Clinical rehabilitation data collection system and method for encephalopathy

Abstract

The invention discloses an encephalopathy clinical rehabilitation data collection system and method, and relates to the field of rehabilitation data processing, and the method comprises the following steps: S1, obtaining original data in real time, and constructing a time-space correlation data set; s2, carrying out noise reduction and smoothing on the original data, marking a potential agent end period, preliminarily screening abnormal segments in rehabilitation training, and quantifying a subjective evaluation text into a symptom level; s3, extracting an exclusive feature set for an agent end phenomenon and rehabilitation abnormity, and realizing situation self-adaptive fusion of multi-modal features by adopting an attention mechanism; and S4, constructing a dichotomy deep learning model to distinguish exception types, and generating an exception point report containing event details in combination with a clinical annotation optimization decision threshold. According to the method, symptom fluctuation details of a patient in a real life scene can be captured, evaluation deviation caused by snapshot type data acquisition is avoided, and intervention direction errors caused by evaluation distortion are reduced.

Description

Technical Field

[0001] This invention relates to the field of rehabilitation data processing technology, and more particularly to brain diseases, specifically a system and method for collecting clinical rehabilitation data for brain diseases. Background Technology

[0002] Parkinson's disease (PD) is a common neurodegenerative disease. Its core pathological mechanism is the degeneration and death of dopaminergic neurons in the substantia nigra, which leads to a decrease in striatal dopamine levels, resulting in motor dysfunction (such as resting tremor, rigidity, bradykinesia, and postural instability) and non-motor symptoms (such as cognitive impairment, depression, and sleep disorders).

[0003] Statistics show that the global prevalence of Parkinson's disease (PD) in people aged 65 and over is approximately 1-2%, and the number of patients is increasing year by year as the population ages. PD patients experience severe impairment in their quality of life, with about 70% experiencing a decline in daily activities as the disease progresses, placing a heavy burden on families and the social healthcare system. Currently, the clinical management of PD focuses on drug therapy (such as levodopa) and rehabilitation training. Rehabilitation training, through targeted exercise interventions (such as balance training and gait correction), can effectively slow functional decline and serves as an important supplement to drug therapy.

[0004] In the rehabilitation process of patients with Parkinson's disease (PD), the collection of clinical data is fundamental to evaluating the effectiveness of interventions and optimizing treatment strategies. Traditional data collection methods are mainly based on intermittent assessments, such as weekly outpatient scale assessments (e.g., UPDRS scores) or single-point data sampling in specific training scenarios. This snapshot-style collection mode often fails to capture real-time dynamic changes in symptoms, leading to the loss of information in critical time windows. This may delay the timing of treatment adjustments, reduce rehabilitation efficiency, and even increase the risk of falls.

[0005] Therefore, it is necessary to improve upon the shortcomings of existing technologies in order to solve the above problems. Summary of the Invention

[0006] This invention overcomes the shortcomings of the prior art and provides a system and method for collecting clinical rehabilitation data for brain diseases.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is: a method for collecting clinical rehabilitation data for brain diseases, comprising the following steps:

[0008] S1. Real-time acquisition of raw data including exercise / physiological data, subjective assessments, and medication records, and the addition of timestamps and contextual labels to construct a spatiotemporally correlated dataset;

[0009] S2. Perform noise reduction and smoothing on the raw data, mark the end of the potential drug period and screen abnormal segments in rehabilitation training, and quantify the subjective assessment text into symptom levels.

[0010] S3. Extract specific feature sets for end-of-dose phenomenon and recovery abnormality respectively, and use attention mechanism to achieve context-adaptive fusion of multimodal features;

[0011] S4. Construct a binary classification deep learning model to distinguish abnormality types, combine clinical annotations to optimize the decision threshold, and generate an anomaly report containing event details.

[0012] In a preferred embodiment of the present invention, in step S1, the motion data includes three-dimensional acceleration, three-dimensional angular velocity, cadence, and gait cycle; the physiological data includes heart rate, electromyographic signals, and skin resistance; the subjective assessment includes a symptom self-rating scale, daily activity ability scores, and subjective feeling descriptions; and the medication record includes the drug name, single dose, medication time, route of administration, and drug type.

[0013] In a preferred embodiment of the present invention, step S2 includes the following sub-steps:

[0014] S21. Wavelet transform algorithm is used to remove high-frequency noise in motion data, and Kalman filter is used to optimize the smoothness of physiological signals.

[0015] S22. Based on the pharmacokinetic model, mark the end-of-phase period of potential agents and extract the motion feature fragments within this period;

[0016] S23. Calculate the deviation of motion indicators through a sliding time window to initially screen abnormal segments that deviate significantly from the baseline;

[0017] S24. Semantically encode the subjective assessment data to map the text descriptions into quantitative symptom levels.

[0018] In a preferred embodiment of the present invention, in step S22, the pharmacokinetic model uses a one-compartment model to describe the drug concentration decay process, as shown in the following formula: ;

[0019] in, After taking the medicine Blood drug concentration at any given time; This is the initial concentration of the drug; It eliminates the rate constant; It is time; It is a natural constant;

[0020] Using the aforementioned pharmacokinetic model, the concentration was calculated to decrease to... Time Specifically, regarding the formula Deformation, Solving : ;

[0021] When marking the end-of-dose period, by the start time and end time Predict and label the end-of-term period of potential agents: ;in, It's time to take the medication.

[0022] In a preferred embodiment of the present invention, in step S23, the difference between the deviation metric window index and the baseline is considered to be greater if the deviation is larger, indicating a higher probability of an anomaly. A standardized deviation formula is used: ;

[0023] in, It is the degree of deviation; It is the motion index value of the current window; It is the baseline indicator mean; It is the standard deviation of the baseline indicator.

[0024] In a preferred embodiment of the present invention, step S3 includes the following sub-steps:

[0025] S31. For the end-of-dose phenomenon, calculate the interval between the abnormal segment and the most recent medication time, extract the slope of the curve of duration after medication versus symptom intensity, and extract the power spectral density of the tremor signal and the motion amplitude attenuation rate in the rigid state.

[0026] S32. For abnormal points in normal rehabilitation, calculate the interval between the abnormal segment and the start / end time of training, extract the correlation coefficient between training intensity and movement error rate, and extract the correlation features between environmental parameters and gait stability based on context labels.

