Brain disease auxiliary evaluation method and system fusing artificial intelligence and holographic physiological function

Abstract

The application discloses a brain disease auxiliary evaluation method and system fusing artificial intelligence and holographic physiological functions, and the method comprises the following steps: acquiring multi-night smart watch data and dividing the multi-night smart watch data into night segments, generating a sleep stage probability vector and a segment quality weight; determining a suspected REM segment, a suspected wake-up segment and a relatively stable sleep segment according to the probability vector, and generating a personal physiological reference value for the night; detecting a night movement event, extracting movement characteristics and autonomic nervous response characteristics, forming a three-segment representation of the movement event, simultaneously generating a wake-up action contrast representation and a reference deviation; inputting the above information into a quality perception multi-task time sequence convolution network to obtain a related abnormality index and a segment credibility score; and outputting an auxiliary evaluation result after similar segment review within the night and multi-night quality screening and summarization.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence-based brain disease auxiliary assessment technology, specifically to a brain disease auxiliary assessment method and system that integrates artificial intelligence and holographic physiological functions. Background Technology

[0002] Rapid eye movement (REM) sleep behavior disorder is associated with early risk assessment of neurodegenerative diseases such as Parkinson's disease and Lewy body dementia. Current clinical assessments typically rely on polysomnography (PSG), video recording, and manual interpretation. While these methods offer high accuracy, they are costly, require sophisticated deployment conditions, and struggle to cover long-term, continuous home sleep scenarios. With the development of wearable devices such as smartwatches, data on triaxial acceleration, angular velocity, photoplethysmography (PPG), heart rate, heart rate variability, and blood oxygen saturation can now be continuously collected at night, providing new data sources for auxiliary assessment of brain disease risks.

[0003] However, existing sleep or motion analysis methods based on wearable devices mostly rely on single sleep stage tags, fixed thresholds, or single-night statistical results, making it difficult to adequately address issues such as sleep staging errors in consumer-grade devices, loosening of the device, motion artifacts, individual differences in physiological baselines, and interference from occasional nighttime movements. Especially for suspected REM (Rapid Eye Movement) abnormal movements, judging solely based on the amplitude or frequency of movements can easily misclassify ordinary actions such as turning over, getting up at night, or adjusting the wristband as abnormal, and it is difficult to characterize the temporal relationship between heart rate, heart rate variability, and blood oxygenation changes before and after the motion event. Summary of the Invention

[0004] This application provides a method and system for assisting in the assessment of brain diseases that integrates artificial intelligence and holographic physiological functions, in order to at least solve some of the technical problems existing in the related technologies described above.

[0005] According to a first aspect of the embodiments of this application, a method for assisting in the assessment of brain diseases that integrates artificial intelligence and holographic physiological functions is provided, comprising: Acquire multi-night smartwatch data, divide each night's data into nighttime segments, and generate a sleep stage probability vector and segment quality weight for each nighttime segment; Based on the probability vector of sleep stages, suspected REM segments, suspected awake segments, and relatively stable sleep segments are identified, and the individual physiological reference value for the night is generated from the relatively stable sleep segments. Motion events are detected from nightly data, multidimensional features and autonomic nervous system response features of motion events are extracted, and a three-segment representation of motion events is generated by combining the individual's physiological reference value for the night. A control representation of conscious actions was generated from motor events within suspected conscious segments, and a reference bias was generated from nighttime segment indicators and individual physiological reference values ​​for the night. The sleep stage probability vector, the three-segment representation of motor events or the zero vector when there are no motor events, the wakeful action control representation, the reference bias and the segment quality weight are concatenated to form the night segment feature vector, and the night segment sequence is formed according to the time sequence. Nighttime segment sequences are input into a quality-aware multi-task temporal convolutional network to obtain a suspected REM motion abnormality index, a motion-autonomic nervous system inconsistency index, and segment credibility scores. The credibility score of the segment is verified by comparing it with similar segments within the night to generate the credibility score of the verified segment. The multi-night data that meets the quality screening criteria are then summarized to output the auxiliary evaluation results.

[0006] As an optional approach, generating the sleep stage probability vector includes: when the smartwatch outputs the confidence score or probability value of each sleep stage, normalizing the confidence score or probability value of each sleep stage to obtain the sleep stage probability vector; when the smartwatch only outputs a single stage label, inputting the root mean square amplitude of acceleration, mean heart rate, and root mean square of continuous difference of the night segment into a pre-trained lightweight classifier to obtain the sleep stage probability vector.

[0007] As an optional approach, identifying suspected REM segments, suspected awake segments, and relatively stable sleep segments includes: recording nighttime segments where the REM component exceeds a preset threshold and is higher than the awake component, light sleep component, and deep sleep component as suspected REM segments; recording nighttime segments where the awake component exceeds a preset threshold and is higher than the REM component, light sleep component, and deep sleep component as suspected awake segments; and selecting nighttime segments from those nighttime segments where the light sleep component or deep sleep component is higher than the other components, where the acceleration activity is in the lowest quartile range of the night, and recording them as relatively stable sleep segments.

[0008] As an optional approach, the individual physiological reference values ​​for the night include resting heart rate reference, heart rate variability reference, blood oxygenation reference, and wrist micro-motion reference. The resting heart rate reference is the median of the heart rate in a relatively stable sleep segment. The heart rate variability reference is the median of the root mean square difference of continuous differences and the median of the standard deviation of adjacent normal heartbeat intervals in a relatively stable sleep segment. The blood oxygenation reference is the median of blood oxygen saturation in a relatively stable sleep segment. The wrist micro-motion reference is the median of the root mean square amplitude of acceleration and the interquartile range in a relatively stable sleep segment.

[0009] As an optional approach, motion event detection includes: calculating the composite amplitude of the triaxial acceleration signal; marking a motion event when the composite amplitude exceeds the resting noise threshold and the duration exceeds the shortest event duration threshold; determining the end of the motion event when the composite amplitude falls below the resting noise threshold and remains quiet for more than a preset interval; and merging short-term activities with adjacent intervals less than the preset interval into the same motion event.

[0010] As an optional approach, the multidimensional features of motion events include event duration, peak value of composite acceleration amplitude, mean value of composite acceleration amplitude, peak value of composite angular velocity amplitude, number of triaxial acceleration direction changes, duration of stillness before the start of the event, duration of stillness after the end of the event, and changes in wrist posture before and after the event; the autonomic nervous system response features include heart rate rise rate, heart rate recovery rate, root mean square change in continuous difference, change in standard deviation of adjacent normal heartbeat interval, short-term changes in blood oxygen saturation, and changes in PPG signal quality.

