Three interpretable intelligent assessment methods for early Parkinson's disease using multi-task remote data

Abstract

The present invention relates to the field of auxiliary medical assessment technology, and specifically discloses an interpretable three-branch intelligent assessment method for early Parkinson's disease under multi-task remote data. By fusing multi-task remote data, introducing a fuzzy partitioning strategy that can handle uncertainty data based on a decision tree, and an interpretable three-branch decision-making mechanism, a joint modeling of multidimensional symptom characteristics is achieved. In the modeling process, with the decision tree division point as the center, learnable fine-tuning parameters and offset parameters are introduced respectively, the division point is continuously fine-tuned and its fuzzy area boundary is optimized, and adaptive fuzzification processing of the decision tree division point is achieved, thereby constructing a fuzzy decision tree, and based on it, by introducing a three-branch decision-making mechanism, it can effectively achieve uncertain processing and sensitive identification of early symptoms, thereby improving assessment accuracy and clinical usability, and supporting early screening and intervention of the disease. Therefore, the present invention has strong flexibility, adaptability, and the ability to process and interpret sample uncertainty.

Description

Technical Field

[0001] The present application relates to the technical field of auxiliary medical assessment, and specifically discloses an interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data. Background Art

[0002] Parkinson's disease is a chronic, progressive neurodegenerative disorder. Its early stages are often accompanied by mild but significant motor and non-motor symptoms. Studies have shown that, compared to irreversible neurological damage in the middle and late stages, early intervention is most effective. Therefore, early and accurate assessment not only helps slow disease progression but also significantly improves patients' quality of life.

[0003] With the development of remote data collection equipment and technology, obtaining and analyzing the symptom characteristics of early Parkinson's disease, such as gait changes, voice fluctuations and handwriting abnormalities, has become more feasible and efficient, providing an important basis for the early intelligent assessment of Parkinson's disease.

[0004] At the data collection level, although remote collection equipment has broadened the data sources, the collection process is still susceptible to interference from factors such as the environment, patient status, and device mode. Traditional evaluation models built based on a single task often have difficulty in sensitively identifying changes in patient status due to their limited generalization capabilities. For this reason, obtaining data by designing multiple task scenarios has become an effective means. For example, insoles with pressure sensors are used to collect gait data of patients during normal walking, walking while listening to music, and walking while communicating; professional sound pickup equipment is used to record speech tasks such as the pronunciation of vowels and consonants; or smart tablets are used to collect handwriting trajectories of spirals, winding lines, and graphics. This type of multi-task remote data contains early representations of multiple dimensions of Parkinson's disease and has important application potential.

[0005] At the data processing level, existing methods for building evaluation models using multi-task remote data simply concatenate data from different tasks and then process them uniformly. This fails to fully eliminate the interference of individual physiological factors (such as height, age, and weight) on features, and also ignores the fact that differences in feature distribution between tasks can weaken the model's ability to extract universal features across tasks. Therefore, integrating the commonalities and differences in multi-task remote data and effectively avoiding the interference of physiological differences are the foundation for building accurate and robust evaluation models.

[0006] At the model construction level, although deep learning methods possess strong feature extraction capabilities, their "black box" nature and lack of interpretability make them difficult to gain physicians' trust. Furthermore, traditional binary classification models often rigidly categorize patients as "sick" or "healthy," which can easily lead to the risk of misjudgment when faced with patients with uncertain conditions. Therefore, in the context of telemedicine, there is an urgent need to develop an intelligent early-stage Parkinson's disease assessment method that can both handle uncertainty and provide good interpretability, providing physicians with more reliable and transparent auxiliary diagnostic support.

[0007] The present invention provides an interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data to solve the above problems. Summary of the Invention

[0008] The purpose of this invention is to construct an interpretable three-branch intelligent assessment method for early Parkinson's disease under multi-task remote data. By integrating multi-task remote data, introducing a fuzzy partitioning strategy that effectively processes uncertain data, and an interpretable three-branch decision-making mechanism, uncertainty processing and sensitive identification of early Parkinson's disease symptoms can be achieved, thereby improving assessment accuracy and clinical usability, and supporting early screening and intervention of the disease.

[0009] To achieve the above objectives, the basic solution of the present invention provides an interpretable three-branch intelligent assessment method for early Parkinson's disease using multi-task remote data, comprising the following steps:

[0010] Collect multi-task remote data to be tested, input it into the interpretable three-branch intelligent evaluation model for processing and analysis, and obtain evaluation results and interpretability reports;

[0011] When the assessment result clearly indicates whether the patient is sick, the output is directly healthy or sick; otherwise, the output is uncertain and the risk of illness is further assessed;

[0012] The method for constructing an explainable three-branch intelligent evaluation model includes the following steps:

[0013] Step S1: collect and process multi-task remote data to obtain a multi-task remote dataset;

[0014] Step S2: Use the multi-task remote dataset as a sample set to input into the C4.5 algorithm to build a clear decision tree;

[0015] Step S3: Using several non-leaf nodes in the clear decision tree as partition points, introducing a learnable fine-tuning parameter for each partition point, and introducing a learnable offset parameter for each partition point on both sides;