[0027] S33. An attention mechanism is used to fuse multimodal features, and dynamic weights are assigned to end-of-dose related features.

[0028] In a preferred embodiment of the present invention, in step S31, the interval duration is calculated by extracting the most recent medication timestamp from the collected medication records and extracting the start timestamp of the marked abnormal segments.

[0029] In the slope extraction, symptom intensity data are collected at 5-minute intervals during the marked potential end-of-dose period. Corresponding to the duration after medication, a straight line is fitted using the least squares method. The slope reflects the trend of symptom intensity change with the duration after medication.

[0030] In the power spectral density extraction, the noise-reduced flutter signal is divided into frames with a frame length of 2 seconds and an overlap rate of 50%. The Welch method is used to calculate the power spectral density of each frame and extract the energy proportion of the 3-7Hz frequency band.

[0031] In the extraction of motion amplitude decay rate, amplitude data of five consecutive repeated movements are extracted from the abnormal segments in the initial screening, and the decay rate of amplitude with the number of movements is calculated to reflect the rate of decline in motor ability caused by stiffness.

[0032] In a preferred embodiment of the present invention, in step S32, the interval between the abnormal segment and the training start / end time is used to determine whether the abnormality is directly related to the training process;

[0033] In the extraction of the correlation coefficient, the Pearson correlation coefficient is used to quantify the association between training load and movement quality and to determine whether the abnormality is caused by the intensity exceeding the patient's tolerance.

[0034] In the aforementioned feature extraction, the influence of ground or lighting environmental factors on gait abnormalities is analyzed, and the influence weights of environmental parameters on gait stability are calculated through multiple linear regression.

[0035] In a preferred embodiment of the present invention, step S4 includes the following sub-steps:

[0036] S41. Construct a binary classification deep learning model, input the fused multimodal features, and output the classification results of end-of-dose phenomenon / normal recovery abnormality points;

[0037] The binary classification deep learning model consists of an LSTM layer, a Dropout layer, a fully connected layer, and a Softmax layer. The Softmax layer outputs probabilities, which are used to convert a 2D vector into class probabilities to facilitate the determination of classification confidence. ;

[0038] in, It is a predicted probability vector; It is the output vector of the fully connected layer; yes The first element of the vector; yes The second element of the vector; It is a natural constant;

[0039] S42. Using clinically labeled data, supervised learning is performed during model training, and the threshold is optimized using F1-score.

[0040] S43. If it is determined to be the end-of-dose phenomenon, simultaneously verify whether the period is in the potential end-of-dose period and whether the symptom level increases with time. If it is determined to be a normal rehabilitation abnormal point, confirm that it is significantly correlated with the training intensity.

[0041] S44. Generate an anomaly classification report that includes the event occurrence time, duration, feature similarity score, and judgment criteria.

[0042] This invention provides a data collection system for a method of collecting clinical rehabilitation data for brain diseases, comprising:

[0043] The data acquisition module is used to acquire raw patient data in real time, including exercise / physiological data, subjective assessment texts, and medication records;

[0044] The spatiotemporal correlation module is used to add unified timestamps and contextual labels to the original data and to build a standardized dataset containing spatiotemporal correlations.

[0045] The data preprocessing module is used to perform high-frequency noise filtering and smoothing optimization of physiological signals in motion data, and to label potential end-of-dose periods related to drug metabolism and quantify symptom levels in subjective assessment texts.

[0046] The feature extraction and fusion module is used to extract specific feature sets for two types of events: end-of-dose phenomenon and abnormal rehabilitation training, and to perform context-adaptive fusion of multimodal features using an attention mechanism.

[0047] The dual-classification decision module is used to distinguish between end-of-dose phenomena and abnormal rehabilitation training types based on a binary classification deep learning model, and generate anomaly reports containing event details.

[0048] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0049] (1) This invention provides a data collection system and method for clinical rehabilitation of brain diseases. By collecting motor / physiological data and combining drug records, subjective assessments and contextual tags, it realizes continuous monitoring of motor function and non-motor symptoms in the daily activities of Parkinson's disease patients. It covers the complete cycle such as the drug action period and the end-of-dose phenomenon period, breaks through the time limitation of traditional intermittent outpatient assessments, and can capture the details of symptom fluctuations in real life scenarios. It avoids assessment bias caused by snapshot-style data collection and fully reflects the intraday / interday variation pattern of patient symptoms. Compared with traditional collection methods that rely on patient subjective recall or fixed time point testing, the dynamic data provided by this invention is closer to the actual functional state of patients, providing an accurate baseline for the formulation of subsequent personalized rehabilitation plans and reducing the error in intervention direction caused by assessment distortion.

[0050] (2) In this invention, the algorithm model is used to extract features from real-time monitoring data, automatically identify the occurrence time, duration and severity of symptoms of end-of-dose phenomenon, and incorporate them as key parameters into the rehabilitation assessment system. This solves the core defect of traditional rehabilitation assessment that ignores fluctuations in drug efficacy, ensures that the assessment data includes the patient's functional trough state, avoids the mismatch between training programs based on drug peak periods and actual tolerance, and can adjust training intensity for end-of-dose periods, further reducing the risk of sports injury and improving the patient's compliance with long-term rehabilitation training.

[0051] (3) In this invention, specific feature sets are extracted for end-of-dose phenomenon and normal rehabilitation abnormality, and dynamic weighted fusion is performed through attention mechanism to accurately separate drug-related and training-related abnormalities. Through differentiated feature extraction, the biomarkers of the two types of abnormalities are focused, and the deep association between features and scenarios is achieved by combining contextual labels, avoiding misjudgment of intervention plan, supporting the clinical formulation of differentiated intervention strategies, and improving the accuracy of rehabilitation plan. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a preferred embodiment of the present invention for a method of collecting clinical rehabilitation data for brain diseases. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0056] Application Overview:

[0057] When the applicant conducted in-depth research on the symptom characteristics of Parkinson's disease during the recovery period, they discovered that the symptom characteristics of Parkinson's disease patients during the recovery period exhibit significant dynamic fluctuations and progressive characteristics. On the one hand, the severity of symptoms gradually worsens with the progression of the disease, and motor and non-motor functions show a progressive decline trend, requiring long-term and continuous rehabilitation intervention. On the other hand, during the stable period of the disease, the patient's symptoms are affected by factors such as drug efficacy, physical condition, and environmental stimuli, showing intraday fluctuations (such as the relief of symptoms during the on-phase and the worsening of symptoms during the off-phase after taking medication) and daytime differences (such as the impact of sleep quality on the next day's motor ability). This fluctuation makes it difficult for static baseline assessments to fully reflect the patient's true functional status.