[0011] As an optional approach, generating a three-segment representation of a motion event includes: defining the time range before the start of the motion event as the first segment, the duration of the motion event as the event segment, and the time range after the end of the motion event as the second segment; the first segment includes the mean of the composite acceleration amplitude, the standard deviation of the composite acceleration amplitude, the mean heart rate, the root mean square of the continuous difference, the standard deviation of the interval between adjacent normal heartbeats, the mean blood oxygen saturation, and the deviation of each indicator relative to the individual's physiological reference value for the night; the event segment includes the multidimensional features of the motion event, the acceleration sequence at each sampling point, and the angular velocity sequence at each sampling point; the second segment includes the heart rate sequence, the root mean square curve of the continuous difference, the blood oxygen saturation curve, and the heart rate recovery time; and the probability vector of the sleep stage is appended to the first segment, the event segment, and the second segment to obtain the three-segment representation of the motion event.

[0012] As an optional approach, generating a conscious action control representation includes: performing multidimensional feature extraction and three-segment representation generation on motion events within suspected conscious segments to obtain a conscious action sample set; when the number of events in the conscious action sample set reaches a preset number, the three-segment representation of motion events in the conscious action sample set is used as the conscious action control representation; when the number of events in the conscious action sample set does not reach the preset number, the duration, number of direction changes, peak acceleration, peak angular velocity, post-action heart rate rise, post-action heart rate recovery time, and post-action root mean square variation of continuous difference are statistically summarized to obtain the conscious action control representation.

[0013] As an alternative approach, the quality-aware multi-task temporal convolutional network includes a dilated causal temporal convolutional backbone, a quality-aware feature aggregation module, a first output head, a second output head, and a third output head. The dilated causal temporal convolutional backbone consists of multiple stacked residual blocks. Each residual block includes two layers of one-dimensional causal dilated convolutions. After each layer of one-dimensional causal dilated convolutions, a weight normalization layer, a modified linear activation function, and a random dropout layer are connected in sequence. The residual blocks output hidden feature sequences through skip connections.

[0014] According to a second aspect of the embodiments of this application, a brain disease auxiliary assessment system integrating artificial intelligence and holographic physiological functions is also provided, comprising: The data acquisition and segment processing module is used to acquire multi-night smartwatch data, divide the data of each night into night segments, and generate a sleep stage probability vector and segment quality weight for each night segment. The segment classification and reference generation module is used to identify suspected REM segments, suspected awake segments, and relatively stable sleep segments based on the probability vector of sleep stages, and to generate the personal physiological reference value for the night from the relatively stable sleep segments. The motion event processing module is used to detect motion events from nightly data, extract multidimensional features and autonomic nervous system response features of motion events, and generate a three-segment representation of motion events by combining the individual's physiological reference values ​​for the night. The control and bias generation module is used to generate a control representation of conscious actions from motor events within suspected conscious segments, and to generate a reference bias from nighttime segment indicators and the individual's physiological reference values ​​for the night. The feature sequence generation module is used to concatenate the sleep stage probability vector, the three-segment representation of motion events or the zero vector when there are no motion events, the wakeful action comparison representation, the reference bias and the segment quality weight into a night segment feature vector, and form a night segment sequence in chronological order. The model inference module is used to input the nighttime segment sequence into the quality-aware multi-task temporal convolutional network to obtain the suspected REM motion abnormality index, the inconsistency index between motion and autonomic nervous system response, and the segment credibility score. The results output module is used to generate a verified segment credibility score by reviewing the segment credibility score with similar segments within the night, and to summarize multi-night data that meet the quality screening criteria, and output auxiliary evaluation results.

[0015] This application, based on multi-night smartwatch data, transforms sleep stages from deterministic labels to probabilistic representations and incorporates segment quality weights into subsequent calculations. This mitigates the impact of sleep staging errors and signal quality fluctuations in consumer-grade devices on assessment results. Furthermore, the system utilizes relatively stable sleep segments from the same wearer throughout the night to form a personal physiological reference, and uses motion events within suspected awakening segments as action comparisons. This ensures that anomaly detection no longer relies solely on fixed thresholds or group averages, but rather more closely reflects the individual's nighttime state. For nighttime motion events, the segmented representation preserves the relationships between heart rate, heart rate variability, blood oxygenation, and wrist movement changes before and after the movement, aiding in identifying mismatches between movement patterns and autonomic nervous system responses. Simultaneously, by combining a quality-perceived temporal model, segment credibility verification, and a multi-night aggregation mechanism, false alarms caused by occasional movements, wearing abnormalities, and low-quality segments can be reduced, improving the continuity and stability of brain disease auxiliary assessment in home settings.

[0016] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Furthermore, no embodiment in this disclosure is required to achieve all the effects described above. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0018] Figure 1 A flowchart illustrating an auxiliary assessment method for brain diseases that integrates artificial intelligence and holographic physiological functions, provided as an embodiment of this disclosure.

[0019] Figure 2 This is a flowchart of motion event detection and multidimensional feature extraction provided in an embodiment of the present disclosure.

[0020] Figure 3 A flowchart illustrating the three-stage representation of motion events provided in this embodiment of the disclosure.

[0021] Figure 4 This is a schematic diagram of the quality-aware multi-task temporal convolutional network model structure provided in an embodiment of this disclosure.

[0022] Figure 5 This is a schematic block diagram of a brain disease auxiliary assessment system that integrates artificial intelligence and holographic physiological functions, provided as an embodiment of this disclosure.

[0023] Figure 6 This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0025] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, collected data, used data, generated data, processed data, etc.) and the data (including but not limited to data used for analysis, stored data, displayed data, collected information, used information, generated information, processed information, etc.) are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0026] According to embodiments of this disclosure, a method for assisting in the assessment of brain diseases by integrating artificial intelligence and holographic physiological functions is provided. The holographic physiological functions refer to a comprehensive representation of multi-channel physiological and motor information collected by devices such as smartwatches, including triaxial acceleration, triaxial angular velocity, photoplethysmography (PPG), heart rate and interval, heart rate variability, blood oxygen saturation, skin temperature, and wrist-off status, as well as the temporal correlation of the above information during sleep. This method is suitable for home sleep scenarios. Before falling asleep, the wearer puts on a smartwatch equipped with a triaxial accelerometer, triaxial angular velocity sensor, PPG sensor, blood oxygen sensor, and skin temperature sensor. The smartwatch continuously collects multi-channel physiological signals throughout the night. The collected data is transmitted via Bluetooth or a wireless network to a paired mobile terminal or uploaded to a cloud server. The server or terminal performs data processing and model inference to identify motor abnormalities related to rapid eye movement sleep behavior disorder (RBD), thereby assisting in the early assessment of neurodegenerative diseases such as Parkinson's disease and Lewy body dementia.