[0016] Step S4: Input each sample in the sample set into the clear decision tree after step S3, and use the minimization of the classification error function as the loss function to iteratively optimize the fine-tuning parameters of each division point position and the offset parameters of the positions on the left and right sides of each division point. By constructing a triangular fuzzy membership function for all the optimized division points and the offset parameters of the positions on the left and right sides, the transformation from the clear decision tree to the fuzzy decision tree is achieved;

[0017] Step S5: Based on the obtained fuzzy decision tree, aggregate the membership of each sample on a path with multiple partition points, normalize it and use it as the weight of the sample on the path, multiply the path weight by the category distribution probability of the sample in the leaf node, and add them together for the same category to determine the weighted probability vector of the sample for different categories;

[0018] Step S6: Introduce a three-branch decision-making mechanism to determine whether the weighted probability vector falls into the interval as the evaluation result, and use the size of the weighted probability vector corresponding to the evaluation result as the disease risk value to construct an interpretable three-branch intelligent evaluation model.

[0019] Further, in the implementation process of step S1, the following steps are included:

[0020] Step S11: Collection of multi-task remote data: collecting multi-task remote data of early Parkinson's patients through remote data sources or smart devices;

[0021] Step S12: Processing of multi-task remote data: including manual feature extraction, physiological feature decorrelation, cross-task feature distribution alignment and data normalization in sequence, and finally obtaining a multi-task remote dataset with unified structure and coordinated features.

[0022] Further, in the implementation process of step S2, the following steps are included:

[0023] Step S21: performing dichotomy processing on all features of the nodes of the clear decision tree to be established;

[0024] Step S22: Using the multi-task remote dataset as the sample set input, a clear decision tree is constructed by running the C4.5 algorithm;

[0025] Step S23: Use a post-pruning strategy to prevent overfitting of the clear decision tree.

[0026] Furthermore, in step S3, the expression for adjusting the position of the division point based on the fine-tuning parameter is as follows:

[0027] ;

[0028] Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the value of the split point of the original clear decision tree, is the dividing point after fine-tuning, is a learnable fine-tuning parameter;

[0029] Offset parameter of the left position of the division point and the offset parameter to the right of the division point They are represented by the following ways:

[0030] ;

[0031] ;

[0032] Where, is a learnable parameter that controls the offset parameter of the left position of the partition point. is a learnable parameter that controls the offset parameter of the right side of the partition point. and The initial value of is 0.

[0033] Furthermore, in the implementation process of step S4, the following steps are included:

[0034] Step S41: Input each sample in the multi-task remote dataset into the clear decision tree after step S3 for forward propagation, passing through each partition point in turn, and constructing a triangular fuzzy membership function based on each partition point and the offset parameters on its left and right sides to obtain the membership degree of each sample to different partition points;

[0035] Step S42: Aggregate the membership of each sample on a path with multiple partition points, normalize it, and use it as the weight of the sample on the path. Multiply the path weight by the category distribution probability of the sample in the leaf node, and then add them together for the same category to determine the weighted probability vector of the sample for different categories. Select the category with the largest weighted probability value as the predicted label of the sample.

[0036] Step S43: Use the error between the true label and the predicted label of the sample to construct a classification error loss function, and optimize three parameters through back propagation and gradient update: the fine-tuning parameter of the partition point, the offset parameter of the position to the left of the partition point, and the offset parameter of the position to the right of the partition point;

[0037] Step S44: Through multiple rounds of training, all parameters are iteratively optimized to minimize the loss function, and finally the fine-tuning parameters of each division point and the offset parameters of the positions on its left and right sides are obtained;

[0038] Step S45: Based on the final optimized fine-tuning parameters and the offset parameters of the positions on the left and right sides, the fine-tuned division points and the offset parameters of the positions on the left and right sides are obtained, and a triangular fuzzy membership function is constructed for each division point to convert each division point into a fuzzy division point, and finally a fuzzy decision tree is obtained.

[0039] Further, in step S41, the triplet of triangular fuzzy membership function is constructed as , therefore, the triangular fuzzy membership function corresponding to each fuzzy partition point is:

[0040] ;

[0041] Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the fine-tuned dividing point, is the offset parameter of the left position of the learnable partition point, is the learnable offset parameter to the right of the partition point, and x is a sample in the multi-task long-range dataset;

[0042] when hour,

[0043] ;

[0044] when hour,

[0045] .

[0046] Furthermore, in step S42, the membership of the aggregated sample x to all partition points on each path is expressed as:

[0047] ;

[0048] Where w m (x) represents the weight of sample x in the mth path, d is the path depth, μ mj (x) is the membership of sample x to the jth partition point in path m;

[0049] Normalized path weight W m Expressed as:

[0050] ;

[0051] Where M is the number of paths, W is m is the normalized weight of sample x in the mth (m=1,2,…,M) path;

[0052] The formula for the weighted probability vector (P0, P1) of sample x for different categories is as follows:

[0053] ;

[0054] Where M is the number of paths, t represents the two categories of healthy or sick, t = 0 or 1, Represents the category distribution probability of the samples in the leaf nodes under the mth path, that is, the category distribution probability of the samples in the leaf nodes in the corresponding clear decision tree;

[0055] The predicted label t of sample x is calculated using the maximum weighted probability formula:

[0056] Predicted label t=max t P t ;

[0057] Where t = 0 or 1 represents the two categories of healthy and sick.

