A handwriting dynamics cascading ensemble learning screening system

Abstract

This invention discloses a handwritten dynamics cascaded integrated learning screening system, belonging to the field of dynamics screening technology. This handwritten dynamics cascaded integrated learning screening system includes: an acquisition and preprocessing module for preprocessing raw dynamics signal data and risk feature data; a handwritten data quality assessment and grading module for determining acquisition quality based on signal coupling evaluation analysis results, executing an interactive prompting mechanism, and entering the risk grading process; a dynamics index discrimination and prompting module for executing the risk grading process based on sensitivity amplification theory analysis results; an abnormal risk scoring module for training a shallow feedforward neural network model; and a risk feature synthesis and early judgment module for early dynamics anomaly discrimination. This system solves the problem of insufficient sensitivity in feature extraction and discrimination of early and subtle abnormalities in existing handwritten dynamics screening systems, making it difficult to achieve high accuracy and reliability in early disease screening.

Description

Technical Field

[0001] This invention relates to the field of dynamic screening technology, specifically a handwritten dynamic cascade integrated learning screening system. Background Technology

[0002] Currently, the diagnosis and assessment of movement disorders in the medical field mainly rely on professional observation, clinical consultation, and standardized scale scoring. However, traditional methods are highly subjective and dependent on expert experience. In recent years, various smart wearable devices and high-precision motion sensing terminals have been widely used to acquire dynamic temporal signals such as displacement, velocity, acceleration, pressure, and angle in individuals' daily activities. By performing spatiotemporal analysis, statistical modeling, and parameter feature extraction on these multi-channel dynamic data, related technologies and products are continuously promoting the digitalization of movement behavior monitoring, the intelligentization of screening methods, and the personalization of health assessments, becoming an important direction for continuous innovation and development in the field.

[0003] For example, the invention patent with publication number CN114582438A discloses a method for predicting the cardiotoxicity of compound hERG based on multidimensional molecular fingerprints. The method includes performing molecular dynamics simulations on compounds in a cardiotoxicity dataset, analyzing the simulation results to calculate a four-dimensional molecular fingerprint based on the molecular dynamics simulation, further combining traditional one-dimensional, two-dimensional, and three-dimensional molecular fingerprints to construct a multidimensional molecular fingerprint feature set, using a feature selection method to screen the optimal feature subset, and using various machine learning algorithms such as random forest, support vector machine, and gradient boosting to establish a cardiotoxicity prediction model for compound hERG. The model performance is evaluated through cross-validation, achieving efficient and intelligent prediction of the cardiotoxicity of compound hERG.

[0004] For example, the invention patent with publication number CN119889670A discloses a method and system for early intelligent screening and warning of acute stroke based on deep learning. It includes acquiring multi-temporal angiography images using pulsed dynamic imaging, constructing a blood flow feature tensor by combining hemodynamic analysis and blood flow velocity vector field data, inputting the data into a dual-flow feature extraction network to obtain a fused hemodynamic feature map, reconstructing the vascular topology through directional filtering and minimum spanning tree algorithm, generating a vascular lesion risk distribution map using a recurrent neural tensor network, obtaining tissue and brain function parameters by combining multispectral imaging and electroencephalogram analysis, calculating the risk level value through adaptive threshold segmentation and hierarchical decision model, and finally generating a clinical intervention plan, thus realizing intelligent screening and graded warning of acute stroke risk.

[0005] Existing technologies have enabled automated modeling and risk assessment of disease-related dynamic signals, molecular features, and imaging data. However, they still have significant shortcomings in areas such as multimodal signal collaborative processing, high-sensitivity identification of early subtle anomalies, accurate extraction of complex features, and cross-individual and cross-scenario generalization capabilities. Current mainstream methods often suffer from high false negative rates and insufficient generalization capabilities when facing practical needs such as multi-source heterogeneous data fusion, fine-grained pathological change detection, feature denoising, and robustness of discrimination. These shortcomings stem from limitations in feature extraction, model adaptability, or low sensitivity to early and weak signals, making it difficult to meet the higher requirements of early intelligent screening and personalized accurate identification.

[0006] Therefore, in order to address the above problems, there is an urgent need for a handwritten dynamics cascaded integrated learning screening system. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a handwriting dynamics cascaded integrated learning screening system, which solves the problem that existing handwriting dynamics screening systems lack the sensitivity for feature extraction and discrimination of early and subtle abnormalities, making it difficult to achieve accuracy and reliability in early disease screening.

[0009] Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solution: a handwritten dynamics cascaded integrated learning screening system, comprising: an acquisition and preprocessing module for acquiring raw dynamics signal data, obtaining risk feature data, and preprocessing the raw dynamics signal data and risk feature data; a handwritten data quality assessment and grading module for performing signal coupling evaluation analysis on the raw dynamics signal data, determining the acquisition quality based on the signal coupling evaluation analysis results, executing an interactive prompting mechanism, and entering the risk grading process; a dynamics index discrimination and prompting module for performing sensitivity amplification theory analysis on the raw dynamics signal data, and executing the risk grading process based on the sensitivity amplification theory analysis results; an anomaly risk scoring module for training a shallow feedforward neural network model on the raw dynamics signal data and outputting a continuous comprehensive risk score; and a risk feature synthesis and early judgment module for performing covariance fluctuation enhancement discrimination on the raw dynamics signal data and risk feature data, and performing early dynamic anomaly discrimination based on the covariance fluctuation enhancement discrimination.

[0011] Furthermore, the specific process of acquiring raw dynamic signal data and obtaining risk characteristic data is as follows: acquiring raw dynamic signal data, which includes: contact event data, time synchronization data, acceleration time-series signals, displacement multi-channel signals, trajectory point sequence data, velocity data, pressure time-series signals, and tilt angle synchronous acquisition data; and obtaining risk characteristic data, which includes: data from the International Standard Clinical Movement Disorder Rating Scale, physiological age, and movement disorder data of the subject's first-degree relatives.

[0012] Furthermore, the specific preprocessing steps for the original dynamic signal data and risk feature data are as follows: First, the original dynamic signal data undergoes multi-channel denoising and feature correlation analysis using wavelet denoising algorithm and mutual information feature screening method to extract dynamic change-sensitive signals and key micro-anomaly parameters. Second, the original dynamic signal data is denoised and feature-dimension reduced using principal component analysis and multi-scale decomposition algorithm to extract the main spatiotemporal dynamic change components and anomaly distribution trends. Third, the original dynamic signal data is dynamically feature-extracted and its change trends highlighted using sliding window differentiation and feature remapping algorithms. Finally, the original dynamic signal data and risk feature data are standardized and normalized using distribution standardization and linear normalization algorithms.