[0027] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0028] Please see Figure 1 , Figure 1 This is a flowchart of a brain disease auxiliary assessment method integrating artificial intelligence and holographic physiological functions according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes steps S1-S7: In step S1, multi-night smartwatch data is acquired, and the data for each night is divided into nighttime segments. Sleep stage probability vectors and segment quality weights are generated for each nighttime segment.

[0029] Specifically, the system acquires smartwatch data from the wearer over several consecutive nights. Each night's data includes three-axis acceleration signals, three-axis angular velocity signals, beat rate and heart rate interval, heart rate variability (HRV) index, blood oxygen saturation (SpO2), skin temperature, wrist-off status indicators, and sleep stage estimates output by the watch's built-in sleep staging algorithm. The system divides each night's data into several night segments along the time axis using fixed windows. The window duration can be configured, for example, to 30 seconds, and adjacent windows do not overlap.

[0030] Consumer smartwatches have inherent errors in sleep stage segmentation, especially at the boundaries between REM sleep and light sleep, and between REM sleep and brief awakening. If the stage labels output by the smartwatch are directly treated as certain facts, subsequent anomaly identification will be directly affected by this error. To address this, this embodiment converts the smartwatch's sleep stage output into a four-dimensional probability vector to replace the deterministic labels, allowing fragments in ambiguous boundary regions to participate in subsequent processing with partial weight, rather than being forcibly assigned to a certain stage.

[0031] Specifically, each component of the four-dimensional probability vector corresponds to the probability of the REM sleep stage, the probability of the awake stage, the probability of the light sleep stage, and the probability of the deep sleep stage, and the sum of the four components is 1. The generation method is divided into two methods depending on the output capability of the watch: when the watch directly outputs the confidence score or probability value of each stage, the system normalizes it to obtain the vector; when the watch only outputs a single stage label without providing confidence, the system inputs three features of the night segment—root mean square magnitude of acceleration, mean heart rate, and root mean square of continuous difference (RMSSD)—into a pre-trained lightweight classifier, which outputs a four-dimensional probability distribution as the sleep stage probability vector. This lightweight classifier can be implemented using logistic regression or a small-scale fully connected network, and is pre-trained on a public dataset with existing polysomnography annotations, with fixed parameters after deployment.

[0032] In subsequent processing, whenever it is necessary to determine whether a segment belongs to the suspected REM stage, the value of the REM component in the probability vector is used as the weight for the night segment in the suspected REM motion anomaly analysis. The higher the REM component, the greater the contribution of the segment to the subsequent REM anomaly determination. Thus, even if some segments are misjudged by the watch, some of their REM features can still enter the subsequent analysis with appropriate weights, thereby mitigating the impact of consumer-grade installment errors.

[0033] In one embodiment, the system simultaneously calculates segment quality weights for each nighttime segment, which are used to adjust the contribution of each segment in subsequent model inference and training. When the wearer is active at night, the PPG signal of a consumer watch may be degraded due to motion artifacts, the accelerometer may reach its upper limit during vigorous movement, and a loose or detached watch may also lead to unreliable data. The role of segment quality weights is to quantify the degree of these interferences.

[0034] Specifically, the fragment quality weight is determined by four factors: PPG signal quality score. The signal-to-noise ratio and peak identification success rate of the PPG waveform are used to obtain the value, which ranges from 0 to 1; the acceleration signal saturation ratio is also used. This refers to the percentage of sampling points where the synthesized acceleration amplitude within a segment reaches the upper limit of the sensor's range, with a value ranging from 0 to 1; skin contact stability factor. This reflects the degree of fluctuation in the contact state between the watch and the skin, taking a value of 1 when the contact is stable and decreasing linearly to 0 when the fluctuation is large; Data Integrity Factor The quality weight is set to 1 when there is no missing data in the segment, and decreases to 0 proportionally when the missing duration exceeds a certain proportion of the total segment duration. The segment quality weight is calculated by multiplying the above four factors: in This represents the proportion of unsaturated samples; the more saturated sampling points there are, the smaller this term becomes. The above formula means that any severe anomaly in any term will cause the overall weight to approach zero. When the wrist-off status flag indicates that the nighttime segment is in a wrist-off state... Set it directly to zero; of the four factors mentioned above, and It is calculated in real time segment by segment from the sensor signal, for example, in this embodiment, From the normalized PPG signal-to-noise ratio With peak recognition success rate The average value is obtained; where, and All normalized to 0 to 1; skin contact stability factor Based on the fluctuation amount output by the skin contact sensor or the baseline fluctuation amount of PPG Sure: in, Maximum permissible contact fluctuation set before deployment; data integrity factor Based on the proportion of missing data within the fragment Sure: in, This is the ratio of the duration of missing data within a segment to the total duration of the segment. The maximum allowable missing percentage set before deployment; It is determined by the ratio of the duration of missing sensor data within a segment to the total duration of the segment.

[0035] In step S2, suspected REM segments, suspected awake segments, and relatively stable sleep segments are determined based on the sleep stage probability vector, and the individual physiological reference value for the night is generated from the relatively stable sleep segments.

[0036] In some embodiments, the system divides nighttime segments of the same night into three categories based on the sleep stage probability vector; specifically, nighttime segments in which the REM component exceeds a preset threshold and is simultaneously higher than the wakefulness component, light sleep component, and deep sleep component are recorded as suspected REM segments; nighttime segments in which the wakefulness component exceeds a preset threshold and is simultaneously higher than the other three components are recorded as suspected wakefulness segments. The above two classification conditions require that a certain component both reach the absolute threshold and be the maximum value among the four components. Therefore, the same segment will not be classified into both suspected REM and suspected wakefulness categories simultaneously. From the nighttime segments where the light sleep component or deep sleep component is higher than the other components, segments with acceleration activity in the lowest quartile range of the night are selected and recorded as relatively stable sleep segments. These segments represent the period with the smallest movement amplitude and the most stable physiological state during the night. The above preset threshold can be configured, for example, between 0.4 and 0.6, and is calibrated before deployment based on the watch model and the accuracy of the phased algorithm. When the number of relatively stable sleep segments is lower than the preset minimum number, or the average segment quality weight of relatively stable sleep segments is lower than the preset quality lower limit, the system will not generate the personal physiological reference value for that night, and will mark that night as a reference record, and will not include it in the subsequent multi-night summary calculation.