[0058] The applicant discovered in clinical practice that the end-of-dose phenomenon, which is prevalent in drug therapy, fundamentally interferes with the collection of rehabilitation data. The end-of-dose phenomenon is a common complication in PD patients taking levodopa long-term, manifesting as a decline in drug efficacy over time, with a significant worsening of motor symptoms (such as stiffness and bradykinesia) or non-motor symptoms (such as anxiety and pain) before the next dose, typically lasting from tens of minutes to several hours. In normal rehabilitation processes, because traditional assessments are mostly conducted during outpatient follow-ups (when patients may be at the peak of drug efficacy), symptom fluctuations caused by the end-of-dose phenomenon are often overlooked. This leads to rehabilitation plans being developed based on idealized symptom states rather than the actual functional lows experienced by patients in their daily lives.

[0059] To address the aforementioned issues, this invention proposes a clinical rehabilitation data collection system and method for encephalopathy. By capturing symptom fluctuations such as end-of-dose phenomenon in real time, it enables precise adaptation of rehabilitation plans, thereby improving the rehabilitation outcomes and quality of life for PD patients.

[0060] like Figure 1 As shown, a method for collecting clinical rehabilitation data for brain diseases includes the following steps:

[0061] S1. Real-time acquisition of raw data including exercise / physiological data, subjective assessments, and medication records, and the addition of timestamps and contextual labels to construct a spatiotemporally correlated dataset;

[0062] S2. Perform noise reduction and smoothing on the raw data, mark the end of the potential drug period and screen abnormal segments in rehabilitation training, and quantify the subjective assessment text into symptom levels.

[0063] S3. Extract specific feature sets for end-of-dose phenomenon and recovery abnormality respectively, and use attention mechanism to achieve context-adaptive fusion of multimodal features;

[0064] S4. Construct a binary classification deep learning model to distinguish abnormality types, combine clinical annotations to optimize the decision threshold, and generate an anomaly report containing event details.

[0065] In some specific implementations, the implementation of step S1 stems from the fact that traditional Parkinson's disease rehabilitation data collection relies on snapshot-style assessments (such as weekly face-to-face consultations) and discrete scale records, which cannot continuously capture the dynamic correlation between symptom fluctuations (such as end-of-dose phenomenon) and drug metabolism cycles and rehabilitation training, resulting in the loss of key time window information. At the same time, multi-source data such as exercise, physiology, and subjective assessment lack spatiotemporal label integration, making it difficult to distinguish symptom triggering factors, resulting in insufficient data correlation and affecting the accuracy of rehabilitation program adjustments.

[0066] Furthermore, by fusing multimodal data and spatiotemporal correlation, and by using a unified timestamp to ensure temporal alignment, a three-dimensional correlation dataset of data, time, and context is constructed. This enables the structured integration and contextualized labeling of multi-source data, thereby laying a data foundation for subsequent accurate identification of abnormality types and significantly improving the continuity and interpretability of rehabilitation assessment.

[0067] Furthermore, the motion data includes three-dimensional acceleration (reflecting the amplitude of limb movement, such as the up-and-down vibration during walking), three-dimensional angular velocity (reflecting the rotational speed of joints, such as the angular velocity of arm swing), cadence (steps per minute), and gait cycle; these are collected in real time through wearable sensing devices worn on the patient's waist, wrist, or lower limbs, with a sampling frequency of no less than 50Hz.

[0068] Physiological data includes heart rate (reflecting physical load), electromyography (EMG) signals (reflecting muscle activity intensity, such as EMG amplitude during stiffness), and skin resistance (reflecting emotional fluctuations); these are collected by sensors integrated into wearable devices, such as photoelectric heart rate sensors to collect heart rate and surface EMG electrodes to collect muscle activity signals, with the sampling frequency synchronized with the motion data.

[0069] Subjective assessment included the Symptom Self-Rating Scale, Daily Activities of Life (ADL) scores, and subjective feelings descriptions; data were collected using the UPDRS-III Patient Self-Rating Version, ADL scales, and patient descriptions.

[0070] Medication records include the drug name, single dose, time of administration, route of administration, and type of drug; they are entered manually by the patient or automatically collected synchronously via a smart pillbox.

[0071] In this embodiment, all acquisition devices are synchronized to UTC standard time via the NTP protocol, with the error controlled within 1 second; for example, the local time of the wearable device is "08:30:05", and after calibration, it is uniformly "08:30:04" (correcting the device time deviation).

[0072] Optionally, the rules for context labeling are defined to automatically label data with context labels based on drug records, training plans, and real-time status.

[0073] For example, the context label types and rules are shown in Table 1.

[0074] Table 1: Context Label Types and Rules

[0075] Contextual tags Define conditions Post-medication window period In the medication record, the period from 0-2 hours after the start of medication administration (peak drug absorption period) is recorded. End of potential agent period In the medication record, the period from 4 to 6 hours after the start of medication administration (the period of lowest drug concentration) is recorded. During training In the rehabilitation training plan, the time from the start to the end of the training... resting state In motion data, periods with acceleration <0.1g and no training plan.

[0076] Example: If a patient takes medication at 08:00 and experiences worsening tremor at 12:30, the data is automatically marked as "potential end of medication period"; if balance training is underway at 16:00, the data is marked as "training in progress".

[0077] In the construction of the spatiotemporal correlation dataset, all data are arranged in ascending order by timestamp to form a time-series data sequence, and for each data sample, the corresponding context label is matched according to the timestamp to form a structured table.

[0078] For example, a structured table of a spatiotemporal correlation dataset is shown in Table 2.