[0058] Furthermore, in the implementation process of step S5, the following steps are included:

[0059] Step S51: After the sample is input into the fuzzy decision tree, the fuzzy partition point membership is calculated: according to the obtained fuzzy decision tree, for any fuzzy partition point, the fuzzy membership of the sample at the fuzzy partition point is calculated using the triangular fuzzy membership function;

[0060] Step S52: Calculate path weight: Aggregate the comprehensive membership values ​​of the samples on each path and normalize them to obtain the weight value of the samples on each path;

[0061] Step S53: Output weighted category probability vector: multiply the path weight and the category distribution probability of the sample in the leaf node, and then add them together under the same category to determine the probability vector of the sample for different categories ,in, represents the health probability, Represents the probability of disease.

[0062] Furthermore, during the implementation of step S6, the three decision-making mechanisms are:

[0063] when , then the output is a certain result, that is, when and When , the output result is: sick, when and When , the output result is: healthy;

[0064] when When , the output result is: uncertain, and the disease assessment value ;

[0065] The confidence threshold ξ ranges from 0.5<ξ<1.

[0066] The principle and effect of this solution are:

[0067] 1. Compared with existing technologies, the present invention is more suitable for telemedicine scenarios: it obtains patients' multi-task behavior data through a remote data acquisition module, and combines it with an intelligent classification model to achieve timely and accurate diagnostic support, effectively reducing patients' dependence on face-to-face medical treatment, and significantly improving the practicality and diagnostic efficiency of telemedicine.

[0068] 2. Compared with the existing technology, the cross-task universality and differentiation joint modeling strategy of the present invention can utilize any one of the multi-task remote data such as gait, voice, and writing to realize the joint modeling of multi-dimensional symptom characteristics. It can not only effectively enhance the comprehensiveness and robustness of the model for early assessment of Parkinson's disease, but also improve the stability and adaptability of the model in different tasks and different populations.

[0069] 3. Compared with the existing technology, the present invention utilizes a learnable fuzzy partitioning mechanism, takes the decision tree partitioning point as the center, introduces learnable fine-tuning parameters and offset parameters, and realizes adaptive fuzzification processing by fine-tuning the partitioning point and optimizing the fuzzy region boundary through forward propagation and backpropagation. Therefore, the present invention has strong flexibility, adaptability and the ability to process and interpret sample uncertainty.

[0070] 4. Compared with existing technologies, this invention features an interpretable evaluation framework, constructs a fuzzy decision tree based on a clear decision tree, and introduces fuzzy partitioning to enhance the ability to handle uncertain data. By introducing a three-branch decision mechanism, it enables flexible discrimination between "healthy-uncertain-sick." Therefore, this invention conforms to medical diagnostic logic and can effectively reduce the risk of misdiagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0072] Figure 1 A flowchart of the interpretable three-limb intelligent assessment method for early Parkinson's disease using multi-task remote data proposed in an embodiment of the present application is shown;

[0073] Figure 2 The technical roadmap of multi-task remote data collection and multi-task remote data processing proposed in the embodiments of the present application is shown;

[0074] Figure 3 The present invention shows the method for interpreting the early stage Parkinson's disease with three intelligent branches under multi-task remote data for any dividing point b.q The triangular fuzzy membership function obtained after fuzzification;

[0075] Figure 4 The figure shows a visualization auxiliary display of multi-task remote gait data used in step S5 of the interpretable three-limb intelligent assessment method for early Parkinson's disease using multi-task remote data proposed in an embodiment of the present application;

[0076] Figure 5 A schematic diagram of an interpretable three-limb intelligent assessment device for early Parkinson's disease under multi-task remote data proposed in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0077] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the specific implementation methods, structures, features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments.

[0078] Three intelligent assessment methods can be used to interpret early Parkinson's disease under multi-task remote data, such as Figures 1 to 4 As shown:

[0079] The steps include:

[0080] Collect multi-task remote data to be tested, input it into the interpretable three-branch intelligent evaluation model for processing and analysis, and obtain evaluation results and interpretability reports;

[0081] When the evaluation result clearly indicates whether the patient is sick, the output is healthy or sick directly; otherwise, the output is uncertain and the risk of illness is further evaluated.

[0082] The method for constructing an interpretable three-branch intelligent evaluation model includes the following steps:

[0083] Step S1: Collect and process multi-task remote data to obtain a multi-task remote dataset.

[0084] like Figure 2 As shown, the specific steps include:

[0085] Step S11: Multi-task remote data collection: Multi-task remote data from early-stage Parkinson's patients is collected via remote data sources or smart devices. The multi-task remote data may include any of the following: using insoles with pressure sensors to collect multi-task gait data from patients walking normally, listening to rhythmic music, and communicating with others; using a microphone (such as a mobile phone or medical-grade sound pickup device) to collect multi-task speech data from patients during vowel and consonant pronunciation tasks; and using a smart tablet to collect multi-task hand-drawn image data using handwritten spirals, winding lines, and simple shapes as templates.

