this state representation and several pre-established respiratory state patterns, including normal breathing pattern, exercise-enhanced breathing pattern, respiratory recovery pattern, and abnormal breathing pattern. Then, based on the state results output by multiple consecutive event nodes, consistency constraint analysis is performed on the prediction results at adjacent time points. When multiple consecutive event nodes point to the same respiratory state pattern, the reliability of the pattern's judgment is improved. If there are significant jumps in the prediction results of adjacent nodes, dynamic correction is performed by combining the evolution trend of the respiratory state of the preceding and following event nodes with the current change in exercise load. This reduces the impact of random noise and single abnormal events on the prediction results, ensuring that the final output respiratory state has good temporal continuity and stability. Furthermore, to further improve the reliability of the prediction results, the confidence level of the current breathing state can be calculated based on the judgment probabilities corresponding to various breathing states output by the prediction network. When the highest judgment probability exceeds a preset confidence threshold, the corresponding breathing state is directly output. If the judgment probabilities of multiple states are close, a joint judgment is made by combining the previous breathing state and the current movement trend to avoid frequent switching of breathing states. Finally, corresponding breathing state judgment information is generated based on the pattern discrimination results, and the breathing state during the movement is output. The breathing state includes at least the current respiratory rate level, the stability of the respiratory rhythm, the respiratory load level, and the respiratory abnormality risk level, and can further predict the development trend of the breathing state over a period of time.

[0063] It should be noted that, based on spatiotemporal heterogeneous graphs, cross-modal feature fusion of respiratory event features from various modalities is performed, and the fused multimodal features are used to predict respiratory status during exercise. This approach can fully utilize the complementary information between different physiological modalities in the temporal and physiological functional dimensions. It maps respiratory-related information, which was originally scattered across multiple sensors such as photoplethysmography pulse wave, blood oxygen saturation, triaxial motion inertia, and skin temperature, into the same spatiotemporal correlation structure for collaborative analysis. This avoids the information fragmentation and missing correlation problems caused by traditional methods that use single physiological signals or simple feature splicing. Furthermore, by relying on spatiotemporal heterogeneous graphs to achieve accurate time alignment and deep cross-modal fusion of respiratory events from different modalities, it can improve the completeness, consistency, and discriminative ability of respiratory status representation during exercise, ultimately improving the accuracy and stability of respiratory status prediction.

[0064] Therefore, this application firstly detects key motion-respiratory events from the temporal data of each modality, and performs time alignment processing on the data of each modality based on the temporal relationship of key motion-respiratory events of all modalities. This transforms the asynchronous physiological data, which was originally caused by differences in sensor sampling frequency, delays in physiological response transmission, and different signal change mechanisms, into correlated data with a unified time reference. Then, using the feature vectors of each modality at different event nodes as nodes, a spatiotemporal heterogeneous graph is constructed. This transforms the originally independent multimodal physiological data during exercise into a unified graph structure with temporal evolution relationships and cross-modal correlation relationships, thus breaking through the limitations of traditional methods that rely solely on time sequence. This method overcomes the limitations of data fusion through sequential or simple feature splicing. Finally, it performs cross-modal feature fusion on respiratory event features of each modality based on spatiotemporal heterogeneous graphs, and uses the fused multimodal fusion features to predict respiratory state during exercise. This fully utilizes the complementary information between different physiological modalities in the temporal and physiological functional dimensions, avoiding the information fragmentation and missing correlation problems caused by traditional methods that use single physiological signals or simple feature splicing. Furthermore, it achieves accurate time alignment and deep cross-modal fusion of respiratory events of different modalities based on spatiotemporal heterogeneous graphs, which can improve the completeness, consistency and discriminative ability of respiratory state representation during exercise, and ultimately improve the accuracy and stability of respiratory state prediction.