[0079] Table 2: Structured Table of Spatiotemporal Correlation Dataset

[0080] Timestamp (UTC) Motion data (acceleration) Physiological data (heart rate) Subjective assessment Drug Records Contextual tags 202*-*-* 08:30:04 <![CDATA[0.32 m / s 2 ]]> 78 times / minute none Levodopa 100mg Post-medication window period 202*-*-* 12:15:20 <![CDATA[0.15 m / s 2 ]]> 92 times / minute Tremors worsened none End of potential agent period 202*-*-* 15:20:10 <![CDATA[0.8 m / s 2 ]]> 105 times / minute none none During training

[0081] In some specific implementations, the implementation of step S2 stems from the fact that the raw data collected in step S1 has problems such as noise interference, difficulty in quantifying subjective evaluation text, and mixing of abnormal fragments with normal fluctuations. Traditional preprocessing methods lack specific processing logic for Parkinson's disease rehabilitation data, resulting in low data quality, high rate of missed or false detection of abnormal events, and difficulty in directly supporting subsequent feature extraction and model training.

[0082] Furthermore, by employing a progressive logic based on data cleaning, time period marking, anomaly screening, and text quantification, the signal-to-noise ratio of the data is effectively improved. By marking the end of the potential agent period, key time windows are focused on, reducing interference from invalid data. Among these, the anomaly screening can compress the amount of data for subsequent analysis by at least 60%, while semantic encoding enables quantitative modeling of subjective evaluation.

[0083] Furthermore, step S2 includes the following sub-steps:

[0084] S21. Wavelet transform algorithm is used to remove high-frequency noise in motion data, and Kalman filter is used to optimize the smoothness of physiological signals.

[0085] S22. Based on the pharmacokinetic model, mark the end-of-phase period of potential agents and extract the motion feature fragments within this period;

[0086] S23. Calculate the deviation of motion indicators through a sliding time window to initially screen abnormal segments that deviate significantly from the baseline;

[0087] S24. Semantically encode the subjective assessment data to map the text descriptions into quantitative symptom levels.

[0088] In this embodiment, in step S21, the wavelet transform algorithm removes high-frequency noise (such as sensor vibration and muscle tremors) from motion signals (such as acceleration and angular velocity) through signal decomposition, threshold denoising and reconstruction, while retaining the effective signals reflecting limb movement.

[0089] The signal decomposition process involves decomposing the original motion data (such as a three-dimensional acceleration sequence) into multiple frequency components from low to high frequency using a wavelet function. For example, an acceleration signal with a sampling frequency of 50 Hz is decomposed into 5 layers, each corresponding to a different frequency band (e.g., layer 1: 25-50 Hz high-frequency noise, layer 5: 0.78-1.56 Hz low-frequency effective signal).

[0090] Threshold denoising involves setting a threshold for the decomposed high-frequency components and setting noise signals with amplitudes below the threshold to zero. The threshold is determined using a heuristic thresholding method (e.g., taking 1.5 times the standard deviation of the signal at that layer). For example, in a certain high-frequency component, the noise fluctuation range is ±0.05 m / s. 2 Set the threshold to 0.08 m / s 2 All ±0.08m / s 2 Signals within a certain range are considered noise and removed.

[0091] The reconstruction process involves retaining the effective low-frequency components and the denoised high-frequency components, and then reconstructing the denoised motion signal using inverse wavelet transform. The formula is as follows: ;in, It is the reconstructed first Layer signals; and These are low-pass and high-pass reconstruction filters, respectively. and These are the low-frequency approximation component and the high-frequency detail component after noise reduction, respectively.

[0092] Furthermore, the optimization of physiological signal smoothness is specifically as follows: define the state variables (such as the current heart rate value) and observation variables (heart rate values ​​collected by the sensor) of the physiological signal, and initialize the state estimate and error covariance matrix; in the prediction phase, based on the state estimate of the previous moment, predict the theoretical state of the current moment (such as predicting the current heart rate based on the heart rate trend of the previous 10 seconds); in the update phase, the predicted value and the current sensor observation value are weighted and fused, and the weights are dynamically adjusted through the error covariance matrix (the larger the error of the observation value, the lower the weight, and the larger the error of the predicted value, the higher the weight), and output the smoothed heart rate value.

[0093] For example, the original electromyographic signal experienced a momentary jump (from 0.3mV to 1.2mV) due to electrode loosening. After Kalman filtering, the jump was corrected to 0.4mV (close to the signal trend before and after the time step), avoiding interference from outliers in subsequent assessments of muscle stiffness.

[0094] In this embodiment, in step S22, the drug metabolism kinetic model uses a one-compartment model to describe the drug concentration decay process, as shown in the following formula: ;

[0095] in, After taking the medicine Blood drug concentration at any given time; This is the initial concentration of the drug; It is the elimination rate constant (related to the half-life). ); It is time; It is a natural constant.

[0096] Furthermore, using a pharmacokinetic model, the concentration was calculated to decrease to... Time Specifically, regarding the formula Deformation, Solving : ;

[0097] At the end of the labeling period, by the start time and end time Predict and label the end-of-term period of potential agents:

[0098] ;

[0099] in, This refers to the time of administration and the end time of medication. The default duration is 1 hour, but it can be adjusted according to individual symptoms.

[0100] It should be noted that, according to clinical data, the half-life of levodopa is about 1-2 hours, the peak effect occurs 1-2 hours after administration, and the end-of-dose phenomenon usually occurs 4-6 hours after administration (the specific time may be adjusted according to individual patient differences, such as 7 hours for elderly patients whose metabolism is slower).

[0101] For example, suppose a patient takes levodopa and then... , , ,but: That is, approximately 4.64 hours after administration, the drug concentration drops below the therapeutic threshold.

[0102] If the patient ,but , The end-of-dose period is marked as 12:38-13:38.

[0103] Furthermore, in the extraction of motion feature segments at the end of the dose period, based on the start time... and end time The corresponding segments are extracted from time-aligned motion data; and tremor features, bradykinesia features, and gait features are extracted for typical symptoms of Parkinson's disease.

[0104] Among them, tremor features are extracted by calculating the dominant frequency energy ratio of the tremor signal (reflecting tremor intensity) and the standard deviation of tremor amplitude (reflecting tremor stability); bradykinesia features are extracted by calculating the mean limb movement speed and the smoothness of the movement trajectory; gait features are extracted by the step length variation coefficient (step length standard deviation / average step length) and step frequency (steps per minute).

[0105] Optionally, the extracted feature values ​​are converted to multiples relative to the baseline period to eliminate individual baseline differences: .

