Intelligent screening method and system for handwritten images of patients with cognitive impairment in the elderly

Abstract

This invention discloses an intelligent screening method and system for handwritten images of elderly patients with cognitive impairment, relating to the fields of artificial intelligence and medical assistance. It acquires handwritten images and writing trajectory data generated by subjects in at least two handwriting tasks, constructs a dynamic motion feature set, stroke order consistency features, trajectory structure features, and shape stability features, performs individual baseline correction on the features, and adaptively assigns weights to the task contributions to form fused features. The fused features are input into a learning model to output a cognitive risk score, along with explanatory information related to the score. By constructing robust high-order variation energy measures and fixed-dimensional structural representations for discrete sampling trajectories and achieving adaptive fusion of task weights, the stability, sensitivity, and interpretability of cognitive impairment screening are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical assistance technology, specifically to an intelligent screening method and system for handwritten images of elderly patients with cognitive impairment. Background Technology

[0002] Early screening for cognitive impairment in the elderly often involves paper-and-pencil tasks such as drawing clocks, copying simple shapes, and transcribing sentences during geriatric, neurological, and community health checkups. Clinical assessments are typically based on the final image, with physicians manually scoring based on factors such as digit placement, shape closure, missing strokes, and overall layout. This method is time-consuming in high-traffic outpatient settings, relies heavily on experience for scoring consistency, and struggles to quantify process-related issues such as increased pauses, inconsistent writing speed, increased shaky transitions, and abnormal stroke order. These process-related issues often appear earlier than the final image, making screening based solely on the final image insufficiently sensitive for mild stages.

[0003] With the application of touch devices and electronic pens in hospital follow-ups, rehabilitation training, and home monitoring, some digital solutions have begun to collect trajectories and extract statistics such as total time, average speed, number of pauses, and average pressure for model judgment. However, in practical applications, there are differences in sampling frequency, sampling jitter, and pressure acquisition capabilities among different terminals, as well as significant differences in the writing amplitude and speed of subjects. This leads to inconsistencies in the feature dimensions and distributions of the same subject on different devices or tasks, causing fluctuations in model output. Furthermore, in multi-task screening, each task covers different dimensions of cognitive ability. Simply concatenating multi-task features into the model makes it difficult to reflect the true contribution of each task to the risk score and hinders the provision of interpretable conclusions that can be verified by doctors. Therefore, the main challenge of current technology lies in how to stably construct comparable, robust, and interpretable handwriting features and fusion mechanisms under real-world conditions of discrete sampling and multi-task coexistence, in order to improve the stability and clinical usability of screening results. Summary of the Invention

[0004] Technical Objective: To address the shortcomings of existing technologies, this invention discloses an intelligent screening method and system for handwritten images of elderly patients with cognitive impairment. By employing task structure description, trajectory quality control, discrete robust stationarity features and fixed-dimensional structure representation, as well as a task adaptive fusion mechanism, the method improves the stability, sensitivity, and interpretability of screening results.

[0005] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent screening of handwritten images of elderly patients with cognitive impairment includes the following steps:

[0007] S1. Based on the task protocol, at least two handwriting tasks are presented on the terminal and handwriting data generated by the subject after completing each handwriting task is collected. The handwriting data includes handwriting images and writing trajectory data.

[0008] S2. Perform preprocessing on the writing trajectory data. The preprocessing includes at least time axis unification, coordinate normalization, and stroke segmentation to obtain the stroke sequence and trajectory sequence corresponding to the handwriting task.

[0009] S3. Construct a feature vector for each handwriting task. The feature vector shall include at least: dynamic motion features reflecting the writing process, stroke order consistency features reflecting the execution plan, trajectory structure features reflecting spatial organization, and shape stability features reflecting the result form.

[0010] S4. Based on the subject difference elimination process, the feature vector is corrected, and based on the task adaptive adjustment, the feature contribution of different handwriting tasks is allocated to obtain the fused features;

[0011] S5. Input the fused features into the learning model and output a cognitive risk score;

[0012] S6. Output the cognitive risk score and the explanatory information, which is used to indicate the task identifier and feature identifier that contribute the most to the cognitive risk score.