[0037] The system uses relatively stable sleep segments to calculate the individual's physiological reference values ​​for the night. For example, it includes four items: resting heart rate reference, which is the median of the heart rate in the relatively stable sleep segments; heart rate variability reference, which is the median of the RMSSD and the standard deviation of adjacent normal heartbeat intervals (SDNN); blood oxygen reference, which is the median of SpO2; and wrist micro-movement reference, which is the median and interquartile range of the root mean square amplitude of acceleration. The median is chosen instead of the mean because the median is not sensitive to occasional outliers. These reference values ​​are all from the same wearer's data from the same night. Using this as a benchmark can eliminate the interference of individual differences and nighttime baseline fluctuations in subsequent steps.

[0038] In step S3, motion events are detected from the nightly data, multidimensional features and autonomic nervous system response features of the motion events are extracted, and a three-segment representation of the motion events is generated by combining the individual's physiological reference value for the night.

[0039] Please see Figure 2 , Figure 2 A flowchart illustrating the motion event detection and multidimensional feature extraction process provided in this disclosure embodiment is shown, such as... Figure 2 As shown in box 201, motion events are detected from the nightly data.

[0040] Specifically, the system calculates the composite amplitude of the triaxial acceleration signal, which is the square root of the sum of the squares of the three-axis components. When the composite amplitude exceeds the resting noise threshold and the duration of the exceedance reaches the minimum event duration threshold, the segment is marked as a motion event. The resting noise threshold is determined based on the statistical level of the acceleration signal within the relatively stable sleep segment of the night, for example, by taking its mean plus several times the standard deviation, to ensure that minor body movements during quiet sleep are not misjudged. The minimum event duration threshold can be configured, for example, from 0.3 seconds to 1 second.

[0041] The criterion for determining the end of an event can be set as the synthesized amplitude falling below the resting noise threshold and remaining quiet for more than a preset interval. If the interval between two short-duration activities is less than the preset interval, they are merged into the same motion event. This can avoid the short pauses in the same turning process being split into multiple events. The system determines the segment to which the motion event belongs based on the time overlap relationship between the motion event and the night segment. When the motion event spans more than two night segments, it is assigned to the night segment with the longest overlap. If the overlap is the same, it is assigned to the night segment where the event started.

[0042] In box 202, for each motion event, the system extracts multidimensional features of the motion event. Specifically, for example, the multidimensional features of the motion event include: event duration; peak value of composite acceleration amplitude; mean value of composite acceleration amplitude; peak value of composite angular velocity amplitude; number of changes in the direction of triaxial acceleration, i.e., the cumulative number of times the acceleration vector direction is significantly deflected during the event, and the angle threshold for deflection judgment is configurable; duration of stillness before the start of the event, i.e., the continuous duration from the end of the previous motion event to the start of this event when the composite amplitude is below the noise threshold; duration of stillness after the end of the event, defined in the same way as the former; and the amount of wrist posture change before and after the event, obtained by calculating the angle between the mean acceleration vectors within a short window before the start of the event and after the end of the event.

[0043] In box 203, autonomic neural response features are extracted within a certain time range before and after each motion event. This range can be configured, for example, from 30 seconds before the event to 60 seconds after the event, and includes six exemplary items: heart rate rise rate, defined as the increase in heart rate relative to the pre-event baseline after the event divided by the rise time; heart rate recovery rate, defined as the decrease in heart rate peak relative to the pre-event baseline heart rate divided by the recovery time, i.e.: in, The peak heart rate within the observation range after the event. Baseline heart rate before the event. The time taken for the heart rate to return from its peak to the pre-event baseline heart rate level; if it does not return to the baseline level within the later observation window, the duration corresponding to the end of that observation window is used as the baseline. .

[0044] RMSSD change, which is the difference between the RMSSD after the event and the RMSSD before the event; SDNN change, defined in the same way; SpO2 short-term change, which is the difference between the lowest SpO2 value after the event and the average SpO2 value before the event; PPG signal quality change; where the baseline heart rate before the event is the average heart rate during the resting period before the event.

[0045] In one embodiment, actions such as ordinary turning over or getting up at night and short, unconscious movements suspected of being associated with RBD may be similar in intensity, but they differ in the temporal relationship between the movement and the autonomic nervous system response. Ordinary turning over typically involves a significant change in body position followed by a moderate increase in heart rate and a slow recovery. In contrast, movements suspected of being associated with RBD tend to have more frequent and irregular changes in direction, and the heart rate recovery pattern after the movement also differs from that of ordinary turning over. If the statistics for movement intensity and physiological indicators are calculated separately within the same time window, this temporal relationship will be lost in the same set of values, making it difficult for the model to distinguish between the two types of events. The three-segment representation structures the movement event and its preceding and following physiological responses along the time axis into three sub-sequences, preserving the aforementioned temporal correspondence.

[0046] Specifically, please refer to Figure 3 , Figure 3 A flowchart illustrating the three-stage representation of motion events provided in an embodiment of this disclosure is shown, as follows: Figure 3 As shown in box 301, the motion event is divided into a pre-event segment, an event segment, and a post-event segment. Specifically, the system defines the time range before the start of the motion event as the pre-event segment, the duration of which can be configured, for example, from 15 to 30 seconds; the duration range of the motion event itself as the event segment; and the time range after the end of the motion event as the post-event segment, the duration of which can be configured, for example, from 30 to 90 seconds. The post-event segment is longer than the pre-event segment because the recovery process of heart rate, HRV, and SpO2 usually requires a longer observation window.

[0047] The first part includes basic indicators and corresponding deviations. The basic indicators include: mean composite acceleration amplitude, standard deviation of composite acceleration amplitude, mean heart rate, RMSSD, SDNN, and mean SpO2. The corresponding deviations include the differences between the above basic indicators and the individual's physiological reference values ​​for the night. Specifically, the deviation of the mean composite acceleration amplitude from the median root mean square amplitude of acceleration in the wrist micro-motion reference, the deviation of the standard deviation of composite acceleration amplitude from the interquartile range in the wrist micro-motion reference, the deviation of the mean heart rate from the resting heart rate reference, the deviations of RMSSD and SDNN from the median RMSSD and SDNN in the heart rate variability reference, respectively, and the deviation of the mean SpO2 from the blood oxygen reference. These deviations reflect the deviation of the wearer's physiological state before the exercise event from the nighttime steady-state baseline.

[0048] The event segment contains the aforementioned multidimensional features of the motion event, as well as the sample-by-sample acceleration sequence and sample-by-sample angular velocity sequence during the event. The sample-by-sample sequence fully preserves the temporal details of the motion pattern, and the temporal distribution of the direction change and the amplitude envelope can be reflected in this sequence.