[0086] Step S12: Processing of multi-task remote data: including manual feature extraction, physiological feature decorrelation, cross-task feature distribution alignment and data normalization, and finally obtaining a multi-task remote dataset with unified structure and coordinated features. Including:

[0087] The processing of multi-task gait data involves collecting multi-task gait data from patients wearing insoles with pressure sensors while walking normally, listening to music, and communicating. This process includes the following four steps:

[0088] Step a121: Manually extracting temporal and spatial features, gait cycle variability, symmetry index, step length, and single / double support time ratio of the multi-task gait data;

[0089] Step a122: Physiological feature de-correlation: Use multiple regression method to remove the influence of physiological features such as height, weight and age on multi-task gait data;

[0090] Step a123: Aligning feature distributions across tasks: Use principal component analysis (PCA) or other feature selection methods to reduce the dimensionality of the gait data for each task, extract shared principal components across tasks to construct a unified feature representation, and retain some task-specific components to enhance individual discriminability, thereby achieving feature distribution alignment across tasks.

[0091] Step a124: Data normalization: Normalize all features, such as using Z-score normalization or Min-Max normalization.

[0092] The processing of multi-task speech data, specifically the processing of multi-task speech data such as vowels and consonants emitted by patients using professional sound pickup equipment, includes the following four steps:

[0093] Step b121: Manually extracting frequency domain features, fundamental frequency features, harmonic features, and glottal dynamic features of the multi-task speech data;

[0094] Step b122: Physiological feature de-correlation: Use a multivariate regression method to remove the influence of physiological features such as height, weight, and age on the multi-task speech data;

[0095] Step b123: Aligning feature distributions across tasks: Use principal component analysis (PCA) or other feature selection methods to reduce the dimensionality of the speech data for each task, extract shared principal components across tasks to construct a unified feature representation, and retain some task-specific components to enhance individual discriminability, thereby achieving feature distribution alignment across tasks.

[0096] Step b124: Data normalization: Normalize all features, such as using Z-score normalization or Min-Max normalization.

[0097] The processing of multi-task hand-drawn image data involves processing the multi-task hand-drawn images of the handwriting trajectories of spiral lines, meandering lines, and standard shapes collected by the patient using a smart tablet. The processing includes the following four steps:

[0098] Step c121: manually extracting trajectory offset, tremor amplitude, line regularity, and morphological statistical features of the multi-task hand-drawn image;

[0099] Step c122: Physiological feature de-correlation: Use multiple regression method to remove the influence of physiological features such as height, weight and age on the multi-task hand-drawn image data;

[0100] Step c123: Aligning feature distributions across tasks: Use principal component analysis (PCA) or other feature selection methods to reduce the dimensionality of the hand-drawn image data for each task, extract shared principal components across tasks to construct a unified feature representation, and retain some task-specific components to enhance individual discriminability, thereby achieving alignment of feature distributions across tasks.

[0101] Step c124: Data normalization: Normalize all features, such as using Z-score normalization or Min-Max normalization.

[0102] By processing different multi-task remote datasets separately, we can obtain multi-task remote datasets with unified structure and coordinated features: gait multi-task remote dataset, speech multi-task remote dataset and hand-drawn image multi-task remote dataset.

[0103] Step S2: Use the multi-task remote dataset as the sample set and input it into the C4.5 algorithm to build a clear decision tree.

[0104] The specific steps include:

[0105] Step S21: Perform dichotomy on all features of the nodes of the clear decision tree to be established: for the discrete attributes of the nodes, use frequency to replace the discrete attribute value to obtain a continuous attribute value; then, use dichotomy (such as threshold-based, equal interval or equal frequency segmentation) to convert the discrete attributes and the original continuous attributes for subsequent evaluation model building.

[0106] Step S22: Using the multi-task remote dataset as the sample set input, the C4.5 algorithm is run to construct a clear decision tree. The C4.5 algorithm uses the information gain ratio to calculate the partitioning ability of each discrete feature, and the discrete feature with the largest information gain ratio is selected for partitioning. This process is recursive until one of the following termination conditions is met: (i) the information gain ratio of the best partitioning discrete feature is lower than the set threshold; (ii) the number of samples in the subset is lower than the set threshold; (iii) all samples in the subset belong to the same category.

[0107] Step S23: Use post-pruning strategy to prevent overfitting of the clear decision tree. Prune some redundant branches through cross-validation or minimizing the validation set error to improve the generalization ability of the model.

[0108] Step S3: Taking several non-leaf nodes in the clear decision tree as division points, introducing a learnable fine-tuning parameter for each division point, and introducing a learnable offset parameter for the position on the left and right sides of each division point.

[0109] Step S31: Extract several non-leaf nodes from the clear decision tree as partitioning points: Extract the partitioning attributes and corresponding partitioning points b of Q non-leaf nodes from the clear decision tree constructed in step S2. q (q=1, 2, …, Q), which serves as the central parameter of the subsequent triangular fuzzy membership function.