[0013] Preferably, the task protocol configures a task structure description for each handwriting task. The task structure description includes at least a standard template and a constraint set, which is used to limit the target components of the handwriting task and the relative relationships between the components.

[0014] Preferably, the time axis unification includes resampling the writing trajectory at different terminal sampling frequencies using a unified time reference; the coordinate normalization includes mapping the trajectory coordinates to a unified coordinate system and eliminating translation and scale differences; and the stroke segmentation includes dividing the trajectory into stroke sequences based on the pen tip contact state or abrupt changes in time intervals.

[0015] Preferably, before constructing the feature vector, a trajectory quality score is calculated for each handwriting task, and the trajectory is processed by removing abnormal segments, interpolating missing segments, or determining invalid task samples based on the trajectory quality score.

[0016] Preferably, the dynamic motion feature includes a robust stationarity feature under discrete sampling conditions. The robust stationarity feature is obtained by constructing a high-order finite difference from the trajectory velocity sequence after denoising and resampling, and normalizing the energy of the high-order finite difference by length and duration.

[0017] Preferably, the stroke order consistency feature is calculated by the degree of violation between the constraint set and the actual order of the stroke sequence, wherein the degree of violation includes at least one of the number of violations and the proportion of violations.

[0018] Preferably, the trajectory structure features are obtained by constructing a fixed-dimensional structure representation from a set of key points. The fixed-dimensional structure representation is jointly determined by the temporal and spatial adjacency relationships of the key points and obtained through vectorization and normalization.

[0019] Preferably, the shape stability feature includes: rigidly aligning the handwritten shape with a standard template, calculating the maximum contour deviation, overlap rate deviation, and direction deviation respectively, combining the three into a comprehensive deviation, and obtaining a stability score based on the comprehensive deviation.

[0020] Preferably, the learning model further outputs a confidence score, which is obtained by calibrating and mapping the model output; the explanatory information is calculated from the internal weight allocation results or feature contribution, and a preset number of items with the highest contribution are output.

[0021] A smart screening system for handwritten images of elderly patients with cognitive impairment, used to implement the smart screening method for handwritten images of elderly patients with cognitive impairment as described above, including:

[0022] The task and acquisition module is used to present the task agreement and acquire handwritten images and writing trajectory data;

[0023] The preprocessing module is used to perform time axis consistency, coordinate normalization, and stroke segmentation on the writing trajectory data;

[0024] The feature construction module is used to construct dynamic motion features, stroke order consistency features, trajectory structure features, and shape stability features.

[0025] The fusion assessment module is used to perform subject difference elimination processing and task adaptive adjustment to generate fusion features, and input the fusion features into the learning model to output a cognitive risk score;

[0026] The output module is used to output cognitive risk scores and explanatory information.

[0027] Beneficial Effects: The intelligent screening method and system for handwritten images of elderly patients with cognitive impairment provided by this invention have the following beneficial effects:

[0028] 1. This invention revolves around a minimal closed loop encompassing task protocol acquisition—trajectory preprocessing—construction of four types of features—difference elimination and task adaptation—learning model scoring—interpretation output. Through time axis consistency, coordinate normalization, and stroke segmentation, the differences in sampling frequency and writing scale across different terminals are unified into comparable data representations. Furthermore, combined with trajectory quality scoring and abnormal segment processing, it can eliminate, interpolate, or invalidate trajectories when noise, disconnections, or missing segments occur during acquisition, preventing low-quality samples from directly entering the model and causing score drift, thus improving screening stability from the source.

[0029] 2. At the feature level, this invention does not rely solely on low-order statistics, but rather constructs robust stationarity features under discrete sampling conditions. It quantifies the fluency changes in the writing process through resampling, denoising, and high-order finite-difference energy normalization, making the features more robust to differences in sampling frequency and trajectory scale. Simultaneously, through stroke order consistency features and fixed-dimensional trajectory structure features, it captures early anomalies from the perspectives of execution planning and spatial organization, respectively, compensating for the shortcomings of relying solely on the final image or speed pauses to identify disordered sequences and structural layout deviations. Furthermore, by combining template alignment and shape stability features obtained through comprehensive deviation decomposition, the final morphological evaluation is unaffected by simple pixel similarity on stroke thickness and slight jitter, improving the discriminative effectiveness for tasks such as copying and drawing clocks.