[0065] In summary, the technical solution adopted in this application can construct a unified key motion respiratory event association mechanism for multimodal physiological time series data, realize accurate time alignment and cross-modal deep fusion of respiratory events of different modalities, and improve the accuracy and stability of respiratory state prediction during exercise.

[0066] In another aspect, in some embodiments, this application provides a sports and health management system, including a sports breathing prediction unit, with reference to... Figure 4 The figure is a schematic diagram of the structure of a motion-breathing prediction unit according to some embodiments of this application. The motion-breathing prediction unit includes: The data acquisition module 201 is used to collect multimodal time-series data related to respiratory function during exercise through smart sensors deployed inside the wearable device; The time alignment module 202 is used to detect key motion and breathing events from each modal time series data, and perform time alignment processing on each modal time series data based on the time series of key motion and breathing events of all modalities. The heterogeneous graph construction module 203 is used to construct a spatiotemporal heterogeneous graph by using the feature vectors of each modality at different event nodes after time alignment as nodes and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights. The breathing prediction module 204 is used to perform cross-modal feature fusion on the breathing event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fusion features characterizing the breathing state, and predict the breathing state during the movement process based on the multimodal fusion features.

[0067] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described intelligent prediction method for motor respiration.

[0068] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing an intelligent prediction method for motion-induced respiration according to some embodiments of this application. The intelligent prediction method for motion-induced respiration in the above embodiments can... Figure 4 The computer device shown is used to implement this, and the computer device includes at least one processor 301, a communication bus 302, a memory 303, and at least one communication interface 304.

[0069] The processor 301 can be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of the intelligent prediction method for motor breathing in this application.

[0070] The communication bus 302 can be used to transmit information between the aforementioned components.

[0071] The memory 303 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 303 may exist independently and be connected to the processor 301 via the communication bus 302. The memory 303 may also be integrated with the processor 301.

[0072] The memory 303 stores program code for executing the scheme of this application, and its execution is controlled by the processor 301. The processor 301 executes the program code stored in the memory 303. The program code may include one or more software modules. In the above embodiments, the determination of the intelligent prediction method for motor respiration can be achieved by the processor 301 and one or more software modules in the program code in the memory 303.

[0073] Communication interface 304 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0074] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0075] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0076] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent prediction method for motion and breathing.

[0077] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0078] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for intelligent prediction of respiration during exercise, characterized in that, Includes the following steps: By using smart sensors deployed inside wearable devices, multimodal time-series data related to respiratory function can be collected during exercise. Key motor and respiratory events are detected from the time series data of each modality, and time alignment is performed on the time series data of each modality based on the time series of key motor and respiratory events of all modalities. Using the feature vectors of each modality at different event nodes after time alignment as nodes, and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights, a spatiotemporal heterogeneous graph is constructed. Based on the spatiotemporal heterogeneous graph, cross-modal feature fusion is performed on the respiratory event features of each modality to obtain multimodal fusion features characterizing the respiratory state. Based on the multimodal fusion features, the respiratory state during the movement process is predicted.

2. The intelligent prediction method for exercise-induced breathing as described in claim 1, characterized in that, The multimodal time-series data includes photoplethysmography pulse wave time-series data, blood oxygen saturation time-series data, triaxial motion inertial time-series data, and skin temperature time-series data.

3. The intelligent prediction method for exercise-induced breathing as described in claim 1, characterized in that, The key motor-respiratory events detected from the time-series data of each modality specifically include: For each mode, denoising and stabilization processing is performed on the modal time series data to eliminate high-frequency interference and suppress baseline drift, resulting in a continuous modal signal sequence. Extract fundamental temporal features from the modal signal sequence to characterize the respiratory cycle structure; Based on the aforementioned basic temporal characteristics, periodic peaks, troughs, and amplitude variation nodes of the modal signal are identified to generate a candidate respiratory event set; Preset timing and amplitude rules are applied to the candidate respiratory event set to filter out key motion respiratory events of the corresponding modality.