[0110] Step S32: introduce a learnable fine-tuning parameter for each division point, and introduce a learnable offset parameter for each division point on the left and right sides: the left offset parameter of the division point and the right offset parameter of the division point , and allows the partition point b q Make fine adjustments. The details are as follows:

[0111] ;

[0112] Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the value of the split point of the original clear decision tree, is the dividing point after fine-tuning, is a learnable fine-tuning parameter;

[0113] Offset parameter of the left position of the division point and the offset parameter to the right of the division point They are represented by the following ways:

[0114] ;

[0115] ;

[0116] Where, is a learnable parameter that controls the offset parameter of the left position of the partition point. is a learnable parameter that controls the offset parameter of the right side of the partition point. and The initial values ​​of are all 0 (i.e., consistent with the clear decision tree).

[0117] In this example, the softplus function has the following effects: (1) ensure >0, >0, that is, the offset parameter is always the effective width; (2) it remains continuous and differentiable, which facilitates gradient descent; (3) it is similar to the nonlinearity of ReLU and has a stable gradient range.

[0118] Step S4: Input each sample in the sample set into the clear decision tree after step S3, and use the minimization of the classification error function as the loss function to iteratively optimize the fine-tuning parameters of each partition point position and the offset parameters of each partition point on both sides. By constructing a triangular fuzzy membership function for all optimized partition points and their left and right offset parameters, the transformation from the clear decision tree to the fuzzy decision tree is achieved. The specific steps include the following:

[0119] Step S41: Input each sample in the multi-task remote data set into the clear decision tree after step S3 for forward propagation, pass through each division point in turn, and construct a triangular fuzzy membership function based on each division point and the offset parameters on its left and right sides to obtain the membership of each sample to different division points.

[0120] According to step S32, the triangular fuzzy membership function parameter triplet is obtained as , so, if Figure 3 As shown, in the forward propagation stage, based on the triple The triangular fuzzy membership function is:

[0121] ;

[0122] Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the fine-tuned dividing point, is the offset parameter of the left position of the learnable partition point, is the learnable offset parameter to the right of the partition point, and x is the input data: a sample in the multi-task long-range dataset.

[0123] Specifically, when hour,

[0124] ;

[0125] when hour,

[0126] .

[0127] Based on the above triangular fuzzy membership function, the membership value of sample x to each fuzzy partition node is calculated. This step uses the triangular fuzzy membership function to fuzzify the partition points in the clear decision tree, thereby constructing a soft partition region with continuous transition characteristics at each partition node.

[0128] Step S42: Aggregate the membership of each sample that contains multiple partition point paths, normalize it and use it as the weight of the sample in the path, multiply the path weight and the category distribution probability of the sample in the leaf node, and then add them together under the same category to determine the weighted probability vector of the sample for different categories, and select the category with the largest weighted probability value as the predicted label of the sample.

[0129] Specifically, the membership of all partition points of sample x on each path is aggregated by T-norm, which is expressed as:

[0130] ;

[0131] Where w m (x) represents the weight of sample x in the mth path, d is the path depth, μ mj (x) is the membership of sample x to the jth partition point in path m;

[0132] Normalized path weight W m Expressed as:

[0133] ;

[0134] Where M is the number of paths, W is m is the normalized weight of sample x in the mth (m=1,2,…,M) path;

[0135] The formula for the weighted probability vector (P0, P1) of sample x for different categories is as follows:

[0136] ;

[0137] Where M is the number of paths, t=0 or 1 represents the two categories of healthy or sick. Represents the category distribution probability of the samples in the leaf nodes under the mth path, that is, the category distribution probability of the samples in the leaf nodes in the corresponding clear decision tree;

[0138] The predicted label t of sample x is calculated using the maximum weighted probability formula:

[0139] Predicted label t=max t P t ;

[0140] Where t = 0 or 1 represents the two categories of healthy and sick.

[0141] The weight of this weighted fusion is derived from the membership value of the path, ensuring that the path with higher membership contributes more to the final classification result.

[0142] According to the above formula, the category with the largest membership and weighted probability value is selected as the predicted category of sample x. The weight of this weighted fusion is derived from the membership value of the path, ensuring that the path with higher membership contributes more to the final classification result.

[0143] Step S43: Use the error between the true label and the predicted label of the sample to construct a classification error loss function, and optimize three parameters through back propagation and gradient update: the fine-tuning parameter of the division point, the offset parameter of the position to the left of the division point, and the offset parameter of the position to the right of the division point.

[0144] During training, the cross entropy loss function is used to measure the error between the prediction and the true label. The cross entropy loss function is:

[0145] ;

[0146] Where, is the probability distribution of the predictions output by the model integrating multiple paths, y i is the one-hot encoding of the true label, and N represents the number of samples in the input dataset.

[0147] In order to improve the generalization performance and robustness of the model, the regularization term of the offset parameter is set to constrain the width of the offset parameter to prevent the fuzzy area from being too wide or too narrow and maintain boundary stability. The regularization term is expressed as:

[0148] ;

[0149] Where λ is the regularization parameter.

[0150] Therefore, the total loss function is:

[0151] ;

[0152] Backpropagation and gradient update: During the backpropagation process, the model automatically calculates the loss function for all learnable parameters using the chain rule. , and Gradient:

[0153] ;

[0154] Through layer-by-layer backpropagation, the gradient is passed from the output layer to each node in sequence. Subsequently, the gradient descent algorithm or its variants (such as Adam, RMSprop, etc.) are used to update the model parameters:

[0155] ;

[0156] ;

[0157] ;

[0158] Where η is the learning rate and n is the current iteration number.