[0030] 3. At the fusion and output level, through subject difference elimination processing and task adaptive adjustment, the model can allocate task contributions according to task quality and discrimination during multi-task screening, and can further output confidence and explanatory information, clearly indicating the task and feature category that contributes the most to the cognitive risk score. This provides clinicians with verifiable evidence, making it easier to quickly locate the source of abnormalities in outpatient or community screening and decide whether to initiate further neuropsychological scale assessments or imaging examinations. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0032] Figure 1 This is a flowchart illustrating the intelligent screening method for handwritten images of elderly patients with cognitive impairment according to the present invention.

[0033] Figure 2 This is a schematic diagram of the intelligent screening system for handwritten images of elderly patients with cognitive impairment according to the present invention;

[0034] Figure 3 This is a schematic diagram of the internal structure of the feature building module of the present invention;

[0035] Figure 4This is a schematic diagram of trajectory quality control and anomaly handling of the present invention;

[0036] Figure 5 This is a schematic diagram showing the relationship between the trajectory quality score Q and the anomaly handling strategy in this invention.

[0037] Figure 6 This is a schematic diagram showing the reliability relationship before and after confidence level calibration of the present invention. Detailed Implementation

[0038] The present invention will now be described more clearly and completely by way of a preferred embodiment in conjunction with the accompanying drawings, but this does not limit the invention to the scope of the described embodiment.

[0039] like Figure 1 As shown, this invention provides an intelligent screening method for handwritten images of elderly patients with cognitive impairment, specifically including the following steps:

[0040] S1. Based on the task protocol, present at least two handwriting tasks on the terminal and collect handwriting data formed by the subject in completing each handwriting task. The handwriting data includes handwriting images and writing trajectory data.

[0041] In one implementation, the terminal presents a clock-drawing task and a graphic-tracing task according to a task protocol, and may further present a sentence-writing task. A task structure description is established for each task, which includes a standard template and a set of constraints. The standard template represents the target shape or constituent elements; the set of constraints represents the relative and sequential relationships between elements. The set of constraints is stored in the form of computable constraint pairs and corresponds one-to-one with the task type.

[0042] The writing trajectory data consists of a sequence of sampling points, each of which contains at least an x-coordinate, a y-coordinate, and a timestamp t. When the terminal has pressure acquisition capabilities, it further includes the pressure value f; when the terminal has contact status acquisition capabilities, it further includes the contact identifier c. After acquisition, a trajectory sequence and handwritten image are generated for each task, and the task identifier, device identifier, and sampling frequency are recorded as metadata for subsequent consistency processing.

[0043] S2. Perform preprocessing on the writing trajectory data. The preprocessing includes at least time axis unification, coordinate normalization, and stroke segmentation to obtain the stroke sequence and trajectory sequence corresponding to the handwriting task.

[0044] Time axis unification is used to unify trajectories with different sampling frequencies to the same time base. It applies uniform time intervals to the trajectories. Resampling is performed to create equally spaced sequences, making velocity, difference, and energy calculations comparable across different terminals. Coordinate normalization is used to eliminate translation and scale differences, including mapping trajectories to a unified coordinate system and normalizing the scale, reducing the impact of differences in writing amplitude on feature dimensions.

[0045] Stroke segmentation is used to form stroke sequences and record stroke order indices. When the terminal provides a contact identifier, a continuous segment from the moment the pen is placed to the moment it is lifted is defined as a stroke. When the terminal does not provide a contact identifier, the breakpoint where the timestamp difference is greater than a segmentation threshold is used as the segmentation point. The threshold is related to the sampling frequency and can be determined by device metadata. After segmentation, a stroke sequence is obtained, where each stroke contains a continuous sampling point subsequence. The order index of the stroke in the current task is also recorded for subsequent order consistency calculations.

[0046] To avoid noise and missing segments directly entering the model and causing scoring instability, a trajectory quality score Q is calculated for each task. The trajectory quality score is jointly determined by the sampling missing rate, jitter intensity, and the proportion of broken strokes. The sampling missing rate is calculated as the proportion of the actual sampling interval deviating from the resampling interval; the jitter intensity is calculated as the energy of the coordinate difference before and after denoising or the energy of high-frequency components; the proportion of broken strokes is calculated as the proportion of segments with abnormally increased stroke intervals accompanied by displacement jumps.