4. The intelligent prediction method for exercise-induced breathing as described in claim 1, characterized in that, Based on the timing of key motor and respiratory events across all modalities, the time alignment processing of the temporal data for each modality specifically includes: Acquire key motion-respiratory events and their corresponding event time identifiers obtained from each modality detection; Key motor respiratory events of different modalities are sorted by time according to the event time identifier; For cross-modal key motion-respiratory events that are within a preset time neighborhood and point to the same respiratory cycle stage, perform event merging processing to generate a unified event node; Organize the event node sequence according to the time order of unified event nodes to construct an event reference timeline; Using the event reference timeline as a reference, time alignment processing is performed on the timing data of each modality.

5. The intelligent prediction method for motor respiration as described in claim 4, characterized in that, Using the event reference timeline as a benchmark, the time alignment processing of the time series data for each modality specifically includes: Map the key motion and breathing events of each modality to the corresponding unified event nodes in the event reference timeline; Calculate the time offset of each modal key motion breathing event relative to the unified event node; Based on the time offset, the time series data of the corresponding mode is time-shifted to generate time-synchronized multimodal time series data.

6. The intelligent prediction method for motor respiration as described in claim 1, characterized in that, Using the feature vectors of each modality at different event nodes after time alignment as nodes, and the fuzzy correlation degree calculated based on event time difference, feature similarity, and event importance as edge weights, a spatiotemporal heterogeneous graph is constructed, specifically including: In time-aligned multimodal time series data, respiratory event feature vectors of the corresponding modality are extracted at each event node, and the respiratory event feature vectors are used as heterogeneous graph nodes for modeling. Based on the event reference timeline, the time interval parameters between event nodes are determined to characterize the temporal adjacency relationship between nodes; Similarity analysis is performed on the respiratory event feature vectors of different modal event nodes to obtain feature similarity parameters used to characterize the degree of cross-modal correlation. The feature similarity parameters are parameters used to characterize the degree of correlation between different modal event nodes at the feature level. Configure event importance parameters for event nodes based on their structural position in the event sequence; The time interval parameter, feature similarity parameter, and event importance parameter are input into the fuzzy inference model, and the fuzzy correlation degree between nodes is generated according to the preset fuzzy rules. Using the fuzzy correlation degree as the edge weight, a connection relationship is established between event nodes to obtain a spatiotemporal heterogeneous graph that integrates temporal correlation and cross-modal correlation.

7. The intelligent prediction method for exercise-induced breathing as described in claim 1, characterized in that, Based on the multimodal fusion features, the prediction of respiratory state during motion specifically includes: Perform feature scale normalization processing on the multimodal fusion features; The normalized multimodal fusion features are input into a pre-defined prediction network; Based on the prediction network, the multimodal fusion features are processed to perform pattern discrimination, generating corresponding breathing state determination information, and then outputting the breathing state during the movement process.

8. A sports and health management system, comprising a sports breathing prediction unit, for executing a sports breathing intelligent prediction method as described in any one of claims 1 to 7, characterized in that, The exercise-breathing prediction unit includes: The data acquisition module is used to collect multimodal time-series data related to respiratory function during exercise through smart sensors deployed inside the wearable device; The time alignment module is used to detect key motion and respiratory events from the time series data of each modality, and to perform time alignment processing on the time series data of each modality based on the time series of key motion and respiratory events of all modalities. The heterogeneous graph construction module is used to construct a spatiotemporal heterogeneous graph by using the feature vectors of each modality at different event nodes after time alignment as nodes and the fuzzy correlation degree calculated based on event time difference, feature similarity and event importance as edge weights. The breathing prediction module is used to perform cross-modal feature fusion on the breathing event features of each modality based on the spatiotemporal heterogeneous graph to obtain multimodal fusion features characterizing the breathing state, and predict the breathing state during the movement process based on the multimodal fusion features.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute a motor breathing intelligent prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a method for intelligent prediction of respiration as described in any one of claims 1 to 7.

Images