[0013] Furthermore, the specific process of signal coupling evaluation and analysis of the original dynamic signal data is as follows: Acquire trajectory point sequence data, contact event data, time synchronization data, acceleration time-series signals, pressure time-series signals, and displacement multi-channel signals; For the trajectory point sequence data, use a connectivity analysis algorithm to calculate the effective trajectory length; for the trajectory point sequence data, use a trajectory segment accumulation algorithm based on Euclidean distance to obtain the total trajectory length; calculate the trajectory integrity rate by comparing the effective trajectory length with the total trajectory length; For the pressure time-series signals and contact event data, use a stroke continuity discrimination algorithm to calculate the ratio of the number of continuous stroke segments to the total number of strokes to obtain the continuous stroke ratio; For the pressure time-series signals and time synchronization data, use a non-touch segment duration accumulation algorithm to calculate the ratio of the accumulated non-touch duration to the total acquisition duration. The following steps are performed: First, obtain the percentage of time spent without pen touch. Second, obtain the instantaneous abnormal fluctuation value from the acceleration and pressure time-series signals using short-time Fourier transform and peak energy accumulation algorithms. Third, obtain the environmental noise from the acceleration, pressure, and displacement multi-channel signals using bandpass filtering and statistical noise power estimation algorithms. Fourth, calculate the product of the trajectory integrity rate and the continuous stroke ratio, use it as the input to the sine function, and take its π-fold sine value to obtain the trajectory stroke-related oscillation term. Fifth, calculate the difference between the instantaneous abnormal fluctuation value and the environmental noise, square it, add 1, and take the natural logarithm to obtain the abnormal noise logarithmic adjustment term. Sixth, calculate the square of the percentage of time spent without pen touch to obtain the energy term of time spent without pen touch. Finally, add the trajectory stroke-related oscillation term, the abnormal noise logarithmic adjustment term, and the energy term of time spent without pen touch and take the absolute value to obtain the acquired signal quality value.

[0014] Further, the specific process of determining the acquisition quality based on the signal coupling evaluation and analysis results, executing the interactive prompt mechanism, and entering the risk grading process is as follows: Real-time comparison of the acquired signal quality value and the acquired signal quality threshold, which includes a primary quality threshold and a secondary quality threshold: When the acquired signal quality value is less than or equal to the secondary quality threshold, it is determined to be a safe interval signal; the risk grading process is directly initiated; When the acquired signal quality value is greater than the secondary quality threshold but less than or equal to the primary quality threshold, it is determined to be a warning interval signal; the interactive prompt mechanism is triggered: an acquisition anomaly prompt is pushed to the user's operation interface, and the user can choose to supplement data acquisition or continue entering the risk grading process. If the user's waiting time threshold is exceeded, the current round of acquired data is blocked, a signal quality warning database is created, and the original dynamic signal data is archived to the signal quality warning database; When the acquired signal quality value is greater than or equal to the primary quality threshold, it is determined to be an invalid interval signal; acquisition is terminated, a pop-up window on the interface prompts that acquisition is invalid, the current round of acquired data is blocked, a signal quality anomaly database is created, and the original dynamic signal data is recorded to the signal quality anomaly database.

[0015] Furthermore, the specific process of performing sensitivity amplification theoretical analysis on the original dynamic signal data is as follows: Acquire acceleration time-series signals, velocity data, pressure time-series signals, acquired signal quality values, trajectory point sequence data, and tilt angle synchronous acquisition data; obtain the average jerk from the acceleration time-series signal using a sliding window differentiation algorithm and a averaging algorithm; perform frequency domain analysis on the velocity and pressure time-series signals using Fast Fourier Transform to calculate the energy sum of a specified high-frequency band and obtain the micro-vibration spectrum energy value; obtain the average velocity from the trajectory point sequence data using displacement differentiation and sampling time-series algorithms and a averaging algorithm; obtain the velocity-pressure coupling coefficient from the velocity and pressure time-series signals using mutual information analysis; obtain the mean trajectory curvature from the trajectory point sequence data using differential geometry algorithms and a averaging algorithm; obtain the velocity-tilt angle coupling coefficient from the velocity and tilt angle synchronous acquisition data using a mutual information algorithm; obtain the average acceleration from the acceleration time-series signal using a averaging algorithm; and extract the spectrum from the acceleration time-series signal using Fast Fourier Transform. The peak value of the main frequency component with the largest amplitude is obtained. The spatial distance between the first and last points and the deviation from the ideal closed trajectory are calculated for the trajectory point sequence data. The trajectory closure error value is obtained by the shortest closed loop matching method. The product of the average jerk and the energy value of the micro-vibration spectrum is calculated to obtain the dynamic amplitude modulation term. The product of the average velocity and the velocity pressure coupling coefficient is subtracted from the product of the mean trajectory curvature and the velocity tilt angle coupling coefficient. The square of the result is added to 1, and the natural logarithm is taken to obtain the velocity coupling curvature linkage logarithm term. The π times sine of the acquired signal quality value is calculated to obtain the signal period fluctuation term. The dynamic amplitude modulation term, the velocity coupling curvature linkage logarithm term, and the signal period fluctuation term are added to obtain the dynamic characteristic comprehensive numerator term. The absolute value of the difference between the average acceleration and the peak value of the main frequency is added to the square of the trajectory closure error value, and the square root is added to 1 to obtain the abnormal characteristic modulation denominator term. The dynamic characteristic comprehensive numerator term is divided by the abnormal characteristic modulation denominator term, and the absolute value is taken to obtain the dynamic sensitivity value.

[0016] Furthermore, the specific process of executing the risk grading procedure based on the sensitivity amplification theory analysis results is as follows: Real-time comparison of kinetic sensitivity values ​​and kinetic sensitivity thresholds, including primary and secondary sensitivity thresholds, is performed. The risk grading procedure is as follows: When the kinetic sensitivity value is greater than or equal to the primary sensitivity threshold, it is determined to be a significant kinetic anomaly; a kinetic anomaly alert is pushed to the user interface, and an assessment report and a report of suspicious characteristic parameters are output to the user. When the kinetic sensitivity value is greater than or equal to the secondary sensitivity threshold but less than the primary sensitivity threshold, it is determined to be an early, weak kinetic anomaly; the original kinetic signal data is pushed to the anomaly risk scoring module. When the kinetic sensitivity value is less than the secondary sensitivity threshold, it is determined that no suspicious kinetic anomaly was detected, and a pop-up window suggests continuing daily health management.

[0017] Further, the specific process of training a shallow feedforward neural network model on the original dynamic signal data to output a continuous comprehensive risk score is as follows: A multi-dimensional input feature vector composed of the acquired signal quality value, dynamic sensitivity value, average velocity, average acceleration, average jerk, micro-vibration spectrum energy value, peak frequency, mean trajectory curvature, trajectory closure error value, velocity-pressure coupling coefficient, and velocity-tilt coupling coefficient is used as training samples. Based on the training samples, the Z-score standardization method is used to map features of different dimensions and ranges to the standard normal distribution interval. A shallow feedforward neural network model is constructed using the standardized training samples. The shallow feedforward neural network contains one or two hidden layers, each using a linear rectified unit as the activation function. The output layer uses a sigmoid activation function to compress the output value of the shallow feedforward neural network model into a unified standard interval. Based on historically judged normal and abnormal dynamic samples, supervised learning is carried out. The neural network parameters are optimized by minimizing the binary cross-entropy loss function to complete the training of the shallow feedforward neural network model. The shallow feedforward neural network outputs a continuous comprehensive risk score. The value ranges from 0 to 1.