[0047] Anomaly handling is performed based on the trajectory quality score. When the length of consecutive missing segments exceeds a threshold, they are identified as invalid segments and removed. When the length of missing segments is within the recoverable range, interpolation is used to complete the segment and it is aligned with the resampling. When the trajectory quality score is lower than the effective threshold, the task sample is identified as invalid and a re-collection is prompted, thereby ensuring the reliability and consistency of the feature construction input.

[0048] S3. Construct a feature vector for each handwriting task. The feature vector shall include at least: dynamic motion features reflecting the writing process, stroke order consistency features reflecting the execution plan, trajectory structure features reflecting spatial organization, and shape stability features reflecting the result form.

[0049] like Figure 3 As shown, the feature construction stage combines the preprocessed trajectory / stroke sequence with the handwritten image input motion features, stroke order consistency features, trajectory structure features, and shape stability features to form a unit, generating and aggregating the task feature vector set.

[0050] Dynamic motion features are used to characterize the kinematic performance of the writing process. In addition to statistics such as average speed, speed variance, number of pauses and total pause duration, this implementation further constructs a robust stationarity feature, Smooth, under discrete sampling conditions to characterize the higher-order variability energy of the writing process and improve cross-terminal comparability.

[0051] Specifically, the velocity vector is first calculated from the resampled and denoised trajectory. and its modulus And then Constructing a third-order finite difference To characterize higher-order changes, and then... The energy E is obtained by summing the squares of the values ​​over time, and then normalizing the energy with respect to the trajectory length L and duration T. Finally This is mapped to a stationarity score, Smooth. The calculation formula is as follows:

[0052]

[0053]

[0054]

[0055]

[0056] in, M is the resampling time interval, and M is the number of sampling points. For denoising coordinates, For velocity vector, For velocity modulus, This is a third-order finite-difference ... For normalized energy, A positive scaling factor. The value of Smooth ranges from 0 to 1, with a larger value indicating a smoother writing process.

[0057] To ensure that the difference calculation is clearly defined at the boundaries, this implementation method... The sequence employs a mirror extension or supplementary point strategy at the boundaries to make... can be This is derived to avoid computational uncertainty caused by missing indexes.

[0058] Stroke order consistency is used to reflect execution planning and sequence control capabilities. This implementation constructs sequence constraints based on the constraint set in the task structure description, and compares the actual stroke order index obtained from stroke segmentation with the sequence constraints to obtain the number of violations and the violation ratio. The number of violations characterizes the absolute degree of constraint violation, while the violation ratio eliminates the impact of differences in constraint set size. A higher violation ratio indicates poorer sequence consistency and can serve as an important input for cognitive risk assessment.

[0059] Trajectory structure features are used to characterize spatial organization and structural errors, avoiding insensitivity to overall layout anomalies caused by relying solely on speed and pause statistics. This implementation first extracts a set of key points from the trajectory. These key points are determined by discrete curvature or rate of change of direction, and are used to characterize turning points and structural nodes. To ensure a fixed dimension in the structural representation, the number of key points is fixed. When key points are insufficient Replace the placeholders in time, when the key points exceed Sort by intensity of change and retain the first one. Subsequently, an adjacency relationship that simultaneously considers temporal and spatial adjacency is constructed. The adjacency relationship is vectorized and normalized to obtain a fixed-dimensional structural representation, which is used as the input of trajectory structural features to the fusion evaluation module and the learning model.

[0060] Shape stability features are used to quantify the stability of the final shape relative to the standard template. This implementation first performs rigid alignment between the handwritten shape and the standard template. This rigid alignment includes translation, rotation, and scale adjustment to eliminate differences in writing position, orientation, and overall scale. After alignment, the maximum contour deviation H, overlap rate I, and directional deviation A are calculated respectively. The maximum contour deviation characterizes and normalizes the worst-case deviation of the contour; the overlap rate characterizes the degree of overlap between the handwritten area and the template area; and the directional deviation characterizes and normalizes the deviation from the main direction. To unify the three types of deviations into a stability metric that can be input into the model, this implementation combines them into a comprehensive deviation D, and maps D to a stability score Shape, calculated as follows:

[0061]

[0062] in, Non-negative weights and , All values ​​can be normalized to the range of 0 to 1. The value of Shape ranges from 0 to 1, and the larger the value, the more stable the shape.