[0049] The latter part includes heart rate sequence, RMSSD change curve, SpO2 change curve, and heart rate recovery time; among them, heart rate recovery time Defined as the time required for heart rate to drop from its peak to the baseline level, this metric reflects the speed at which the autonomic nervous system returns to a steady state after an action.

[0050] In box 302, the sleep stage probability vector of the segment containing the motion event is appended to the end of the preceding, event, and following segments, forming a complete three-segment representation of the motion event. Specifically, to ensure that the three-segment representation of the motion event can serve as a fixed-dimensional component of the feature vectors of subsequent nighttime segments, the system resamples the sampled-point sequences, heart rate sequences, and change curves in the preceding, event, and following segments according to a preset sampling rate and preset length; portions exceeding the preset length are truncated in chronological order, and portions insufficient for the preset length are padded with zeros; statistical features are directly concatenated according to their original definitions. After the above processing, the three-segment representation of each motion event has the same dimension. The appended probability vector enables the subsequent model to understand the sleep stage context of the event when processing each feature segment; for example, the analytical meaning of the same morphological movement occurring in a segment with a higher REM probability and a segment with a higher wakefulness probability is different.

[0051] In step S4, a conscious action control representation is generated from the suspected conscious motor events, and a reference bias is generated from the nighttime segment indicators and the individual's physiological reference values ​​for the night.

[0052] Specifically, the system also performs the aforementioned multidimensional feature extraction of motion events and the aforementioned three-segment representation generation of motion events on motion events detected within suspected conscious segments, thereby obtaining a sample set of ordinary actions such as turning over, getting up at night, and adjusting the watch position by the wearer in a conscious state during the night, i.e., a conscious action sample set.

[0053] The generation method of the conscious action control representation varies depending on the number of events in the conscious action sample set. Specifically, when the number of events in the conscious action sample set reaches a preset number, the system sorts the fixed-dimensional three-segment representation of each motion event in the set according to the event occurrence time, and truncates or fills with zeros according to the preset maximum sample size to form a fixed-dimensional conscious action control representation. At the same time, the system calculates the median and interquartile range of the duration, number of direction changes, peak acceleration, peak angular velocity, post-action heart rate rise, post-action heart rate recovery time, and post-action RMSSD change for each event in the set to obtain the statistical characteristics of conscious actions.

[0054] When the number of events in the set of conscious actions is greater than zero but less than the preset number, the system calculates the median and interquartile range of the above indicators for each event in the set to obtain a fixed-dimensional conscious action comparison representation, and uses the statistical summary value of this fixed dimension as the conscious action statistical feature; when the set of conscious actions is empty, the system uses the zero vector of the same dimension as the statistical summary value as the conscious action comparison representation, and marks the conscious action statistical feature as invalid; the conscious action comparison representations generated in the above two valid cases are shared among the segments within the same night.

[0055] This allows subsequent models to compare waking actions from the same wearer on the same night when determining whether motion events in suspected REM fragments are abnormal, eliminating the influence of factors such as individual sleeping posture, watch tightness, and mattress firmness in control comparisons.

[0056] The system also generates a reference bias for each nighttime segment. The reference bias is the difference between the heart rate, RMSSD, SDNN, SpO2, and wrist micromotion level within that segment and the corresponding items in the individual's physiological reference values ​​for that night. The wrist micromotion level includes the deviation of the mean of the composite acceleration amplitude relative to the median of the wrist micromotion reference, and the deviation of the fluctuation level of the composite acceleration amplitude relative to the interquartile range of the wrist micromotion reference, reflecting the degree of deviation of that segment from the steady-state baseline of the same night.

[0057] In step S5, the sleep stage probability vector, the three-segment representation of motion events or the zero vector when there are no motion events, the wakefulness action control representation, the reference bias and the segment quality weight are concatenated into a night segment feature vector, and a night segment sequence is formed according to the time sequence.

[0058] Specifically, for each nighttime segment, the following components are concatenated to form a fixed-dimensional nighttime segment feature vector: a sleep stage probability vector; if there is a motion event within the segment, the three-segment representation of that motion event is taken; if there are multiple motion events within the segment, the fixed-dimensional three-segment representations of each motion event within the segment are weighted and averaged according to the event duration to obtain a segment-level three-segment representation of motion events; if there is no motion event, it is replaced by a zero vector of the same dimension; a wakefulness action comparison representation; a reference bias; and segment quality weights. For segments without motion events, the zero vector ensures that all segment feature vectors have the same dimension, and the model can distinguish whether there is a motion event by the difference between the zero vector and non-zero vectors. After concatenation, all nighttime segment feature vectors are arranged in chronological order to form a nighttime segment sequence as model input.

[0059] In step S6, the nighttime segment sequence is input into a quality-aware multi-task temporal convolutional network to obtain a suspected REM motion abnormality index, a motion and autonomic nervous system response inconsistency index, and a segment credibility score.

[0060] In some embodiments, please refer to Figure 4 , Figure 4 A schematic diagram of the quality-aware multi-task temporal convolutional network model structure provided in this disclosure embodiment is shown, such as... Figure 4 As shown, the model includes a dilated causal temporal convolutional backbone, a quality-aware feature aggregation module, a first output head, a second output head, and a third output head.

[0061] Specifically, the dilated causal temporal convolution backbone adopts a temporal convolutional network (TCN) structure, consisting of multiple residual blocks stacked sequentially. The number of residual blocks can be configured, for example, to be four. Each residual block contains two layers of one-dimensional causal dilated convolutions, where causality means that during each convolution operation, the output at each time step depends only on the input of the current and previous time steps, without using information from future segments. Each causal dilated convolution is followed by a weight normalization layer, a modified linear activation function (ReLU), and a dropout layer. Weight normalization decomposes the weight vector of the convolution kernel into directional and magnitude components for optimization, which helps to accelerate training convergence. The dropout layer randomly sets some channels to zero with a certain probability during training, reducing overfitting.

[0062] The inflation coefficient of each residual block increases exponentially, for example, 1, 2, 4, 8 in sequence. The kernel size can be configured as an integer between 3 and 7. The inflation coefficient refers to the spacing between elements of the kernel. The exponential growth causes the receptive field of the network to expand exponentially with the number of layers, covering a long time range of hundreds to thousands of segments with a low number of parameters. Each residual block has a skip connection: if the input dimension and the output dimension are not the same, the input is transformed by a one-dimensional convolutional layer with a kernel size of 1 and then added element by element to the output of the main path of the residual block; if the dimensions are the same, they are added directly.