[0106] For example, the mean tremor amplitude during the baseline period (1-2 hours after medication) was 0.2 m / s. 2 The final velocity of the agent was 0.6 m / s. 2 Then the relative tremor amplitude = 0.6 / 0.2 = 3.0 (indicating that the tremor has worsened by 3 times).

[0107] In this embodiment, in step S23, the difference between the deviation metric window index and the baseline is used. The greater the deviation, the higher the probability of an anomaly. A standardized deviation formula is adopted:

[0108] ;

[0109] in, It is the degree of deviation; This is the motion index value of the current window (e.g., average tremor amplitude of 0.42 m / s). 2 ); It is the baseline indicator mean; It is the standard deviation of the baseline indicator;

[0110] The deviation threshold is set as the baseline mean + 3 times the standard deviation, which corresponds to a deviation of 3 (covering 99.7% of the normal data distribution). If the deviation of a window is ≥3, the window is marked as an abnormal window; otherwise, it is a normal window. When the number of consecutive abnormal windows is ≥3, these windows are merged into an abnormal segment, and their start and end timestamps are recorded.

[0111] For example, the deviation of window 1 (12:38:00-05) is 2.4 (<3, normal); the deviation of window 2 (12:38:02-07) is 3.2 (≥3, abnormal); the deviation of window 3 (12:38:04-09) is 3.5 (≥3, abnormal); and the deviation of window 4 (12:38:06-11) is 3.1 (≥3, abnormal).

[0112] Three consecutive abnormal windows (windows 2-4) are merged into an abnormal segment: 12:38:02-12:38:11 (lasting 9 seconds).

[0113] In this embodiment, in step S24, a mapping rule between symptom keywords and severity levels is established as a translation manual for semantic encoding.

[0114] For example, focusing on motor symptoms: Based on clinical criteria for Parkinson's disease (such as the UPDRS scale), core motor symptoms including tremor, rigidity, bradykinesia, and frozen gait are defined, as well as core non-motor symptoms including fatigue, pain, and low mood; and each symptom is divided into three levels: 1=mild, 2=moderate, and 3=severe. Each level is labeled with descriptive keywords by neurologists, and the mapping rules are shown in Table 3.

[0115] Table 3: Mapping Rules between Symptom Keywords and Severity Levels

[0116] Symptom types Mild (Level 1) keywords Moderate (Level 2) keywords Severe (Level 3) keywords Tremor Slight shaking, occasionally occurs Continuous shaking affects the ability to hold light objects. Violent shaking, unable to hold objects stiff Slight stiffness, slightly slower movement Significant stiffness and limited mobility Severe stiffness, unable to move independently frozen gait Occasional pauses, requiring quick, short steps Frequent pauses, requiring prompts from others Completely frozen, unable to move

[0117] In some specific implementations, the implementation of step S3 stems from the problem that traditional rehabilitation data collection lacks a differentiated identification mechanism for drug-related end-of-dose phenomena and training-related rehabilitation abnormalities. The clinical treatment paths for the two in drug adjustment and training optimization are completely different. Because traditional methods do not design extraction strategies for their characteristic differences, the risk of misjudging intervention plans is high. At the same time, data collection does not fully combine contextual labels such as medication window and training status for feature correlation analysis, making it difficult to analyze the triggering factors of symptom fluctuations, further reducing the accuracy of decision-making.

[0118] Furthermore, a two-stage logic of differentiated feature extraction and context-adaptive fusion is used to achieve accurate differentiation, effectively solving the problem of feature confusion between the two types of abnormal points. This significantly improves the accuracy of automatic classification of end-of-dose phenomenon and rehabilitation abnormal points. At the same time, the correlation information between symptoms and drugs, training, and environment is preserved during feature fusion, laying the foundation for subsequent abnormal point tracing (such as verifying the rationality of the end-of-dose phenomenon time period and the correlation between training intensity and rehabilitation abnormal points), and directly supporting the accurate formulation of clinical intervention plans.

[0119] Furthermore, step S3 includes the following sub-steps:

[0120] S31. For the end-of-dose phenomenon, calculate the interval between the abnormal segment and the most recent medication time, extract the slope of the curve of duration after medication versus symptom intensity, and extract the power spectral density of the tremor signal and the motion amplitude attenuation rate in the rigid state.

[0121] S32. For abnormal points in normal rehabilitation, calculate the interval between the abnormal segment and the start / end time of training, extract the correlation coefficient between training intensity and movement error rate, and extract the correlation features between environmental parameters and gait stability based on context labels.

[0122] S33. An attention mechanism is used to fuse multimodal features, and dynamic weights are assigned to end-of-dose related features.

[0123] In this embodiment, in step S31, the interval duration is calculated by extracting the most recent medication timestamp from the collected medication records and extracting the start timestamp of the abnormal segments marked in step S23. ;

[0124] in, It is the start time of the abnormal segment; This is the most recent time of medication; 3600 is the conversion factor from seconds to hours.

[0125] In slope extraction, symptom intensity data were collected at 5-minute intervals during the potential end-of-dose period marked in step S22, corresponding to the duration after medication administration. A straight line was fitted using the least squares method, and the slope reflects the trend of symptom intensity change with the duration after medication administration. The formula is as follows: ;

[0126] in, It is the slope of the curve (reflecting the rate of change in symptom intensity; a positive value indicates that the symptoms worsen over time). It refers to the number of data points; It is the first Duration after medication administration at each sampling point; It is the first Symptom intensity at each sampling point.

[0127] In the power spectral density extraction, the tremor signal after noise reduction in step S21 is framed, with a frame length of 2 seconds and an overlap rate of 50%. The Welch method is used to calculate the power spectral density (PSD) of each frame, and the energy proportion of the 3-7Hz frequency band (the characteristic frequency band of Parkinson's disease tremor) is extracted. The formula is as follows: ;

[0128] in, It is frequency Power spectral density at; It is the number of intra-frame sampling points; It is the sampling frequency; It is a signal The results of the Fast Fourier Transform; It is the modulo square operation of complex numbers.

[0129] In the extraction of motion amplitude decay rate, from the abnormal segments initially screened in step S23, amplitude data of five consecutive repetitive movements are extracted, and the decay rate of amplitude with the number of movements is calculated to reflect the rate of decline in motor ability caused by stiffness. The formula is as follows: ;

[0130] in, It is the motion amplitude decay rate; This is the range of motion for the first movement; It is the first The range of motion.