[0063] S4. Based on the subject difference elimination process, the feature vector is corrected, and based on the task adaptive adjustment, the feature contribution of different handwriting tasks is allocated to obtain the fused features.

[0064] To reduce the impact of individual and equipment differences, this implementation method constructs a baseline vector for the subjects. The baseline vector is derived from the subject's historical handwritten data or reference statistics from the same stratum population, and its dimension is consistent with that of the feature vector. For each task's feature vector... Performing interpolation correction yields This eliminates stable individual writing habit deviations.

[0065] To reflect the differences in contributions from multiple tasks and improve interpretability, this implementation calculates task weights for each task. The task weights are determined by the task quality score and the discrimination score, and the sum of the task weights is 1. For ease of implementation, a feasible normalized expression is given below:

[0066]

[0067] in, The task score represents the fusion of task quality and discrimination. A positive scaling factor. Fusion features. It can be formed by the correction features of each task under the action of task weights, so that the contribution of different tasks to the final score can be explicitly allocated and interpreted.

[0068] S5. Input the fused features into the learning model and output a cognitive risk score.

[0069] The learning model receives fused features and outputs a cognitive risk score, Score, which is a real number from 0 to 100. To indicate the reliability of the conclusions, the learning model further outputs an uncalibrated confidence score. The confidence calibration module will Mapped to calibration confidence This ensures that the calibration confidence level is consistent with the actual accuracy on the validation set, thereby facilitating stratified treatment by physicians.

[0070] S6. Output the cognitive risk score and the explanatory information, which is used to indicate the task identifier and feature identifier that contribute the most to the cognitive risk score.

[0071] The output module also outputs explanatory information. This explanatory information indicates the set of task identifiers and feature identifiers that contribute the most to the score. The contribution can be obtained from the model's internal weight allocation or feature contribution ranking, and a preset number of items are output to help physicians quickly locate the source of the abnormality, thereby improving the interpretability and clinical usability of the conclusions.

[0072] like Figure 2 As shown, the present invention also provides an intelligent screening system for handwritten images of elderly patients with cognitive impairment, used to implement the intelligent screening method for handwritten images of elderly patients with cognitive impairment as described above, including:

[0073] The task and acquisition module is used to present the task agreement and acquire handwritten images and writing trajectory data.

[0074] In one embodiment, the task and acquisition module includes a task presentation unit, an input acquisition unit, and a metadata recording unit. The task presentation unit is used to display task prompts, standard templates, or reference structures on the terminal according to the task protocol, and associate a task identifier with each task; the input acquisition unit is used to acquire a sequence of sampling points from a touch screen, electronic pen, or graphics tablet, where each sampling point at least includes coordinate values ​​and a timestamp; the metadata recording unit is used to record the device identifier, sampling frequency, resolution, and task start / end time, and saves the trajectory sequence and corresponding handwritten images in a unified data structure to support traceable processing by subsequent modules.

[0075] The preprocessing module is used to perform time axis consistency, coordinate normalization, and stroke segmentation on the writing trajectory data.

[0076] In one embodiment, the preprocessing module includes a resampling unit, a denoising unit, a coordinate normalization unit, and a stroke segmentation unit, and may further include a trajectory quality control unit. The resampling unit unifies trajectories with different sampling frequencies to the same time reference, mapping the trajectories to equal time interval sequences through interpolation; the denoising unit suppresses acquisition jitter and spike noise, outputting smooth trajectories; the coordinate normalization unit eliminates translation and scale differences, mapping the trajectories to a unified coordinate system and normalizing the scale; the stroke segmentation unit segments the trajectory into stroke sequences based on changes in contact state or abrupt changes in time intervals and labels the stroke order index. The trajectory quality control unit generates a trajectory quality score based on the sampling missing rate, jitter intensity, and the proportion of broken strokes, and triggers abnormal segment removal, short missing segment imputation, or invalid task sample determination accordingly, thereby ensuring the reliability and consistency of the feature construction input.