[0063] Skip connections allow gradients to bypass the main path and propagate directly back during training, mitigating gradient vanishing in deep networks; the number of hidden channels in each residual block can be configured, for example, between 64 and 256; the backbone ultimately outputs a vector of length [missing information]. Hidden feature sequences ,in The total number of valid segments for that night, each This is the feature vector of the corresponding segment after processing by all residual blocks.

[0064] Specifically, the quality-aware feature aggregation module adjusts feature propagation using fragment quality weights at two locations. The first location is at the output of each residual block; the system will adjust the feature propagation at each time step. Corresponding hidden features Multiplying by the modulation term yields the modulation hidden feature, and the modulation term is calculated as the segment quality weight. With the first learnable scalar The product plus the second learnable scalar ,Right now . and During model training, updates are performed via backpropagation, with initial values ​​set to 1 and 0 respectively; when the quality weight of a certain segment... At lower levels, the modulation term tends to be close to The features of this segment are suppressed when propagating to subsequent residual blocks; when When it approaches 1, the modulation term approaches The features remain largely unaffected. This allows the noise features generated by low-quality segments to gradually weaken during the layer-by-layer propagation of long sequence convolutions, preventing localized noise caused by a watch loosening when a wearer turns over from polluting the overall features throughout the night.

[0065] The second position is after the final output of the main branch; the system performs weighted pooling on the hidden feature sequences according to fragment quality weights: in To prevent small constants with a denominator of zero, This refers to the quality-weighted overnight summary feature; weighted pooling ensures that high-quality segments contribute more to the overnight summary while reducing the impact of low-quality segments. For use by the first and second output heads.

[0066] The three output heads share the backbone parameters, but each has its own independent output layer parameters. Specifically, the first output head first performs weighted pooling on the hidden features corresponding to each suspected REM segment to obtain the suspected REM aggregated normalized features. : in, This is a collection of suspected REM fragments from the same night. For the first REM components of each night segment, This is the segment quality weight. No. The hidden features of the first night segment are output by the dilated causal temporal convolution backbone, corresponding to the first night segment. Hidden feature vectors of nighttime segments, To prevent small constants with a denominator of zero; the first output head receives a weighted, overnight summary of quality characteristics. and suspected REM summary features The two are concatenated and passed through two fully connected layers and a terminal sigmoid function to output a suspected REM motion anomaly index; if If empty, then Take the zero vector of the same dimension; this formula is passed Simultaneously considering the likelihood of a fragment belonging to the REM stage and the quality of the fragment data, fragments with a higher likelihood of REM and better data quality contribute more to the suspected REM summary features.

[0067] The second output head first performs weighted pooling on the hidden features corresponding to the motion events to obtain the summative features of the motion events. : in, This is a compilation of the events of the night's activities. For sports events The nighttime segment, For sports events The segment quality weight of the night segment; The output of the dilated causal temporal convolution backbone is related to motion events. The hidden feature vector corresponding to the night segment; To prevent small constants with a denominator of zero; the second output head receives... and The output motor inconsistency index with autonomic nervous system response; if If empty, then Take the zero vector of the same dimension; this formula uses the segment quality weight to perform weighted pooling on the hidden features corresponding to the motion event, so that the motion event segments with higher data quality contribute more to the summary features of the motion event.

[0068] The third output head receives the hidden features of each segment. After passing through a fully connected layer and a sigmoid activation function, each segment is independently output with a segment confidence score of 0 to 1. This score reflects the extent to which the data of the segment is suitable as an auxiliary evaluation basis. That is, although the segment is not hard excluded by the quality weight, there may still be soft interferences such as a slightly loose wristband or occasional jumps in heart rate extraction. The third output head learns to identify such situations through training and gives a low score.

[0069] In some embodiments, the neural network training process can optionally be divided into two stages. Specifically, the first stage is self-supervised pre-training: using a large amount of unlabeled nighttime watch data from normal individuals, by occluding some channel features (e.g., occluding heart rate recovery sequences or some motion event features), the model predicts the occluded content based on the remaining channels, enabling the TCN backbone to learn the general temporal relationship between movement and autonomic nervous system responses during normal sleep. The second stage is supervised fine-tuning: using a small dataset with polysomnography confirmation or clinical assessment annotations, the three output heads and the backbone are jointly fine-tuned; wherein, the total model loss is... ,in and The binary cross-entropy losses for the first and second tasks are respectively. The mean squared error loss for the third task; , , The preset loss weight coefficients can be configured, for example, to be 1.0, 0.8, or 0.5. Training uses the Adam optimizer, with the learning rate in the fine-tuning phase being lower than that in the pre-training phase.

[0070] In step S7, the credibility score of the segment is verified by the similar segments within the night to generate the credibility score of the verified segment, and the multi-night data that meets the quality screening conditions is summarized to output the auxiliary evaluation results.

[0071] In one embodiment, abnormal movements associated with RBD (Recurrent Mode Deformation) tend to occur multiple times within the same night, exhibiting pattern consistency; while occasional large movements, such as an arm hitting the headboard or a sudden twitch of the body, are isolated. Based on this characteristic, the system extracts hidden features corresponding to each motion event in suspected REM (Recurrent Mode Deformation) segments within the same night and calculates the pairwise cosine similarity between these hidden features. Cosine similarity measures the degree of proximity of two vectors in direction, with a value ranging from -1 to 1. The closer the value is to 1, the more similar the feature representations of the two events are.

[0072] When the cosine similarity of two motion events exceeds a preset similarity threshold, the system records these two events as similar motion events and multiplies their respective segment credibility scores by a gain factor greater than 1 with an upper limit to obtain the verification segment credibility score. The gain factor is positively correlated with the number of similar events that night; that is, the more similar events, the greater the gain. An upper limit is set to prevent extreme amplification, which can be configured, for example, to 1.5. The specific values ​​of the gain factor and the similarity threshold are calibrated before deployment based on the balance point between the false positive rate and the false negative rate on the validation set. Specifically, in this embodiment, the gain factor... Determine using the following formula: in, For example, take 1.5; The number of events that constitute similar motion events to the current motion event; This refers to the growth factor determined based on the validation set before deployment.

[0073] For each motion event, the system uses the segment credibility score of the night segment to which the motion event belongs as the segment credibility score of the motion event; the night segment to which the motion event belongs is determined according to the time overlap relationship between the motion event and the night segment.