[0018] Furthermore, the specific process for covariance fluctuation enhancement discrimination of the original dynamic signal data and risk characteristic data is as follows: Obtain the comprehensive risk score, data from the International Standard Clinical Movement Disorder Rating Scale (ICSR), physiological age, and movement disorder data of the subject's first-degree relatives; for the movement disorder data of the subject's first-degree relatives, obtain the family history risk value through a binary mapping algorithm; based on the data from the ICSR, perform speed calculation algorithms, stroke total length geometric algorithms, and trajectory curvature variance algorithms on the trajectory point sequence data to obtain writing speed, handwriting size, and line smoothness. Regarding writing speed... The handwriting size and line smoothness were quantitatively assessed separately, and the scores of the scoring scale were obtained through a standardized weighted summation algorithm. The covariance of the comprehensive risk score and the scoring scale score was calculated, multiplied by the physiological age plus 1, and then multiplied by the standard deviation of the product sequence of the family history risk value and the comprehensive risk score to obtain the risk dynamic enhancement term. The product of the scoring scale score and the physiological age was calculated, the square of the comprehensive risk score was subtracted, and the absolute value was taken to obtain the feature joint suppression term. The risk dynamic enhancement term was subtracted from the feature joint suppression term, and the absolute value was taken to obtain the comprehensive risk feature discrimination value.

[0019] Furthermore, the specific process for early dynamic anomaly identification based on covariance fluctuation enhancement discrimination is as follows: real-time comparison of the comprehensive discrimination value of risk characteristics and the comprehensive discrimination threshold of risk characteristics: when the comprehensive discrimination value of risk characteristics is less than the comprehensive discrimination threshold of risk characteristics, it is determined that no early dynamic anomaly has been detected, and no anomaly is prompted to the user; a pop-up window suggests continuing daily health management. When the comprehensive discrimination value of risk characteristics is greater than or equal to the comprehensive discrimination threshold of risk characteristics, it is determined that an early dynamic anomaly exists. The early dynamic anomaly prompt is pushed to the user's operation interface, and an assessment report and a report of suspicious characteristic parameters are output to the user.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) This invention, through the dynamic signal quality screening and hierarchical management mechanism, can realize the real-time identification and processing of abnormal signals during the acquisition stage, thereby achieving the effect of controllable signal acquisition quality throughout the entire process, and effectively solving the problem that data acquisition anomalies cannot be classified and intervened in a timely manner in the prior art.

[0023] (2) This invention improves the sensitivity and recognition rate of early and weak motor disorders by using a multi-level and multi-channel dynamic characteristic grading method, thereby achieving a significant improvement in the sensitivity of early disease screening and effectively solving the problem of difficulty in accurately identifying early abnormalities in the prior art.

[0024] (3) This invention, by innovatively introducing an integrated intelligent discrimination model, realizes the comprehensive analysis and intelligent scoring of multi-feature and multi-dimensional risk factors, thereby achieving the effect of continuity and automation in disease screening and discrimination, and effectively solves the limitations of single model and discrete output in the existing technology.

[0025] (4) This invention effectively integrates dynamic performance and medical risk factors through synergistic analysis with clinical risk characteristics, thereby achieving a comprehensive risk assessment effect of multi-source information combination, and effectively solving the problem that multiple types of risk characteristics cannot be synergistically analyzed in the prior art.

[0026] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0027] Figure 1 This is a structural diagram of a handwritten dynamics cascaded integrated learning screening system according to the present invention;

[0028] Figure 2 This is a line graph showing the distribution of the dynamic characteristic parameters of the present invention;

[0029] Figure 3This is a schematic diagram of the shallow feedforward neural network structure of the present invention. Detailed Implementation

[0030] 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.

[0031] Please see Figures 1-3 This invention provides a technical solution: a handwritten dynamics cascaded integrated learning screening system, comprising: an acquisition and preprocessing module for acquiring raw dynamics signal data, obtaining risk feature data, and preprocessing the raw dynamics signal data and risk feature data; a handwritten data quality assessment and grading module for performing signal coupling evaluation analysis on the raw dynamics signal data, determining the acquisition quality based on the signal coupling evaluation analysis results, executing an interactive prompting mechanism, and entering a risk grading process; a dynamics index discrimination and prompting module for performing sensitivity amplification theory analysis on the raw dynamics signal data, and executing a risk grading process based on the sensitivity amplification theory analysis results; an abnormal risk scoring module for training a shallow feedforward neural network model on the raw dynamics signal data and outputting a continuous comprehensive risk score; and a risk feature synthesis and early judgment module for performing covariance fluctuation enhancement discrimination on the raw dynamics signal data and risk feature data, and performing early dynamics anomaly discrimination based on the covariance fluctuation enhancement discrimination.

[0032] Specifically, the process of acquiring raw dynamic signal data and obtaining risk characteristic data is as follows: raw dynamic signal data is acquired, including: touch event data, time synchronization data, acceleration time-series signals, displacement multi-channel signals, trajectory point sequence data, velocity data, pressure time-series signals, and tilt angle synchronous acquisition data. The handwriting dynamic acquisition terminal integrates the acquisition, which integrates a high-precision capacitive touch panel, MEMS acceleration and angular velocity sensors, a high-sensitivity pressure sensor, and a multi-channel synchronous acquisition and edge processing unit. It supports a sampling frequency of not less than 100Hz and a time synchronization error of not more than 1ms, realizing the synchronous acquisition and real-time reporting of multiple parameters such as handwriting, gestures, pressure, and posture.

[0033] Risk characteristic data is obtained, including: data from the International Standard Clinical Movement Disorder Rating Scale, physiological age, and movement disorder data of the subject's first-degree relatives. Through scale assessment input, it is ensured that all scale data, physiological age, and family history information can be associated with and archived with the dynamic signals according to a unified test number.

[0034] This implementation plan utilizes a dedicated handwritten dynamics acquisition terminal and customized assessment to achieve centralized collection and standardized archiving of dynamics signals and risk characteristic data, effectively improving the comprehensiveness and consistency of data acquisition. The dynamics terminal can simultaneously acquire key signals and ensure acquisition frequency and time accuracy, maximizing the reproduction of the subject's true motion characteristics. Risk characteristic information is entered through a standardized process, ensuring a unique identifier association with the dynamics data. This step provides a high-quality, structured input data foundation for subsequent data processing, feature analysis, and intelligent screening, enhancing scientific rigor, standardization, and reliability.

[0035] Specifically, wavelet denoising algorithm and mutual information feature screening method are used to perform multi-channel denoising and feature correlation analysis on the original dynamic signal data, extracting dynamic change sensitive signals and key micro-anomaly parameters, effectively suppressing environmental noise and occasional interference in different channels, and improving the accuracy and robustness of subsequent feature analysis. Principal component analysis and multi-scale decomposition algorithm are used to denoise and reduce the feature dimension of the original dynamic signal data, extracting the main spatiotemporal dynamic change components and anomaly distribution trends, significantly reducing the redundancy of high-dimensional original data, and retaining the core feature components that can characterize early anomalies. Sliding window differentiation and feature remapping algorithm are used to extract dynamic features and highlight change trends in the original dynamic signal data, realizing sensitive detection and visualization of micro-motion anomalies and non-stationary changes. Distribution standardization and linear normalization algorithm are used to standardize and normalize the original dynamic signal data and risk feature data, ensuring the comparability of data from different collection sources, dimensions and distributions, and providing a unified data foundation for subsequent modeling analysis and risk assessment.