[0077] like Figure 4 As shown, trajectory quality control and anomaly handling include processes such as quality index calculation, quality score generation, threshold determination, and invalidation determination or repair processing to ensure the quality of trajectory inputs entering stroke segmentation and feature construction.

[0078] The trajectory quality score Q is used to characterize the usability of the trajectory input for subsequent feature construction. Q can be constructed based on multiple quality indicators: the sampling missing rate represents the proportion of missing or interrupted sampling points, the pen break anomaly ratio represents the proportion of unexpected pen break events, and the jitter intensity represents the intensity of minor, repeated oscillations of the trajectory within a short timescale. After normalizing each indicator, Q is obtained through weighted convergence or nonlinear mapping, and then... , This serves as a threshold for tiered processing, including invalidation detection, repair, and direct approval.

[0079] like Figure 5 As shown, a correlation can be established between the trajectory quality score Q and the processing strategy: when Q is below the first threshold... The task sample is deemed invalid when Q is in a certain state. With the second threshold When Q is not less than 1, perform outlier removal or imputation repair on short missing segments; It directly proceeds to stroke segmentation and feature construction processing, thereby maintaining the consistency of feature construction input under different acquisition qualities.

[0080] The feature construction module is used to construct dynamic motion features, stroke order consistency features, trajectory structure features, and shape stability features.

[0081] In one embodiment, the feature construction module includes a motion feature construction unit, a stroke order consistency feature construction unit, a trajectory structure feature construction unit, and a shape stability feature construction unit. The motion feature construction unit calculates indicators such as speed, pauses, and higher-order change energy on the resampled and denoised trajectory, and normalizes the energy indicators by length and duration to form a robust stability representation comparable across terminals. The stroke order consistency feature construction unit compares the actual stroke order with the constraints in the task structure description to obtain the number and proportion of violations. The trajectory structure feature construction unit extracts a set of key points from the trajectory and constructs temporal and spatial adjacency relationships to form a fixed-dimensional structural representation vector. The shape stability feature construction unit rigidly aligns the handwritten shape with a standard template and calculates components such as maximum contour deviation, overlap rate deviation, and direction deviation, combining them to obtain a shape stability representation. The features output by each unit are normalized and then aggregated into a task feature vector according to the task identifier.

[0082] like Figure 3 As shown, the feature construction module takes the preprocessed trajectory and image as input, and generates task feature vectors through the construction units of motion features, stroke order consistency features, trajectory structure features and shape stability features, respectively. After feature normalization and task aggregation, the module outputs the task feature set.

[0083] The fusion assessment module is used to perform subject difference elimination processing and task adaptive adjustment to generate fusion features, and input the fusion features into the learning model to output a cognitive risk score.

[0084] In one embodiment, the fusion assessment module includes a difference elimination unit, a task weight calculation unit, a fusion feature generation unit, and a model inference unit. The difference elimination unit performs baseline correction on the subject's features, making the features insensitive to individual stable writing habits and device differences. The task weight calculation unit generates task weights based on task quality scores and task discrimination, ensuring the weights sum to 1, thereby explicitly allocating the contribution of each task in a multi-task scenario. The fusion feature generation unit weights and combines the feature vectors of each task according to the task weights to form a fusion feature. The model inference unit inputs the fusion feature into a learning model to obtain a cognitive risk score and can further output an uncalibrated confidence level.

[0085] The output module is used to output cognitive risk scores and explanatory information.

[0086] In one embodiment, the output module includes a scoring mapping unit, a confidence calibration unit, and an interpretation generation unit. The scoring mapping unit maps the model output to a cognitive risk score with preset dimensions; the confidence calibration unit performs calibration mapping on uncalibrated confidence scores to ensure consistency between the confidence scores and the validation set accuracy, facilitating stratified treatment by physicians; the interpretation generation unit generates explanatory information based on the model's internal weight allocation results or feature contribution calculation results. The explanatory information at least indicates the task identifier and feature identifier with the highest contribution, and is output to the terminal interface or electronic medical record interface.

[0087] like Figure 6 As shown, by establishing a mapping relationship between prediction confidence and empirical accuracy on the validation set, the uncalibrated confidence can be... Calibration function Mapped to confidence level ,make It is more consistent with the actual accuracy rate, so as to enable stratified treatment and review triggering based on confidence thresholds on the physician side.