[0074] Specifically, when the cosine similarity of a certain motion event with all other suspected REM segments in the same night does not exceed the similarity threshold, and the distance between the event segment feature vector and the conscious action statistical feature is less than the preset distance threshold, it indicates that the event is both isolated and similar to ordinary conscious actions. The system multiplies its segment credibility score by a decay factor less than 1 to reduce it, thus obtaining the verification segment credibility score. If the conscious action statistical feature is marked as invalid, the decay judgment based on the conscious action statistical feature is not performed.

[0075] In this embodiment, the attenuation factor Determine using the following formula: in, The distance between the event segment feature vector and the statistical features of effective conscious actions. For the preset distance threshold, For the preset lower limit; when hour, The system multiplies the segment credibility score by The credibility score of the reviewed segment is obtained.

[0076] The aforementioned distance can be Euclidean distance, which reflects the degree of similarity between the isolated event and a normal conscious action in the feature space. The smaller the distance, the more the event resembles a normal action, and the more justifiable the reason for reducing its credibility. The preset distance threshold can also be calibrated based on the validation set before deployment, which will not be elaborated here.

[0077] In one embodiment, for motion events that do not fall into either of the above two categories—that is, events that are neither similar events nor simultaneously satisfy the conditions of being isolated and resembling conscious actions—their segment confidence scores remain unchanged and are directly used as the review segment confidence scores. After review, recurring similar abnormal patterns receive higher confidence, while isolated events with features resembling ordinary actions are suppressed, thereby reducing false alarms caused by single, occasional events.

[0078] In some embodiments, auxiliary evaluation results are output after summarizing multi-night data that meets the quality screening criteria. Specifically, the quality screening criteria include: whether the effective wearing time of the night meets the minimum requirement, whether the total duration of suspected REM segments meets the minimum requirement, and whether the average segment quality weight of the whole night meets the minimum requirement; the above minimum requirements are preset according to the actual use scenario before system deployment; nights that do not meet the quality screening criteria, as well as nights where the number or quality of relatively stable sleep segments is insufficient to generate the personal physiological reference value for the night, are marked as reference records and are not included in the summary calculation.

[0079] For multi-night data that meets the criteria, the system calculates the cross-night mean and coefficient of variation for the suspected REM movement abnormality index and the cross-night mean and coefficient of variation for the movement-autonomic nervous system inconsistency index. The coefficient of variation, the ratio of the standard deviation to the mean, is used to measure the dispersion of the results for each night. Based on the above cross-night mean, coefficient of variation, and verification segment reliability score, the system generates an auxiliary evaluation result, which includes: the mean of the suspected REM movement abnormality index, the mean of the movement-autonomic nervous system inconsistency index, the cross-night consistency rating, and the data quality sufficiency rating. The cross-night consistency rating is determined by the coefficient of variation; the lower the coefficient of variation, the higher the stability of the abnormal pattern across different nights, and the higher the rating. The data quality sufficiency rating is determined by the number of valid nights included in the summary and the average segment quality weight for each night. Optionally, the system requires that at least a preset number of nights meeting the criteria be accumulated before outputting a formal evaluation report.

[0080] Please see Figure 5 , Figure 5This is a structural block diagram of a brain disease auxiliary assessment system integrating artificial intelligence and holographic physiological functions, provided in an embodiment of this application. Figure 5 As shown, the system includes: The data acquisition and segment processing module 501 is used to acquire multi-night smartwatch data, divide the data of each night into night segments, and generate a sleep stage probability vector and segment quality weight for each night segment. The segment classification and reference generation module 502 is used to determine suspected REM segments, suspected awake segments, and relatively stable sleep segments based on the probability vector of sleep stages, and to generate the personal physiological reference value for the night from the relatively stable sleep segments. The motion event processing module 503 is used to detect motion events from nightly data, extract multidimensional features and autonomic nervous system response features of motion events, and generate a three-segment representation of motion events by combining the individual's physiological reference value for the night. The control and bias generation module 504 is used to generate a control representation of conscious actions from motor events within suspected conscious segments, and to generate a reference bias from nighttime segment indicators and the individual's physiological reference values ​​for the night. The feature sequence generation module 505 is used to concatenate the sleep stage probability vector, the three-segment representation of motion events or the zero vector when there are no motion events, the wakeful action comparison representation, the reference bias and the segment quality weight into a night segment feature vector, and form a night segment sequence according to time sequence. The model inference module 506 is used to input the nighttime segment sequence into the quality-aware multi-task temporal convolutional network to obtain the suspected REM motion abnormality index, the inconsistency index between motion and autonomic nervous system response, and the segment credibility score. The results output module 507 is used to generate a verified segment credibility score by reviewing the segment credibility score with similar segments within the night, and to summarize multi-night data that meet the quality screening conditions to output auxiliary evaluation results.

[0081] Each processing unit and / or module in the embodiments of this application can be implemented by an analog circuit that implements the functions described in the embodiments of this application, or by software that executes the functions described in the embodiments of this application.

[0082] Based on the same inventive concept, this application also provides an electronic device, the method corresponding to which can be the method in the foregoing embodiments, and its problem-solving principle is similar to that method. For example... Figure 6 As shown, Figure 6This is a schematic block diagram of an electronic device provided in an embodiment of the present disclosure. The device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the methods and / or technical solutions of the foregoing embodiments of the present application.

[0083] In particular, the methods and / or embodiments in this application can be implemented as computer software programs. For example, the embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. When the computer program is executed by a processor, it performs the functions defined in the methods of this application.

[0084] Another embodiment of this application provides a storage medium storing computer program instructions thereon, which can be executed by a processor to implement the methods and / or technical solutions of any one or more embodiments of this application described above.

[0085] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. A brain disease auxiliary assessment method integrating artificial intelligence and holographic physiological functions, characterized in that, include: Acquire multi-night smartwatch data, divide each night's data into nighttime segments, and generate a sleep stage probability vector and segment quality weight for each nighttime segment; Based on the probability vector of sleep stages, suspected REM segments, suspected awake segments, and relatively stable sleep segments are identified, and the individual physiological reference value for the night is generated from the relatively stable sleep segments. Motion events are detected from nightly data, multidimensional features and autonomic nervous system response features of motion events are extracted, and a three-segment representation of motion events is generated by combining the individual's physiological reference value for the night. A control representation of conscious actions was generated from motor events within suspected conscious segments, and a reference bias was generated from nighttime segment indicators and individual physiological reference values ​​for the night. The sleep stage probability vector, the three-segment representation of motor events or the zero vector when there are no motor events, the wakeful action control representation, the reference bias and the segment quality weight are concatenated to form the night segment feature vector, and the night segment sequence is formed according to the time sequence. Nighttime segment sequences are input into a quality-aware multi-task temporal convolutional network to obtain a suspected REM motion abnormality index, a motion-autonomic nervous system inconsistency index, and segment credibility scores. The credibility score of the segment is verified by comparing it with similar segments within the night to generate the credibility score of the verified segment. The multi-night data that meets the quality screening criteria are then summarized to output the auxiliary evaluation results.