[0159] Through this optimization process, the model continuously adjusts the parameter values ​​to minimize the loss function and improve the evaluation performance.

[0160] Step S44: Through multiple rounds of training, all parameters are iteratively optimized to minimize the loss function, and finally the fine-tuning parameters of each division point and the offset parameters of the positions on its left and right sides are obtained.

[0161] Multiple rounds of training and fine-tuning: Through multiple rounds of training (epochs), the model gradually adjusts all parameters to minimize the loss function. In each round of training, an optimization algorithm (such as Adam) is used to continuously adjust the three parameters until the change in the loss function approaches zero or reaches a preset stopping condition. Finally, the three optimal parameters are obtained:

[0162] ;

[0163] ;

[0164] ;

[0165] Where, Based on the partition point b q After fine-tuning, is the learnable parameter that controls the offset parameter on the left side of the partition point after optimization. It is a learnable parameter that controls the offset parameter to the right of the partition point after optimization.

[0166] Step S45: Based on the final optimized fine-tuning parameters and the offset parameters of the positions on the left and right sides, the fine-tuned division points and the offset parameters of the positions on the left and right sides are obtained, and a triangular fuzzy membership function is constructed for each division point to convert each division point into a fuzzy division point, and finally a fuzzy decision tree is obtained.

[0167] The decision tree fuzzification method based on learnable offset parameters can effectively avoid the problems of traditional fuzzy decision trees that rely on heuristic search, such as easy falling into local optimality, difficulty in parallel training, and poor generalization. By realizing differentiable and learnable modeling of the left and right offsets of the partition points, it can significantly improve the training efficiency, robustness and generalization ability of the model.

[0168] Step S5: According to the obtained fuzzy decision tree, aggregate the membership of each sample containing multiple partition point paths, normalize it and use it as the weight of the sample in the path, multiply the path weight and the category distribution probability of the sample in the leaf node, and then add them together under the same category to determine the weighted probability vector of the sample for different categories.

[0169] The specific steps include:

[0170] Step S51: After the sample is input into the fuzzy decision tree, the fuzzy partition point membership is calculated: According to the obtained fuzzy decision tree, for any fuzzy partition point, the fuzzy membership of the sample on the fuzzy partition point is calculated using the triangular fuzzy membership function. According to the obtained optimal parameters, the fine-tuned partition point is calculated. , the offset parameter of its left position and the offset parameter to the right :

[0171] ;

[0172] ;

[0173] ;

[0174] For each sample x, starting from the root node, traverse all paths with non-zero membership to the sample (paths are fuzzy, so a sample may belong to multiple paths). , calculate the fuzzy membership of the sample at the node based on the following triangular fuzzy membership function:

[0175] ;

[0176] Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the triangular fuzzy membership function corresponding to each non-leaf node q in the final fuzzy decision tree, Based on the division point b after fine-tuning, is the offset parameter of the left position of the partition point that can be learned after optimization, It is the offset parameter of the right side of the partition point that can be learned after optimization.

[0177] Specifically, when hour,

[0178] ;

[0179] when hour,

[0180] ;

[0181] Step S52: Calculate path weight: Aggregate the comprehensive membership values ​​of the samples on each path and normalize them to obtain the weight value of the sample on each path: Use T-norm to aggregate the membership of all nodes on the path. The comprehensive membership of sample x on each path from the root node to the leaf node is expressed as:

[0182] ;

[0183] Where, represents the weight of sample x in the mth path, d is the path depth, is the membership of sample x to the jth partition point under the final fuzzy decision tree path m;

[0184] Normalized path weight Expressed as:

[0185] ;

[0186] Where M is the number of paths, is the normalized weight of the sample x in the mth (m=1,2,…,M) path of the final fuzzy decision tree.

[0187] Step S53: Output weighted category probability vector: multiply the path weight and the category distribution probability of the sample in the leaf node, and then add them together under the same category to determine the probability vector of the sample for different categories ,in, represents the health probability, The probability of disease is:

[0188] The weighted classification probability formula for sample x is:

[0189] ;

[0190] In the formula, t represents the two categories of healthy or sick, t=0 or 1, It represents the category distribution probability of the samples in the leaf nodes under the mth path, that is, the category distribution probability of the samples in the leaf nodes in the corresponding clear decision tree.

[0191] Step S6: Introduce a three-branch decision-making mechanism to determine whether the weighted probability vector falls into the interval as the evaluation result, and use the size of the weighted probability vector corresponding to the evaluation result as the disease risk value to construct an interpretable three-branch intelligent evaluation model.

[0192] Specifically, we first set a confidence threshold ξ (0.5 < ξ < 1) based on expert opinions to define the boundary between "clear" and "uncertain". The three decisions are as follows:

[0193] when When , the output is determined, including: and When the output result is "illness", when and The output result is "healthy";

[0194] when Output result "uncertain & disease assessment value ”.

[0195] In this embodiment, the three-branch intelligent evaluation method for early Parkinson's disease under multi-task remote data also includes generating a model decision path report and a visual auxiliary display. The model decision path report is an interpretability report.