[0088] In one embodiment, the system further includes a data storage / management module for storing task data, feature data, model parameters and log information, and providing retrieval and reading capabilities to the fusion evaluation module.

[0089] The data storage / management module is used to persistently store the collected task data, preprocessing results, feature data, model parameters and log information, and provides retrieval interfaces by subject identifier, task identifier and time range to support follow-up comparison, model updates and audit trails.

[0090] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent screening of handwritten images of elderly patients with cognitive impairment, characterized in that, Specifically, the following steps are included: S1. Based on the task protocol, at least two handwriting tasks are presented on the terminal and handwriting data generated by the subject after completing each handwriting task is collected. The handwriting data includes handwriting images and writing trajectory data. S2. Perform preprocessing on the writing trajectory data. The preprocessing includes at least time axis unification, coordinate normalization, and stroke segmentation to obtain the stroke sequence and trajectory sequence corresponding to the handwriting task. S3. Construct a feature vector for each handwriting task. The feature vector shall include at least: dynamic motion features reflecting the writing process, stroke order consistency features reflecting the execution plan, trajectory structure features reflecting spatial organization, and shape stability features reflecting the result form. S4. Based on the subject difference elimination process, the feature vector is corrected, and based on the task adaptive adjustment, the feature contribution of different handwriting tasks is allocated to obtain the fused features; S5. Input the fused features into the learning model and output a cognitive risk score; S6. Output the cognitive risk score and the explanatory information, which is used to indicate the task identifier and feature identifier that contribute the most to the cognitive risk score.

2. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The task protocol configures a task structure description for each handwriting task. The task structure description includes at least a standard template and a set of constraints. The set of constraints is used to limit the target components of the handwriting task and the relative relationships between the components.

3. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The time axis unification includes resampling the writing trajectory at different terminal sampling frequencies using a unified time reference; the coordinate normalization includes mapping the trajectory coordinates to a unified coordinate system and eliminating translation and scale differences; the stroke segmentation includes dividing the trajectory into stroke sequences based on the pen tip contact state or abrupt changes in time intervals.

4. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, Before constructing the feature vector, a trajectory quality score is calculated for each handwriting task, and the trajectory is processed by removing abnormal segments, interpolating missing segments, or determining invalid task samples based on the trajectory quality score.

5. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The dynamic motion features include robust stationarity features under discrete sampling conditions. These robust stationarity features are obtained by constructing a high-order finite difference from the trajectory velocity sequence after denoising and resampling, and by normalizing the length and duration of the high-order finite difference energy.

6. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The stroke order consistency feature is calculated by the degree of violation between the constraint set and the actual order of the stroke sequence, wherein the degree of violation includes at least one of the number of violations and the proportion of violations.

7. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The trajectory structure features are obtained by constructing a fixed-dimensional structure representation from the set of key points. The fixed-dimensional structure representation is jointly determined by the temporal and spatial adjacency relationships of the key points and obtained through vectorization and normalization.

8. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The shape stability features include: rigidly aligning the handwritten shape with a standard template, calculating the maximum contour deviation, overlap rate deviation, and orientation deviation respectively, and combining the three into a comprehensive deviation, and obtaining a stability score based on the comprehensive deviation.

9. The intelligent screening method for handwritten images of elderly patients with cognitive impairment according to claim 1, characterized in that, The learning model further outputs a confidence score, which is obtained by calibrating and mapping the model output; the explanatory information is calculated from the internal weight allocation results or feature contribution, and outputs a preset number of items with the highest contribution.

10. A handwritten image intelligent screening system for elderly patients with cognitive impairment, characterized in that, A method for intelligent screening of handwritten images of elderly patients with cognitive impairment as described in any one of claims 1-9 includes: The task and acquisition module is used to present the task agreement and acquire handwritten images and writing trajectory data; The preprocessing module is used to perform time axis consistency, coordinate normalization, and stroke segmentation on the writing trajectory data; The feature construction module is used to construct dynamic motion features, stroke order consistency features, trajectory structure features, and shape stability features. The fusion assessment module is used to perform subject difference elimination processing and task adaptive adjustment to generate fusion features, and input the fusion features into the learning model to output a cognitive risk score; The output module is used to output cognitive risk scores and explanatory information.

Images