[0131] It should be noted that the end-of-dose phenomenon is strongly correlated with the drug metabolism cycle. By quantifying the dynamic changes of symptoms over time after medication by measuring the interval duration and slope, the temporal pattern of drug concentration decrease and symptom aggravation can be captured. The 3-7Hz characteristic frequency band of Parkinson's disease tremor and the amplitude decay caused by rigidity are biomarkers that distinguish it from other abnormalities such as training fatigue. Specific identification can be achieved through power spectral density and motion amplitude decay rate.

[0132] In this embodiment, in step S32, the interval between the abnormal segment and the training start / end time is used to determine whether the abnormality is directly related to the training process. The formula for calculating the interval is as follows:

[0133] ;

[0134] in, It is the start time of the abnormal segment; and These are the start and end times of the most recent training session.

[0135] In the correlation coefficient extraction, the Pearson correlation coefficient is used to quantify the association between training load and movement quality, and to determine whether the abnormality is caused by the intensity exceeding the patient's tolerance.

[0136] ;

[0137] in, It is the correlation coefficient (with a value range of [-1, 1], and a positive value indicates that the error rate increases as the intensity increases). It is data to quantity; It is the first Training intensity within a time window; It is the first Action error rate for each time window.

[0138] In the feature extraction process, the influence of ground or lighting environmental factors on gait anomalies is analyzed. The weights of environmental parameters on gait stability are calculated using multiple linear regression, as shown in the following formula:

[0139] ;

[0140] in, It is the step size variation coefficient (reflecting gait stability; the lower the value, the more stable the gait). and These are the regression coefficients (weights) for ground smoothness and light intensity, respectively, representing the weights of each unit change in this environmental parameter. The impact; It is a constant term; It is an error term; It refers to the flatness of the ground (1-5 levels, 1 = extremely uneven, 5 = extremely smooth). It refers to light intensity.

[0141] It should be noted that by extracting multi-dimensional features from abnormal points in normal rehabilitation, we can accurately characterize abnormalities related to rehabilitation training. This not only solves the problem that traditional methods cannot distinguish the causes of abnormalities, but also provides an interpretable basis for subsequent clinical decision-making.

[0142] In this embodiment, in step S33, the multimodal features include the end-of-dose phenomenon-related features extracted in step S31 and the normal recovery abnormality-related features extracted in step S32.

[0143] Optionally, before multimodal feature fusion, each feature needs to be Z-score standardized to eliminate the influence of dimensions. ;

[0144] in, It is the first The first sample One original feature; and They are the first The mean and standard deviation of each feature.

[0145] Furthermore, during multimodal feature fusion, the standardized feature matrix will be... As keys and values, the end-of-dose indicator vector, keys, and values ​​are mapped to the same dimension through linear transformation;

[0146] Feature weights are calculated using scaled dot product attention. The focus is on enhancing the feature weights that match the indicator vector at the end of the dosing period:

[0147] ;

[0148] in, It is a query vector used to indicate the scenario that the model should currently focus on; It is the transpose of the key vectors; It is the dimension of the key, used to scale the inner product result to avoid gradient vanishing.

[0149] Specifically, the dynamic weighting logic is as follows: when the end-of-dose period indicator vector is equal to [1,0] (end-of-dose period), the key corresponding to the extracted end-of-dose features will have a higher similarity to the query and will receive a greater weight after Softmax.

[0150] Furthermore, each feature value is multiplied by its corresponding weight to obtain the final fused feature.

[0151] It should be noted that the attention mechanism simulates the diagnostic logic of clinicians, achieving dynamic focusing of multimodal features. This not only leverages the diagnostic value of end-of-dose features at specific times but also avoids interference from irrelevant features.

[0152] In some specific implementations, the implementation of step S4 stems from the clinical need for rehabilitation abnormality classification. After feature extraction and attention fusion in step S3, it is necessary to further map the multimodal features into two categories of labels: end-of-dose phenomenon or normal rehabilitation abnormality points, in order to support subsequent precise intervention.

[0153] Specifically, traditional classification methods (such as SVM and ordinary neural networks) are difficult to handle high-dimensional dynamic features, and the two types of abnormalities have overlapping features (such as both may have increased action error rates), resulting in a high misdiagnosis rate and making it difficult to meet the clinical requirements for accurate localization of abnormal causes.

[0154] Furthermore, by adopting an LSTM and fully connected layer binary classification architecture based on the feature-category mapping principle, the accuracy of abnormal classification is significantly improved, thereby enhancing the model's comprehensive classification accuracy for end-of-dose phenomena and abnormal points in normal rehabilitation, effectively reducing the risk of incorrect adjustment of rehabilitation plans due to misdiagnosis.

[0155] Furthermore, step S4 includes the following sub-steps:

[0156] S41. Construct a binary classification deep learning model, input the fused multimodal features, and output the classification results of end-of-dose phenomenon / normal recovery abnormality points;

[0157] S42. Using clinically labeled data, supervised learning is performed during model training, and the threshold is optimized using F1-score.

[0158] S43. If it is determined to be the end-of-dose phenomenon, simultaneously verify whether the period is in the potential end-of-dose period and whether the symptom level increases with time. If it is determined to be a normal rehabilitation abnormal point, confirm that it is significantly correlated with the training intensity.

[0159] S44. Generate an anomaly classification report that includes the event occurrence time, duration, feature similarity score, and judgment criteria.

[0160] In this embodiment, in step S41, the structure of the binary classification deep learning model includes an LSTM layer (capturing time dependencies), a Dropout layer (preventing overfitting), a fully connected layer (feature mapping), and a Softmax layer (outputting classification probabilities).

[0161] Specifically, the LSTM layer is used for feature extraction to learn the dynamic changes of features over time. The hidden layer has a dimension of 64, and it is a bidirectional LSTM (simultaneously focusing on features from past and future time steps), outputting only the features from the last time step.

[0162] Dropout layer regularization is used to randomly drop some neurons to prevent the model from over-relying on a certain feature;

[0163] The fully connected layer feature mapping is used to map the high-dimensional features output by the LSTM to binary classification probabilities; the first layer has a 64-dimensional input and a 32-dimensional output, with the ReLU activation function; the second layer has a 32-dimensional input and a 2-dimensional output (corresponding to two class labels).