2. The method according to claim 1, characterized in that, Generating the sleep stage probability vector includes: when the smartwatch outputs the confidence score or probability value of each sleep stage, normalizing the confidence score or probability value of each sleep stage to obtain the sleep stage probability vector; when the smartwatch only outputs a single stage label, inputting the root mean square amplitude of acceleration, mean heart rate, and root mean square of continuous difference of the night segment into a pre-trained lightweight classifier to obtain the sleep stage probability vector.

3. The method according to claim 2, characterized in that, The identification of suspected REM sleep segments, suspected awake sleep segments, and relatively stable sleep segments includes: nighttime segments in which the REM component exceeds a preset threshold and is higher than the awake, light sleep, and deep sleep components are recorded as suspected REM sleep segments; nighttime segments in which the awake component exceeds a preset threshold and is higher than the REM component, light sleep, and deep sleep components are recorded as suspected awake sleep segments; and from nighttime segments in which the light sleep component or deep sleep component is higher than the other components, nighttime segments in which the acceleration activity is in the lowest quartile range of the night are selected and recorded as relatively stable sleep segments.

4. The method according to claim 3, characterized in that, The individual physiological reference values ​​for the night include resting heart rate reference, heart rate variability reference, blood oxygenation reference, and wrist micro-movement reference. The resting heart rate reference is the median heart rate of a relatively stable sleep segment. The heart rate variability reference is the median of the root mean square difference of continuous differences and the median of the standard deviation of adjacent normal heartbeat intervals in a relatively stable sleep segment. The blood oxygenation reference is the median blood oxygen saturation in a relatively stable sleep segment. The wrist micro-movement reference is the median and interquartile range of the root mean square amplitude of acceleration in a relatively stable sleep segment.

5. The method according to claim 4, characterized in that, The motion event detection process includes: calculating the composite amplitude of the triaxial acceleration signal; marking a motion event when the composite amplitude exceeds the resting noise threshold and the duration exceeds the shortest event duration threshold; determining the end of the motion event when the composite amplitude falls below the resting noise threshold and remains quiet for more than a preset interval; and merging short-term activities with adjacent intervals less than the preset interval into the same motion event.

6. The method according to claim 5, characterized in that, Multidimensional features of motion events include event duration, peak value of composite acceleration amplitude, mean value of composite acceleration amplitude, peak value of composite angular velocity amplitude, number of triaxial acceleration direction changes, duration of stillness before the start of the event, duration of stillness after the end of the event, and changes in wrist posture before and after the event; autonomic nervous system response features include heart rate rise rate, heart rate recovery rate, root mean square change in continuous difference, standard deviation change in adjacent normal heartbeat intervals, short-term changes in blood oxygen saturation, and changes in PPG signal quality.

7. The method according to claim 6, characterized in that, The three-segment representation of a motion event is generated as follows: the time range before the start of the motion event is defined as the first segment, the duration of the motion event is defined as the event segment, and the time range after the end of the motion event is defined as the second segment. The first segment includes the mean of the composite acceleration amplitude, the standard deviation of the composite acceleration amplitude, the mean heart rate, the root mean square of the continuous difference, the standard deviation of the interval between adjacent normal heartbeats, the mean blood oxygen saturation, and the deviation of each indicator relative to the individual's physiological reference value for the night. The event segment includes the multidimensional features of the motion event, the acceleration sequence at each sampling point, and the angular velocity sequence at each sampling point. The second segment includes the heart rate sequence, the root mean square curve of the continuous difference, the blood oxygen saturation curve, and the heart rate recovery time. The probability vector of the sleep stage is appended to the first segment, the event segment, and the second segment to obtain the three-segment representation of the motion event.

8. The method according to claim 7, characterized in that, The generation of a conscious action control representation includes: performing multidimensional feature extraction and three-segment representation generation of motion events within suspected conscious segments to obtain a conscious action sample set; when the number of events in the conscious action sample set reaches a preset number, the three-segment representation of motion events in the conscious action sample set is used as the conscious action control representation; when the number of events in the conscious action sample set does not reach the preset number, the duration, number of direction changes, peak acceleration, peak angular velocity, post-action heart rate rise, post-action heart rate recovery time, and post-action root mean square variation of continuous difference are statistically summarized to obtain the conscious action control representation.

9. The method according to claim 8, characterized in that, The quality-aware multi-task temporal convolutional network includes a dilated causal temporal convolutional backbone, a quality-aware feature aggregation module, a first output head, a second output head, and a third output head. The dilated causal temporal convolutional backbone consists of multiple stacked residual blocks. Each residual block includes two layers of one-dimensional causal dilated convolutions. After each layer of one-dimensional causal dilated convolutions, a weight normalization layer, a modified linear activation function, and a random dropout layer are connected in sequence. The residual blocks output hidden feature sequences through skip connections.

10. A brain disease auxiliary assessment system integrating artificial intelligence and holographic physiological functions, characterized in that, include: The data acquisition and segment processing module is used to acquire multi-night smartwatch data, divide the data of each night into night segments, and generate a sleep stage probability vector and segment quality weight for each night segment. The segment classification and reference generation module is used to identify suspected REM segments, suspected awake segments, and relatively stable sleep segments based on the probability vector of sleep stages, and to generate the personal physiological reference value for the night from the relatively stable sleep segments. The motion event processing module is used to detect motion events from nightly data, extract multidimensional features and autonomic nervous system response features of motion events, and generate a three-segment representation of motion events by combining the individual's physiological reference values ​​for the night. The control and bias generation module is used to generate a control representation of conscious actions from motor events within suspected conscious segments, and to generate a reference bias from nighttime segment indicators and the individual's physiological reference values ​​for the night. The feature sequence generation module is used to concatenate the sleep stage probability vector, the three-segment representation of motion events or the zero vector when there are no motion events, the wakeful action comparison representation, the reference bias and the segment quality weight into a night segment feature vector, and form a night segment sequence in chronological order. The model inference module is used to input the nighttime segment sequence into the quality-aware multi-task temporal convolutional network to obtain the suspected REM motion abnormality index, the inconsistency index between motion and autonomic nervous system response, and the segment credibility score. The results output module is used to generate a verified segment credibility score by reviewing the segment credibility score with similar segments within the night, and to summarize multi-night data that meet the quality screening criteria, and output auxiliary evaluation results.

Images