[0036] In this implementation plan, a variety of signal processing and feature engineering methods were used to systematically preprocess and optimize the collected dynamic signal data and risk feature data, significantly improving the effectiveness, comparability, and anomaly sensitivity of the data. After denoising, dimensionality reduction, feature extraction, and standardization normalization, the key information reflecting motion anomalies was fully preserved, while redundancy and noise interference were greatly reduced. This provides a high-quality, structured data foundation for subsequent anomaly screening and risk assessment, further enhancing the scientific rigor and accuracy of the overall analysis.

[0037] Specifically, the process of signal coupling evaluation and analysis of the raw dynamic signal data is as follows: Track point sequence data, contact event data, time synchronization data, acceleration time-series signals, pressure time-series signals, and displacement multi-channel signals are acquired. All data undergoes unified acquisition numbering and time base calibration to ensure consistency and traceability in subsequent processing. The effective trajectory length of the track point sequence data is statistically analyzed using a connectivity analysis algorithm, effectively reflecting the continuity and trajectory characteristics during motion. The total trajectory length is obtained from the track point sequence data using a trajectory segment accumulation algorithm based on Euclidean distance, comprehensively covering all stroke segments and the total amount of motion paths. The ratio of the effective trajectory length to the total trajectory length is calculated to obtain the trajectory integrity rate, used to measure the integrity of handwriting acquisition and abnormal interruption situations. The pressure time-series signals and contact event data are analyzed... The data is analyzed using a stroke continuity discrimination algorithm to calculate the ratio of the number of consecutive stroke segments to the total number of strokes, resulting in a consecutive stroke ratio that reflects the stability and continuity of the subject's writing process. For pressure timing signals and time synchronization data, an algorithm for accumulating non-pen-touching segment duration is used to calculate the ratio of the accumulated non-pen-touching time to the total acquisition time, resulting in a non-pen-touching time percentage that reflects interruptions and pauses during the writing process. For acceleration and pressure timing signals, short-time Fourier transform and energy peak accumulation algorithms are used to obtain instantaneous abnormal fluctuation values, which are used to identify sudden high-frequency abnormal changes and jitter characteristics during writing. For acceleration, pressure, and displacement multi-channel signals, bandpass filtering and statistical noise power estimation algorithms are used to obtain environmental noise, separating the real signal from external interference components and enhancing feature robustness.

[0038] The product of trajectory integrity rate and continuous stroke ratio is used as the input of the sine function. Taking its sine value multiplied by π yields the trajectory-stroke correlation oscillation term. Utilizing the nonlinear mapping characteristic of the sine function over the input interval [0,1], the product of trajectory integrity rate and continuous stroke ratio is transformed into an output with periodicity and amplification sensitivity. This helps enhance the ability to distinguish changes in the coupling relationship between trajectory and stroke, highlighting the response to anomalies. The difference between instantaneous anomaly fluctuation value and environmental noise is calculated, squared, added to 1, and then the natural logarithm is taken to obtain the anomaly noise logarithmic adjustment term, significantly improving the sensitivity to sudden anomalies coupled with changes in background noise. The square of the proportion of time without pen touch is calculated to obtain the non-pen touch time energy term, further amplifying the pause anomaly information during the writing process. The trajectory-stroke correlation oscillation term, the anomaly noise logarithmic adjustment term, and the non-pen touch time energy term are added together and their absolute values ​​are taken to obtain the acquired signal quality value. The specific calculation formula is as follows:

[0039] ;

[0040] In the formula, This indicates the quality value of the acquired signal and comprehensively evaluates the overall effectiveness of the signal. This represents the trajectory completeness rate, reflecting the continuity and completeness of the trajectory. It indicates the ratio of consecutive strokes, reflecting the stability of continuous stroke writing; This indicates the percentage of time the pen was not touched, measuring the proportion of pauses and invalid data collection. It represents instantaneous abnormal fluctuation values, characterizing sudden abnormal changes in the signal; This represents environmental noise and is used to quantify the level of interference during the data acquisition process.

[0041] This implementation scheme effectively enhances the scientific rigor and sensitivity of signal integrity and validity assessment through comprehensive feature analysis and coupled evaluation of multidimensional dynamic signals. Compared to traditional single-index and single-channel judgment methods, this approach can fully capture subtle anomalies and complex fluctuations in the signal, significantly improving the ability to identify anomalous data, noise interference, and missing data. Simultaneously, the quantified signal quality evaluation results provide a more solid and reliable data foundation for subsequent data screening, anomaly re-acquisition, and risk analysis, contributing to improved automated processing capabilities and intelligent judgment levels.

[0042] Specifically, the process of determining the acquisition quality based on the signal coupling evaluation and analysis results, executing the interactive prompt mechanism, and entering the risk grading process is as follows:

[0043] The system compares the quality values ​​of the acquired signals with the quality thresholds in real time. The quality thresholds include primary and secondary quality thresholds, enabling hierarchical dynamic judgment and closed-loop management of the quality of the acquired data.

[0044] When the quality value of the acquired signal is less than or equal to the secondary quality threshold, it is determined to be a signal in the safe range, indicating that the acquired data is of high quality and has strong reliability, and can be directly used in subsequent analysis; it can directly enter the risk classification process to ensure the validity and high reliability of the analysis input data.

[0045] When the acquired signal quality value is greater than the secondary quality threshold but less than or equal to the primary quality threshold, it is determined to be a warning interval signal, indicating that the current data quality has a certain risk but is still usable. An interactive prompt mechanism is triggered: an abnormal acquisition prompt is pushed to the user interface, guiding the user to pay attention to the signal quality and make their own decision on whether to reacquire data. The user can choose to reacquire data or continue into the risk grading process, enhancing the flexibility and data assurance of the acquisition process. If the user's waiting time threshold is exceeded, the data acquired in this round is blocked to prevent low-quality data from affecting subsequent analysis results. A signal quality warning database is created, and the original dynamic signal data is archived in the signal quality warning database to achieve traceable management and subsequent optimization analysis.

[0046] When the acquired signal quality value is greater than or equal to the first-level quality threshold, it is determined to be an invalid interval signal, indicating that the acquired data is seriously abnormal and distorted and cannot be used for analysis; the acquisition is terminated, and a pop-up window on the interface prompts that the acquisition is invalid, informing the user of the reason for the acquisition failure immediately, blocking the data acquired in this round, ensuring that the input data is always within a controllable range, creating a signal quality anomaly database and recording the original dynamic signal data into the signal quality anomaly database, which is convenient for subsequent tracking, review and quality improvement.

[0047] In this implementation plan, this step achieves intelligent grading and process control of the acquired signal quality through hierarchical threshold discrimination and dynamic closed-loop management. It can identify the reliability and risk level of data based on the acquired signal quality values, and accordingly take appropriate data flow, prompting interaction, and archiving measures to ensure that only high-quality, highly reliable data enters the subsequent analysis process. For data with risk anomalies, timely prompts are provided, flexible processing is implemented, and a hierarchical database is established, effectively preventing invalid and low-quality data from affecting the overall analysis results. This significantly improves the standardization, stability, and traceability of data processing, providing a solid and reliable signal quality guarantee for subsequent risk identification and intelligent analysis.