[0164] The Softmax layer outputs probabilities, which are used to convert a 2D vector into class probabilities (summing to 1), facilitating the determination of classification confidence.

[0165] ;

[0166] in, It is a predicted probability vector; It is the output vector of the fully connected layer; yes The first element of the vector (corresponding to the original score of the abnormal point in normal rehabilitation); yes The second element of the vector (corresponding to the original fraction of the end-of-dose phenomenon); It is a natural constant.

[0167] Furthermore, the loss function for the binary classification deep learning model uses binary cross-entropy:

[0168] ;

[0169] in, It is the sample size; It is the first The true label of each sample; It is the first The probability of each sample being predicted as end-of-dose phenomenon; It is the first The probability of each sample being predicted as a normal recovery abnormality.

[0170] In this embodiment, in step S42, the data source is: the rehabilitation training data of 1200 Parkinson's patients are included, and the data is jointly labeled by 3 neurologists according to the "Clinical Diagnostic Criteria for Movement Disorders"; and samples with fuzzy labels are removed, missing values ​​are processed, and data cleaning is performed.

[0171] The dataset is divided into a 6:2:2 ratio: a training set (720 cases) for learning model parameters (input features and expert-annotated labels); a validation set (240 cases) for threshold optimization (independent of the training process to avoid overfitting); and a test set (240 cases) for final performance evaluation (simulating real clinical scenarios).

[0172] Specifically, using clinically labeled data as a supervisory signal, the model parameters are optimized through backpropagation, enabling the model to learn the mapping patterns between features and labels.

[0173] The threshold search logic in threshold optimization is as follows: the classification threshold is the probability threshold for the model to determine the end-of-dose phenomenon (e.g., when the threshold = 0.5, the predicted probability ≥ 0.5 is judged as the end of the dose). In clinical practice, missed diagnosis is more harmful than misdiagnosis, but the two need to be balanced. Therefore, 81 candidate thresholds are tested in the interval [0.1, 0.9] with a step size of 0.01. The F1-score corresponding to each threshold is calculated, and the candidate threshold corresponding to the maximum value is selected as the final threshold.

[0174] In this embodiment, in step S43, if the phenomenon is determined to be end-of-dose, the classification result is confirmed to be consistent with the clinicopathological characteristics of end-of-dose in Parkinson's patients by matching time patterns and testing symptom progression trends.

[0175] The specific time pattern matching is as follows: obtain the time of the most recent dose of anti-Parkinson's drug, record the start time of the abnormality, calculate the interval between the time of the dose and the start time of the abnormality. If the interval is within the range of 4 to 6 hours, it is determined to be in the potential end-of-dose period; otherwise, it is an unexpected period.

[0176] The symptom progression trend test is specifically conducted by extracting the symptom level score every 5 minutes during the abnormal period and observing whether the score gradually increases over time without significant decline. A trend test (such as the Mann-Kendall test) is used to determine whether the trend is significantly increasing. If the test results show a significant trend, the test is passed.

[0177] Furthermore, if a normal rehabilitation abnormality is identified, the correlation analysis between training intensity and the abnormality can confirm that the abnormality is caused by excessively high or low training intensity.

[0178] The correlation between training intensity and outliers is verified by recording the symptom level of the current outlier and the symptom levels of the last 10 normal training sessions. The correlation between intensity and symptom level is calculated through correlation analysis as in step S32. If there is a significant correlation, the verification is passed.

[0179] For example, in step S44, the template for the anomaly classification report is shown in Table 4.

[0180] Table 4: Report Template Examples

[0181] Module content Patient Information ID: P2025071701, Diagnosis: Parkinson's disease, Duration of illness: 3 years Error Details Event time: 202*-*-* 10:00 Duration: 80 minutes Feature similarity score: 96 points (highly similar) Judgment basis 1. Time: 5.5 hours after medication (4-6 hours potential end-of-dose window); 2. Characteristics: 96 points of similarity to historical typical cases; 3. Verification: Significant trend of symptom escalation.

[0182] This invention provides a data collection system for a method of collecting clinical rehabilitation data for brain diseases, comprising:

[0183] The data acquisition module is used to acquire raw patient data in real time, including exercise / physiological data, subjective assessment texts, and medication records;

[0184] The spatiotemporal correlation module is used to add unified timestamps and contextual labels to the original data and to build a standardized dataset containing spatiotemporal correlations.

[0185] The data preprocessing module is used to perform high-frequency noise filtering and smoothing optimization of physiological signals in motion data, and to label potential end-of-dose periods related to drug metabolism and quantify symptom levels in subjective assessment texts.

[0186] The feature extraction and fusion module is used to extract specific feature sets for two types of events: end-of-dose phenomenon and abnormal rehabilitation training, and to perform context-adaptive fusion of multimodal features using an attention mechanism.

[0187] The dual-classification decision module is used to distinguish between end-of-dose phenomena and abnormal rehabilitation training types based on a binary classification deep learning model, and generate anomaly reports containing event details.

[0188] It should be noted that the brain disease clinical rehabilitation data collection system can implement the steps in the brain disease clinical rehabilitation data collection method in the above embodiments and achieve the same technical effect. Refer to the description in the above embodiments, which will not be elaborated here.

[0189] The above description is based on the preferred embodiments of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the invention is defined by the appended claims rather than the foregoing description, and all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0190] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for collecting clinical rehabilitation data for brain diseases, characterized in that, Includes the following steps: S1. Real-time acquisition of raw data including exercise / physiological data, subjective assessments, and medication records, and the addition of timestamps and contextual labels to construct a spatiotemporally correlated dataset; S2. Perform noise reduction and smoothing on the raw data, mark the end of the potential drug period and screen abnormal segments in rehabilitation training, and quantify the subjective assessment text into symptom levels. S3. Extract specific feature sets for end-of-dose phenomenon and recovery abnormality respectively, and use attention mechanism to achieve context-adaptive fusion of multimodal features; S4. Construct a binary classification deep learning model to distinguish abnormality types, combine clinical annotations to optimize the decision threshold, and generate an anomaly report containing event details.