[0196] Model Decision Path Report: Extract the weight of a sample (one test subject) across all paths and select the top-K paths with the highest weights for the disease category as the key paths for the model's decision. This report integrates all key paths, node contributions, and membership to generate a structured, interpretable evaluation report.

[0197] The report content includes: the confidence level and final diagnostic conclusion determined based on expert opinions; the direction and affiliation of the sample in each path; and the role of each key feature in the judgment.

[0198] Visualization auxiliary display: Figure 4 As shown, the position of the patient's characteristics in the fuzzy division interval, the path weight, and the final judgment result are graphically displayed to facilitate doctors to quickly understand the model behavior and decision-making basis.

[0199] In this embodiment, an interpretable three-limb intelligent assessment device for early Parkinson's disease under multi-task remote data is also provided. It should be understood that the device is used to execute a method for constructing an interpretable three-limb intelligent assessment model and to execute an interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data.

[0200] like Figure 5As shown, it specifically includes a multi-task remote data acquisition module for executing a method for constructing an interpretable three-branch intelligent evaluation model, a multi-task remote data processing module, a model construction module, and a model training module and a disease assessment module for executing an interpretable three-branch intelligent evaluation method for early Parkinson's disease under multi-task remote data.

[0201] The multi-task remote data acquisition module collects multi-task remote data from early-stage Parkinson's patients via remote data sources or smart devices. This multi-task remote data includes any of the following: multi-task gait data collected using insoles with pressure sensors during normal walking, walking while listening to rhythmic music, and walking while communicating with others; multi-task speech data collected using a microphone (such as a mobile phone or medical-grade sound pickup device) during vowel and consonant pronunciation tasks; and multi-task hand-drawn image data collected using a smart tablet using templates of handwritten spirals, winding lines, and simple shapes. The multi-task remote data processing module then performs manual feature extraction, physiological feature de-correlation, cross-task feature distribution alignment, and data normalization on this multi-task remote data, ultimately generating a multi-task remote dataset with a unified structure and coordinated features.

[0202] The model construction module is specifically divided into: the first construction module: constructing a clear decision tree. Use the C4.5 algorithm to construct an initial clear decision tree for the processed multi-task remote data set; the second construction module: fuzzification of the decision tree. Take each division point in the clear decision tree as the initial center of the triangular fuzzy membership function, and introduce learnable offset parameters for its position on its left and right sides respectively. Through the forward propagation and backpropagation mechanisms, the fine-tuned division point position and the offset parameters of the corresponding left and right positions are combined to enable the model to adaptively adjust the fuzzy range of the division point. Construct a triangular fuzzy membership function based on the optimized parameters to achieve the fuzzification of the clear decision tree. The third construction module: construct an intelligent evaluation model based on a three-branch fuzzy decision tree. According to the obtained fuzzy decision tree, calculate the probability vector of the sample in each category. By introducing a three-branch decision mechanism, an intelligent evaluation model based on a three-branch fuzzy decision tree is constructed.

[0203] The model training module refers to training the three-branch intelligent evaluation model of interpretability through a multi-task remote data set to meet the usage standards.

[0204] The disease assessment module refers to inputting the multi-task remote data of a new patient, and using an intelligent assessment model based on a three-branch fuzzy decision tree to output assessment results, model decision path reports and visual auxiliary displays.

[0205] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to equivalent embodiments using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. An interpretable three-branch intelligent assessment method for early Parkinson's disease using multi-task remote data, characterized by: The steps include: Collect multi-task remote data to be tested, input it into the interpretable three-branch intelligent evaluation model for processing and analysis, and obtain evaluation results and interpretability reports; When the assessment result clearly indicates whether the patient is sick, the output is directly healthy or sick; otherwise, the output is uncertain and the risk of illness is further assessed; The method for constructing an explainable three-branch intelligent evaluation model includes the following steps: Step S1: collect and process multi-task remote data to obtain a multi-task remote dataset; Step S2: Use the multi-task remote dataset as a sample set to input into the C4.5 algorithm to build a clear decision tree; Step S3: Using several non-leaf nodes in the clear decision tree as partition points, introducing a learnable fine-tuning parameter for each partition point, and introducing a learnable offset parameter for each partition point on both sides; Step S4: Input each sample in the sample set into the clear decision tree after step S3, and use the minimization of the classification error function as the loss function to iteratively optimize the fine-tuning parameters of each division point position and the offset parameters of the positions on the left and right sides of each division point. By constructing a triangular fuzzy membership function for all the optimized division points and the offset parameters of the positions on the left and right sides, the transformation from the clear decision tree to the fuzzy decision tree is achieved; Step S5: Based on the obtained fuzzy decision tree, aggregate the membership of each sample on a path with multiple partition points, normalize it and use it as the weight of the sample on the path, multiply the path weight by the category distribution probability of the sample in the leaf node, and add them together for the same category to determine the weighted probability vector of the sample for different categories; Step S6: Introduce a three-branch decision-making mechanism to determine whether the weighted probability vector falls into the interval as the evaluation result, and use the size of the weighted probability vector corresponding to the evaluation result as the disease risk value to construct an interpretable three-branch intelligent evaluation model.

2. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1 is characterized in that: In the implementation process of step S1, the following steps are included: Step S11: Collection of multi-task remote data: collecting multi-task remote data of early Parkinson's patients through remote data sources or smart devices; Step S12: Processing of multi-task remote data: including manual feature extraction, physiological feature decorrelation, cross-task feature distribution alignment and data normalization in sequence, and finally obtaining a multi-task remote dataset with unified structure and coordinated features.

3. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1 is characterized in that: In the implementation process of step S2, the following steps are included: Step S21: performing dichotomy processing on all features of the nodes of the clear decision tree to be established; Step S22: Using the multi-task remote dataset as the sample set input, a clear decision tree is constructed by running the C4.5 algorithm; Step S23: Use a post-pruning strategy to prevent overfitting of the clear decision tree.

4. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1 is characterized in that: In step S3, the expression for adjusting the position of the division point based on the fine-tuning parameter is as follows: ； Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the value of the split point of the original clear decision tree, is the dividing point after fine-tuning, is a learnable fine-tuning parameter; Offset parameter of the left position of the division point and the offset parameter to the right of the division point They are represented by the following ways: ； ； Where, is a learnable parameter that controls the offset parameter of the left position of the partition point. is a learnable parameter that controls the offset parameter of the right side of the partition point. and The initial value of is 0.

5. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1 is characterized in that: The implementation process of step S4 includes the following steps: Step S41: Input each sample in the multi-task remote dataset into the clear decision tree after step S3 for forward propagation, passing through each partition point in turn, and constructing a triangular fuzzy membership function based on each partition point and the offset parameters on its left and right sides to obtain the membership degree of each sample to different partition points; Step S42: Aggregate the membership of each sample on a path with multiple partition points, normalize it, and use it as the weight of the sample on the path. Multiply the path weight by the category distribution probability of the sample in the leaf node, and then add them together for the same category to determine the weighted probability vector of the sample for different categories. Select the category with the largest weighted probability value as the predicted label of the sample. Step S43: Use the error between the true label and the predicted label of the sample to construct a classification error loss function, and optimize three parameters through back propagation and gradient update: the fine-tuning parameter of the partition point, the offset parameter of the position to the left of the partition point, and the offset parameter of the position to the right of the partition point; Step S44: Through multiple rounds of training, all parameters are iteratively optimized to minimize the loss function, and finally the fine-tuning parameters of each division point and the offset parameters of the positions on its left and right sides are obtained; Step S45: Based on the final optimized fine-tuning parameters and the offset parameters of the positions on the left and right sides, the fine-tuned division points and the offset parameters of the positions on the left and right sides are obtained, and a triangular fuzzy membership function is constructed for each division point to convert each division point into a fuzzy division point, and finally a fuzzy decision tree is obtained.

6. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 5, characterized in that: In step S41, the triplet of triangular fuzzy membership function is expressed as , therefore, the triangular fuzzy membership function corresponding to each fuzzy partition point is: ； Where q=1, 2, ..., Q, Q is the number of non-leaf nodes, is the fine-tuned dividing point, is the offset parameter of the left position of the learnable partition point, is the learnable offset parameter to the right of the partition point, and x is a sample in the multi-task long-range dataset; when hour, ； when hour, 。 7. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 5, characterized in that: In step S42, the membership of the aggregated sample x to all partition points on each path is expressed as: ； Where w m (x) represents the weight of sample x in the mth path, d is the path depth, μ mj (x) is the membership of sample x to the jth partition point in path m; Normalized path weight W m Expressed as: ； Where M is the number of paths, W is m is the normalized weight of sample x in the mth path, m=1,2,…,M; The formula for the weighted probability vector (P0, P1) of sample x for different categories is as follows: ； Where M is the number of paths, t represents the two categories of healthy or sick, t = 0 or 1, Represents the category distribution probability of the samples in the leaf nodes under the mth path, that is, the category distribution probability of the samples in the leaf nodes in the corresponding clear decision tree; The predicted label t of sample x is calculated using the maximum weighted probability formula: Predicted label t=max t P t ; Where t = 0 or 1 represents the two categories of healthy and sick.

8. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1 is characterized in that: The implementation process of step S5 includes the following steps: Step S51: After the sample is input into the fuzzy decision tree, the fuzzy partition point membership is calculated: according to the obtained fuzzy decision tree, for any fuzzy partition point, the fuzzy membership of the sample at the fuzzy partition point is calculated using the triangular fuzzy membership function; Step S52: Calculate path weight: Aggregate the comprehensive membership values ​​of the samples on each path and normalize them to obtain the weight value of the samples on each path; Step S53: Output weighted category probability vector: multiply the path weight and the category distribution probability of the sample in the leaf node, and then add them together under the same category to determine the probability vector of the sample for different categories ,in, represents the health probability, Represents the probability of disease.

9. The interpretable three-limb intelligent assessment method for early Parkinson's disease under multi-task remote data according to claim 1, characterized in that: During the implementation of step S6, the three decision-making mechanisms are: when , then the output is a certain result, that is, when and When , the output result is: sick, when and When , the output result is: healthy; when When , the output result is: uncertain, and the disease assessment value ; Among them, the confidence threshold ξ range is: 0.5<ξ<1, represents the health probability.

Images