[0048] Specifically, the process of performing sensitivity amplification theoretical analysis on the original dynamic signal data is as follows: Acceleration time-series signals, velocity data, pressure time-series signals, acquired signal quality values, trajectory point sequence data, and tilt angle synchronously acquired data are obtained. These multi-source data collectively reflect the dynamic process and subtle changes in the subject's movement behavior. The acceleration time-series signals are analyzed using a sliding window differentiation algorithm and a averaging algorithm to obtain the average jerk, revealing the intensity and trend of short-term movement rate changes. Velocity data and pressure time-series signals are analyzed in the frequency domain using Fast Fourier Transform to calculate the energy sum of specified high-frequency segments, obtaining the micro-tremor spectrum energy value, which is used to sensitively capture high-frequency jitter and subtle abnormal tremor characteristics. The trajectory point sequence data are analyzed using displacement differentiation and sampling time-series algorithms and averaging algorithms to obtain the average velocity, characterizing the overall movement rhythm and writing fluency. The velocity-pressure coupling coefficient is obtained from the velocity data and pressure time-series signals using mutual information analysis, reflecting the correlation strength between velocity changes and synchronous fluctuations in hand pressure. It is an indicator of comprehensive motor control and neural regulation capabilities; the mean curvature of the trajectory is obtained from the trajectory point sequence data through differential geometry and averaging algorithms, which characterizes the curvature complexity of the handwriting trajectory and is helpful for identifying abnormal writing patterns; the velocity-tilt coupling coefficient is obtained from the velocity data and tilt angle synchronous acquisition data through mutual information algorithms, which is used to measure the synergistic relationship between movement speed and hand posture changes; the average acceleration is obtained from the acceleration time sequence signal through averaging algorithms, which reflects the intensity of the overall motion state; the spectrum is extracted from the acceleration time sequence signal through fast Fourier transform, and the peak value of the main frequency component with the largest amplitude is obtained, which reflects the typical periodic characteristics of motion and helps to identify abnormal rhythms; the spatial distance between the first and last points and the deviation from the ideal closed trajectory are calculated from the trajectory point sequence data, and the trajectory closure error value is obtained through the shortest closed loop matching method, which measures the closure accuracy and spatial consistency of the writing action.

[0049] The product of average jerk and micro-vibration spectrum energy value is calculated to obtain the dynamic amplitude modulation term, highlighting the sensitivity to the coupling between drastic acceleration changes and high-frequency micro-vibrations. The product of average velocity and velocity-pressure coupling coefficient, minus the product of mean trajectory curvature and velocity tilt angle coupling coefficient, squared, incremented by 1, and then the natural logarithm is taken to obtain the velocity coupling curvature linkage logarithmic term. This composite term amplifies the linkage between abnormal motion control and spatial deformation through nonlinear coupling, providing differentiated highlighting capabilities for potential anomalies. The π-fold sine of the acquired signal quality value is calculated to obtain the signal periodic fluctuation term. Introducing a periodic function mapping can enhance the response and identification of the rhythmicity of signal fluctuations. It combines nonlinearity and sensitivity amplification effects; by adding the dynamic amplitude modulation term, the velocity coupling curvature linkage logarithmic term, and the signal periodic fluctuation term, a comprehensive numerator of dynamic characteristics is obtained, realizing the fusion expression of multiple dynamic anomalies and enhancing the anomaly aggregation effect; the absolute value of the difference between the average acceleration and the peak frequency is added to the square of the trajectory closure error value, and the square root of the result is added to 1 to obtain the anomaly characteristic modulation denominator, which helps to dynamically adjust the denominator weight and smooth the impact of extreme outliers on the overall sensitivity; the dynamic sensitivity value is obtained by dividing the comprehensive numerator of dynamic characteristics by the anomaly characteristic modulation denominator and finally taking the absolute value. The specific calculation formula is as follows:

[0050] ;

[0051] In the formula, This represents the dynamic sensitivity value, used to quantitatively reflect the overall sensitivity to motion abnormalities in the subject. It represents the average jerk, reflecting the degree of drastic change in motion speed over a short period of time; It represents the energy value of the micro-vibration spectrum, characterizing the intensity of high-frequency fine jitter and tremor; It represents the average speed and is used to measure the fluency of the overall writing motion. This represents the velocity-pressure coupling coefficient, reflecting the synchronous relationship between velocity and pressure changes; It represents the mean curvature of the trajectory, describing the curvature complexity of the motion trajectory; This represents the velocity tilt angle coupling coefficient, which measures the correlation between velocity and changes in hand posture. This indicates the quality value of the acquired signal, comprehensively reflecting the completeness and validity of the acquired data; It represents the average acceleration, depicting the intensity and dynamism of the overall motion; It represents the peak frequency and identifies the dominant rhythmic features in writing; It represents the trajectory closure error value, reflecting the accuracy and spatial consistency of the trajectory closure.

[0052] In this embodiment, Table 1 is a comparison table of handwriting dynamic parameters, which records in detail the normalized data of key dynamic characteristic parameters such as average jerk, micro-vibration spectrum energy, average velocity, and mean trajectory curvature for different object numbers. This data is used to quantify the multidimensional dynamic performance of different objects during the handwriting process. The average jerk and micro-vibration spectrum energy parameters for each object are the results of normalization algorithm processing. The acquired signal quality value, average acceleration, peak frequency, and trajectory closure error parameters not mentioned in the table have been uniformly fixed at 0.40 to eliminate the influence of fluctuations in other variables and ensure the comparability of the main dynamic characteristics in a horizontal comparison. Taking object number 1 as an example, the average jerk is 0.32, the micro-vibration spectrum energy is 0.41, the average velocity is 0.20, and the mean trajectory curvature is 0.13. Object number 3 shows outstanding performance in the micro-vibration spectrum energy parameter, indicating that it has more significant dynamic fluctuation characteristics during the handwriting process.

[0053] Table 1 Comparison of Handwritten Dynamics Parameters

[0054]

[0055] like Figure 2 The figure shows a line graph illustrating the distribution of the dynamic characteristic parameters of this invention, visually demonstrating the changing trends of each dynamic parameter across different objects. As can be seen from the graph, the micro-vibration spectrum energy is highest in object number 3, reflecting significant fluctuations in the dynamic signal during handwriting. In contrast, object number 4 shows generally lower values ​​for each parameter, indicating more stable dynamic performance during handwriting. The remaining parameters show significant differences in distribution among different objects, indicating a certain degree of heterogeneity in the dynamic characteristics of individuals. Since the acquired signal quality value, average acceleration, peak frequency, and trajectory closure error are not shown in this figure and have been uniformly fixed at 0.40, the line graph trend more prominently reflects the relative distribution characteristics of the main dynamic parameters among different objects, providing a data foundation for subsequent object difference analysis and risk identification.

[0056] In this implementation plan, this step achieves comprehensive quantification and acute discrimination of abnormal movement manifestations in subjects through multi-dimensional dynamic feature extraction and sensitivity amplification theory analysis. It systematically integrates multi-source data such as acceleration, velocity, pressure, and trajectory, and through composite algorithms and differentiated index design, it can amplify minute abnormal fluctuations and reveal the coupling relationships of various complex movement characteristics, effectively improving the detection capability of early movement disorders and abnormal states. The analysis results not only provide a scientific basis for comparing differences in dynamic characteristics between samples but also ensure the comparability and stability of various features under normalization processing. Ultimately, this step provides solid and reliable multi-parameter data support and sensitivity assurance for subsequent risk grading, intelligent discrimination, and movement disorder screening, improving overall detection efficiency and practical value.