2. The method for collecting clinical rehabilitation data for brain diseases according to claim 1, characterized in that: In step S1, the motion data includes three-dimensional acceleration, three-dimensional angular velocity, cadence, and gait cycle; the physiological data includes heart rate, electromyographic signals, and skin resistance; the subjective assessment includes a symptom self-rating scale, daily activity ability scores, and subjective feeling descriptions; and the medication record includes the drug name, single dose, time of administration, route of administration, and drug type.

3. The method for collecting clinical rehabilitation data for brain diseases according to claim 1, characterized in that: Step S2 includes the following sub-steps: S21. Wavelet transform algorithm is used to remove high-frequency noise in motion data, and Kalman filter is used to optimize the smoothness of physiological signals. S22. Based on the pharmacokinetic model, mark the end-of-phase period of potential agents and extract the motion feature fragments within this period; S23. Calculate the deviation of motion indicators through a sliding time window to initially screen abnormal segments that deviate significantly from the baseline; S24. Semantically encode the subjective assessment data to map the text descriptions into quantitative symptom levels.

4. The method for collecting clinical rehabilitation data for encephalopathy according to claim 3, characterized in that: In step S22, the drug metabolism kinetic model uses a one-compartment model to describe the drug concentration decay process, as shown in the following formula: ; in, After taking the medicine Blood drug concentration at any given time; This is the initial concentration of the drug; It eliminates the rate constant; It is time; It is a natural constant; Using the aforementioned pharmacokinetic model, the concentration was calculated to decrease to... Time Specifically, regarding the formula Deformation, Solving : ; When marking the end-of-dose period, by the start time and end time Predict and label the end-of-term period of potential agents: ;in, It's time to take the medication.

5. The method for collecting clinical rehabilitation data for encephalopathy according to claim 3, characterized in that: In step S23, the deviation metric window index differs from the baseline; the greater the deviation, the higher the probability of an anomaly. A standardized deviation formula is used: ; in, It is the degree of deviation; It is the motion index value of the current window; It is the baseline indicator mean; It is the standard deviation of the baseline indicator.

6. The method for collecting clinical rehabilitation data for encephalopathy according to claim 1, characterized in that: Step S3 includes the following sub-steps: S31. For the end-of-dose phenomenon, calculate the interval between the abnormal segment and the most recent medication time, extract the slope of the curve of duration after medication versus symptom intensity, and extract the power spectral density of the tremor signal and the motion amplitude attenuation rate in the rigid state. S32. For abnormal points in normal rehabilitation, calculate the interval between the abnormal segment and the start / end time of training, extract the correlation coefficient between training intensity and movement error rate, and extract the correlation features between environmental parameters and gait stability based on context labels. S33. An attention mechanism is used to fuse multimodal features, and dynamic weights are assigned to end-of-dose related features.

7. The method for collecting clinical rehabilitation data for encephalopathy according to claim 6, characterized in that: In step S31, the interval duration is calculated by extracting the most recent medication timestamp from the collected medication records and extracting the start timestamp of the marked abnormal segments. In the slope extraction, symptom intensity data are collected at 5-minute intervals during the marked potential end-of-dose period. Corresponding to the duration after medication, a straight line is fitted using the least squares method. The slope reflects the trend of symptom intensity change with the duration after medication. In the power spectral density extraction, the noise-reduced flutter signal is divided into frames with a frame length of 2 seconds and an overlap rate of 50%. The Welch method is used to calculate the power spectral density of each frame and extract the energy proportion of the 3-7Hz frequency band. In the extraction of motion amplitude decay rate, amplitude data of five consecutive repeated movements are extracted from the abnormal segments in the initial screening, and the decay rate of amplitude with the number of movements is calculated to reflect the rate of decline in motor ability caused by stiffness.

8. A method for collecting clinical rehabilitation data for encephalopathy according to claim 6, characterized in that: In step S32, the interval between the abnormal segment and the training start / end time is used to determine whether the abnormality is directly related to the training process; In the extraction of the correlation coefficient, the Pearson correlation coefficient is used to quantify the association between training load and movement quality and to determine whether the abnormality is caused by the intensity exceeding the patient's tolerance. In the aforementioned feature extraction, the influence of ground or lighting environmental factors on gait abnormalities is analyzed, and the influence weights of environmental parameters on gait stability are calculated through multiple linear regression.

9. A method for collecting clinical rehabilitation data for encephalopathy according to claim 1, characterized in that: Step S4 includes the following sub-steps: S41. Construct a binary classification deep learning model, input the fused multimodal features, and output the classification results of end-of-dose phenomenon / normal recovery abnormality points; The binary classification deep learning model consists of an LSTM layer, a Dropout layer, a fully connected layer, and a Softmax layer. The Softmax layer outputs probabilities, which are used to convert a 2D vector into class probabilities to facilitate the determination of classification confidence. ; in, It is a prediction probability vector; It is the output vector of the fully connected layer; yes The first element of the vector; yes The second element of the vector; It is a natural constant; S42. Using clinically labeled data, supervised learning is performed during model training, and the threshold is optimized using F1-score. S43. If it is determined to be the end-of-dose phenomenon, simultaneously verify whether the period is in the potential end-of-dose period and whether the symptom level increases with time. If it is determined to be a normal rehabilitation abnormal point, confirm that it is significantly correlated with the training intensity. S44. Generate an anomaly classification report that includes the event occurrence time, duration, feature similarity score, and judgment criteria.

10. A data collection system based on any one of claims 1-9 for collecting clinical rehabilitation data of brain diseases, characterized in that, include: The data acquisition module is used to acquire raw patient data in real time, including exercise / physiological data, subjective assessment texts, and medication records; The spatiotemporal correlation module is used to add unified timestamps and contextual labels to the original data and to build a standardized dataset containing spatiotemporal correlations. The data preprocessing module is used to perform high-frequency noise filtering and smoothing optimization of physiological signals in motion data, and to label potential end-of-dose periods related to drug metabolism and quantify symptom levels in subjective assessment texts. The feature extraction and fusion module is used to extract specific feature sets for two types of events: end-of-dose phenomenon and abnormal rehabilitation training, and to perform context-adaptive fusion of multimodal features using an attention mechanism. The dual-classification decision module is used to distinguish between end-of-dose phenomena and abnormal rehabilitation training types based on a binary classification deep learning model, and generate anomaly reports containing event details.

Images