[0057] Specifically, the risk grading process based on the analysis results of the sensitivity amplification theory is as follows: real-time comparison of dynamic sensitivity values ​​and dynamic sensitivity thresholds, including primary and secondary sensitivity thresholds, to achieve graded identification and process-oriented management of motion anomaly risks.

[0058] When the dynamic sensitivity value is greater than or equal to the first-level sensitivity threshold, it is judged as a significant dynamic abnormality, indicating that the current subject has obvious movement abnormalities and risk of impairment. The dynamic abnormality prompt is pushed to the user operation interface to promptly remind the user to pay attention to the abnormal situation. An assessment report and a report of suspicious characteristic parameters are output to the user, providing detailed data support for further clinical assessment and early intervention.

[0059] When the kinetic sensitivity value is greater than or equal to the secondary sensitivity threshold and less than the primary sensitivity threshold, it is judged as an early weak kinetic abnormality, indicating that the subject has slight abnormal fluctuations and early risk signals. The original kinetic signal data is pushed to the abnormal risk scoring module for further multi-parameter risk fusion analysis and accurate assessment.

[0060] When the kinetic sensitivity value is less than the secondary sensitivity threshold, it is determined that no suspicious kinetic abnormality has been detected, indicating that the subject's current kinetic state is within the normal range. A pop-up window suggests continuing daily health management, providing users with positive and healthy lifestyle guidance and avoiding unnecessary anxiety and intervention.

[0061] This implementation plan establishes a scientific and detailed risk grading system for motion abnormalities by setting dual sensitivity thresholds. Based on real-time calculated motion sensitivity values, it categorizes subjects into three states: significant abnormalities, early and weak abnormalities, and normal status, and intelligently pushes different assessment and management measures accordingly. For users with significant abnormalities, clear prompts and detailed reports are provided immediately, helping them to pay attention to their health as early as possible and obtain targeted clinical intervention suggestions. For early and weak abnormalities, a more in-depth risk fusion analysis process is initiated to achieve accurate identification and dynamic tracking of subclinical states or hidden risks. For users with normal motion performance, health management suggestions are proactively provided via pop-up windows, strengthening health awareness while reducing unnecessary psychological burden and misjudgment interventions. This mechanism greatly improves the sensitivity, targeting, and scientific rigor of motion abnormality detection, making the entire risk screening and health service process more intelligent, closed-loop, and personalized, providing efficient, accurate, and full-cycle support for users at different risk levels.

[0062] Specifically, the process of training a shallow feedforward neural network model on the original dynamic signal data to output a continuous comprehensive risk score is as follows: A multi-dimensional input feature vector is constructed from the acquired signal quality value, dynamic sensitivity value, average velocity, average acceleration, average jerk, micro-vibration spectrum energy value, peak frequency, mean trajectory curvature, trajectory closure error, velocity-pressure coupling coefficient, and velocity-tilt coupling coefficient as training samples. These features originate from upstream multi-algorithm fusion and in-depth analysis of dynamic behavior, fully reflecting the multi-dimensional heterogeneity and potential risks of the examinee's motion performance. Based on the training samples, the Z-score standardization method is used to map features of different dimensions and ranges to the standard normal distribution interval, achieving data scale uniformity when multiple features are input collaboratively, thus improving the stability and generalization ability of the model training. The standardized training samples are then used to construct a shallow feedforward neural network model. This model contains one or two hidden layers. Each hidden layer, through collaborative design with upstream feature construction, can achieve effective combination and nonlinear mapping between features, balancing discriminative ability and model interpretability. Each hidden layer uses a linear rectified unit as the activation function, and the output layer uses a sigmoid activation function. This compresses the output values ​​of the shallow feedforward neural network model into a unified standard range, ensuring that the continuous risk score remains within the [0,1] range, facilitating seamless integration with medical diagnostic standards. Supervised learning is conducted based on historical normal and abnormal dynamic samples. The neural network parameters are optimized by minimizing the binary cross-entropy loss function. This training mechanism effectively utilizes known labels to strengthen the model's discriminative boundaries and effectively suppresses overfitting. Ultimately, the shallow feedforward neural network model outputs a continuous comprehensive risk score, ranging from 0 to 1. Figure 3 The diagram shows the principle of a shallow feedforward neural network structure. It takes 11-dimensional upstream dynamics and risk features as input. First, it achieves data scale unification through Z-score standardization. Then, it completes the nonlinear mapping of high-order features through 1 to 2 layers of ReLU activated hidden layers. Finally, the Sigmoid output layer generates a continuous risk score with a value range of [0, 1]. This realizes the hierarchical fusion of multi-parameter dynamic features and intelligent risk assessment.

[0063] This implementation scheme achieves accurate, hierarchical, and continuous intelligent discrimination of motion abnormality risk through efficient collaboration between a shallow feedforward neural network and upstream multi-feature construction. Relying on the full expression and standardized mapping of multi-dimensional input features, it effectively integrates key dynamic information from different sources and scales, enhancing the sensitivity to potential motion disorders and subtle abnormalities while ensuring model interpretability and stability. The shallow structure combined with meticulous feature design achieves screening sensitivity and early warning effects approaching those of deep networks, even under conditions of limited model complexity. The final output of a continuous comprehensive risk score provides a scientific and intuitive basis for clinical risk assessment and personalized health management, significantly enhancing its practicality and reliability in large-scale, early screening scenarios.

[0064] Specifically, the process of enhancing covariance fluctuation discrimination for raw dynamic signal data and risk characteristic data is as follows: A comprehensive risk score, data from the International Standard Clinical Movement Disorder Rating Scale (ICSR), physiological age, and movement disorder data of the examinee's first-degree relatives are obtained. These multi-source data collectively cover individual risk, genetic background, and motor behavior performance. For the movement disorder data of the examinee's first-degree relatives, a binary mapping algorithm is used to obtain a family history risk value, transforming complex family history information into quantifiable risk indicators for subsequent fusion analysis with other features. Based on the ICSR data, speed calculation algorithms, stroke length geometric algorithms, and trajectory curvature variance algorithms are used on the trajectory point sequence data to obtain writing speed, handwriting size, and line fluency. This achieves linkage verification between the clinical rating scale and objective movement data. Writing speed, handwriting size, and line fluency are quantitatively assessed item by item, and a standardized weighted summation algorithm is used to obtain the scoring scale score, ensuring a balanced contribution of different item features and improving the scientific rigor of the overall score.

[0065] The covariance of the comprehensive risk score and the rating scale score is calculated and multiplied by the physiological age plus 1. Leveraging the amplifying effect of age on risk covariance, this more acutely captures age-related risk trends. This is then multiplied by the standard deviation of the product sequence of family history risk value and comprehensive risk score to further integrate the volatility of genetic risk with individual risk variation, achieving dynamic enhancement of multiple risk factors and yielding a dynamic risk enhancement term. The product of the rating scale score and the physiological age is calculated, and the square of the comprehensive risk score is subtracted. The absolute value is then used to smooth extreme abnormalities and unreasonable outcomes through joint feature suppression, yielding a joint feature suppression term. Finally, the joint feature suppression term is subtracted from the dynamic risk enhancement term, and the absolute value is taken to obtain the comprehensive risk feature discriminant value. The specific calculation formula is as follows:

[0066]

[0067] In the formula, This represents the comprehensive risk characteristic discrimination value, which measures the overall risk level under the integration of multiple factors; This represents the overall risk score, reflecting the overall risk of the current dynamic anomaly; The score on the rating scale reflects the overall risk of the current dynamic anomaly. Indicates chronological age, adjusted for the impact of age on risk; This indicates a family history risk score, reflecting genetically related risk factors. It represents the standard deviation of the product sequence, integrating the volatility of genetic risk with individual risk variation.

[0068] This implementation plan integrates multidimensional information such as comprehensive risk scores, scale assessments, age, and family history. Utilizing covariance amplification and joint suppression mechanisms, it dynamically and sensitively assesses the overall risk of examinees. This not only accurately characterizes an individual's multi-source risk status but also highlights the actual impact of age and key genetic factors on risk trends, effectively improving the early identification capability of high-risk individuals. The resulting comprehensive risk characteristic discrimination value provides more scientific and reliable data support for precise classification, personalized intervention, and long-term health management, significantly enhancing the early warning effect and practical application value.

[0069] Specifically, the process of identifying early dynamic anomalies based on the covariance fluctuation enhancement discrimination is as follows: Real-time comparison of the comprehensive discrimination value and the comprehensive discrimination threshold of risk characteristics to achieve final intelligent discrimination and graded alerts for early dynamic anomalies.

[0070] When the comprehensive risk characteristic discrimination value is less than the comprehensive risk characteristic discrimination threshold, it is determined that no early dynamic abnormality was detected, indicating that the subject's current risk status is good and there are no obvious abnormal manifestations. No abnormality is indicated to the user, and a pop-up window suggests continuing daily health management and encourages the user to maintain good living habits to achieve normalized self-monitoring of health risks.

[0071] When the comprehensive risk characteristic discrimination value is greater than or equal to the comprehensive risk characteristic discrimination threshold, it is determined that there is an early dynamic anomaly, indicating that potential risk signals and early abnormal manifestations have been identified. The early dynamic anomaly prompt is pushed to the user's operation interface to remind the user to pay attention to their health status in a timely manner. An assessment report and a report of suspicious characteristic parameters are output to the user to provide detailed decision-making basis for subsequent professional assessment and personalized intervention.

[0072] This implementation plan achieves accurate identification and tiered feedback for early motor abnormalities by dynamically comparing the comprehensive risk characteristic discriminant value with the threshold. It can identify the subject's current comprehensive risk status, effectively screening out individuals without abnormalities, reducing false alarms and over-intervention, while also issuing immediate warnings to users at risk of early abnormalities and pushing detailed assessment reports and key parameters to help users achieve self-health management and early intervention. Overall, this mechanism significantly improves the sensitivity and response efficiency to early abnormalities, providing scientific and reliable decision support for intelligent and personalized movement disorder screening and tiered health management.

[0073] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0074] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A handwritten dynamics cascaded integrated learning screening system, characterized in that, include: The acquisition and preprocessing module is used to acquire raw dynamic signal data, obtain risk characteristic data, and preprocess the raw dynamic signal data and risk characteristic data. The handwritten data quality assessment and grading module is used to perform signal coupling evaluation analysis on the raw dynamic signal data, determine the acquisition quality based on the signal coupling evaluation analysis results, execute an interactive prompt mechanism, and enter the risk grading process. The dynamic index discrimination and prompting module is used to perform sensitivity amplification theory analysis on the raw dynamic signal data and execute a risk classification process based on the results of the sensitivity amplification theory analysis. The specific process of sensitivity amplification theory analysis includes: acquiring acceleration time-series signals, velocity data, pressure time-series signals, acquired signal quality values, trajectory point sequence data, and tilt angle synchronous acquisition data; obtaining the average jerk from the acceleration time-series signal using a sliding window differential algorithm and a mean algorithm; performing frequency domain analysis on the velocity and pressure time-series signals using Fast Fourier Transform to calculate the energy sum of a specified high-frequency band and obtain the micro-vibration spectrum energy value; obtaining the average velocity from the trajectory point sequence data using displacement differential and sampling time-series algorithms and a mean algorithm; obtaining the velocity-pressure coupling coefficient from the velocity and pressure time-series signals using mutual information analysis; obtaining the mean trajectory curvature from the trajectory point sequence data using differential geometry algorithms and a mean algorithm; obtaining the velocity-tilt coupling coefficient from the velocity and tilt angle synchronous acquisition data using a mutual information algorithm; and obtaining the average acceleration from the acceleration time-series signal using a mean algorithm. The following steps are performed: The acceleration time-series signal is extracted using Fast Fourier Transform (FFT), and the peak value of the dominant frequency component with the largest amplitude is obtained. The spatial distance between the first and last points and the deviation from the ideal closed trajectory are calculated for the trajectory point sequence data. The trajectory closure error value is obtained using the shortest closed-loop matching method. The product of the average jerk and the micro-vibration spectrum energy value is calculated to obtain the dynamic amplitude modulation term. The product of the average velocity and the velocity-pressure coupling coefficient is subtracted from the product of the mean trajectory curvature and the velocity tilt angle coupling coefficient. The square of the result is added to 1, and the natural logarithm is taken to obtain the velocity coupling curvature linkage logarithm term. The π-fold sine of the acquired signal quality value is calculated to obtain the signal period fluctuation term. The dynamic amplitude modulation term, the velocity coupling curvature linkage logarithm term, and the signal period fluctuation term are added together to obtain the dynamic characteristic comprehensive numerator term. The absolute value of the difference between the average acceleration and the peak value of the dominant frequency is added to the square of the trajectory closure error value, and the square root is added to 1 to obtain the abnormal characteristic modulation denominator term. The dynamic characteristic comprehensive numerator term is divided by the abnormal characteristic modulation denominator term, and the absolute value is taken to obtain the dynamic sensitivity value. The abnormal risk scoring module is used to train a shallow feedforward neural network model on the raw dynamic signal data and output a continuous comprehensive risk score. The risk feature integration and early judgment module is used to perform covariance fluctuation enhancement discrimination on the original dynamic signal data and risk feature data, and to perform early dynamic anomaly discrimination based on the covariance fluctuation enhancement discrimination.

2. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process for acquiring raw dynamic signal data and obtaining risk characteristic data is as follows: The raw dynamic signal data is collected, including: contact event data, time synchronization data, acceleration timing signals, displacement multi-channel signals, trajectory point sequence data, velocity data, pressure timing signals, and tilt angle synchronous acquisition data. Obtain risk characteristic data, which includes: data from the International Standard Clinical Movement Disorder Rating Scale, physiological age, and movement disorder data of the subject's first-degree relatives.

3. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process for preprocessing the raw dynamic signal data and risk characteristic data is as follows: By employing wavelet denoising algorithms and mutual information feature filtering methods, multi-channel denoising and feature correlation analysis are performed on the original dynamic signal data to extract dynamic change-sensitive signals and key micro-anomaly parameters. Principal component analysis and multi-scale decomposition algorithms are used to denoise and reduce the dimensionality of features in the original dynamic signal data, extracting the main spatiotemporal dynamic change components and anomaly distribution trends. Sliding window differentiation and feature remapping algorithms are used to extract dynamic features and highlight change trends in the original dynamic signal data. Distribution standardization and linear normalization algorithms are used to standardize and normalize the original dynamic signal data and risk feature data.

4. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process of performing signal coupling evaluation analysis on the original dynamic signal data is as follows: The system acquires trajectory point sequence data, touch event data, time synchronization data, acceleration timing signals, pressure timing signals, and displacement multi-channel signals. For the trajectory point sequence data, a connectivity analysis algorithm is used to calculate the effective trajectory length. For the trajectory point sequence data, a trajectory segment accumulation algorithm based on Euclidean distance is used to obtain the total trajectory length. The ratio of the effective trajectory length to the total trajectory length is calculated to obtain the trajectory integrity rate. For the pressure timing signals and touch event data, a stroke continuity discrimination algorithm is used to calculate the ratio of the number of consecutive stroke segments to the total number of strokes, obtaining the consecutive stroke ratio. For the pressure timing signals and time synchronization data, a non-touch segment duration accumulation algorithm is used to calculate the ratio of the accumulated non-touch duration to the total acquisition time, obtaining the non-touch duration percentage. Instantaneous abnormal fluctuation values ​​are obtained from acceleration and pressure time-series signals using short-time Fourier transform and peak energy accumulation algorithms; environmental noise is obtained from acceleration, pressure, and displacement multi-channel signals using bandpass filtering and statistical noise power estimation algorithms. The product of trajectory integrity rate and continuous stroke ratio is calculated and used as the input of the sin function. The sine value is then multiplied by π to obtain the trajectory stroke correlation oscillation term. Calculate the difference between the instantaneous abnormal fluctuation value and the environmental noise, square it, add 1, and then take the natural logarithm to obtain the abnormal noise logarithmic adjustment term; Calculate the square of the percentage of time not touched the pen to obtain the energy term for the period of time not touched the pen; The quality value of the acquired signal is obtained by adding the trajectory stroke association oscillation term, the abnormal noise logarithmic adjustment term, and the energy term of the non-pen touch duration.

5. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process of determining the acquisition quality based on the signal coupling evaluation and analysis results, executing the interactive prompt mechanism, and entering the risk classification process is as follows: Real-time comparison of acquired signal quality values ​​and acquired signal quality thresholds, including primary quality thresholds and secondary quality thresholds: When the quality value of the acquired signal is less than or equal to the secondary quality threshold, it is determined to be a signal in the safe zone; and it directly enters the risk classification process. When the quality value of the acquired signal is greater than the secondary quality threshold and less than or equal to the primary quality threshold, it is determined to be a warning interval signal. Triggering interactive prompt mechanism: Push the abnormal acquisition prompt to the user operation interface. The user can choose to reacquire data and continue to enter the risk classification process. If the user's waiting time threshold is exceeded, the data acquisition in this round will be blocked, a signal quality warning database will be created, and the original dynamic signal data will be archived to the signal quality warning database. When the acquired signal quality value is greater than or equal to the first-level quality threshold, it is determined to be an invalid interval signal; acquisition is terminated, a pop-up window on the interface prompts that acquisition is invalid, the data acquired in this round is blocked, a signal quality anomaly database is created, and the original dynamic signal data is recorded to the signal quality anomaly database.

6. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process of implementing the risk grading procedure based on the analysis results of the sensitivity amplification theory is as follows: Real-time comparison of kinetic sensitivity values ​​and kinetic sensitivity thresholds, including primary and secondary sensitivity thresholds, and execution of risk grading procedures: When the dynamic sensitivity value is greater than or equal to the first-level sensitivity threshold, it is determined that the dynamic abnormality is significant. The dynamic abnormality prompt will be pushed to the user operation interface, and an evaluation report and a report of suspicious characteristic parameters will be output to the user. When the dynamic sensitivity value is greater than or equal to the secondary sensitivity threshold and less than the primary sensitivity threshold, it is judged as an early weak dynamic anomaly. The raw dynamic signal data is pushed to the anomaly risk scoring module; When the kinetic sensitivity value is less than the secondary sensitivity threshold, it is determined that no suspicious kinetic abnormality has been detected, and a pop-up window suggests continuing daily health management.

7. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process of training a shallow feedforward neural network model on the original dynamic signal data to output a continuous comprehensive risk score is as follows: A multi-dimensional input feature vector, composed of the acquired signal quality value, dynamic sensitivity value, average velocity, average acceleration, average jerk, micro-vibration spectrum energy value, peak frequency, mean trajectory curvature, trajectory closure error value, velocity-pressure coupling coefficient, and velocity-tilt coupling coefficient, is used as training samples. Based on the training samples, the Z-score standardization method is used to map features of different dimensions and ranges to a standard normal distribution interval. A shallow feedforward neural network model is constructed using the standardized training samples. The shallow feedforward neural network contains one or two hidden layers, each using a linear rectified unit as the activation function, and the output layer uses a sigmoid activation function. The output value of the shallow feedforward neural network model is compressed to a unified standard interval. Supervised learning is carried out based on historically determined normal and abnormal dynamic samples. The neural network parameters are optimized by minimizing the binary cross-entropy loss function to complete the training of the shallow feedforward neural network model. The shallow feedforward neural network outputs a continuous comprehensive risk score. The value ranges from 0 to 1.

8. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process for performing covariance fluctuation enhancement discrimination on the original dynamic signal data and risk characteristic data is as follows: The study obtained comprehensive risk scores, data from the International Standard Clinical Movement Disorder Rating Scale, physiological age, and movement disorder data of the examinee's first-degree relatives; for the movement disorder data of the examinee's first-degree relatives, a family history risk value was obtained through a binary mapping algorithm. Based on data from the International Standard Clinical Movement Disorder Assessment Scale, writing speed, handwriting size, and line fluency are obtained by using speed calculation algorithms, stroke total length geometric algorithms, and trajectory curvature variance algorithms on trajectory point sequence data. Writing speed, handwriting size, and line fluency are then quantitatively assessed separately, and the scoring scale scores are obtained through a standardized weighted summation algorithm. Calculate the covariance of the overall risk score and the score of the rating scale, multiply the product of the physiological age plus 1, and then multiply it by the standard deviation of the product sequence of the family history risk value and the overall risk score to obtain the risk dynamic enhancement term; Calculate the product of the rating scale score and the physiological age, subtract the square of the comprehensive risk score, and take the absolute value to obtain the joint suppression term of the features; subtract the joint suppression term of the features from the risk dynamic enhancement term, and take the absolute value to obtain the comprehensive discrimination value of the risk features.

9. The handwriting dynamics cascaded integrated learning screening system according to claim 1, characterized in that: The specific process for early dynamic anomaly detection based on covariance fluctuation enhancement discrimination is as follows: Real-time comparison of comprehensive risk characteristic discrimination value and comprehensive risk characteristic discrimination threshold: When the comprehensive risk characteristic judgment value is less than the comprehensive risk characteristic judgment threshold, it is determined that no early dynamic abnormality has been detected. The user is not notified that there is an abnormality, and a pop-up window suggests continuing daily health management. When the comprehensive risk characteristic judgment value is greater than or equal to the comprehensive risk characteristic judgment threshold, it is determined that there is an early dynamic anomaly; Early dynamic anomaly alerts are pushed to the user interface, and evaluation reports and reports of suspicious characteristic parameters are output to the user.

Images