Parkinson's disease patient subtype intelligent typing method based on machine learning

Abstract

The invention discloses a Parkinson's disease patient subtype intelligent typing method based on machine learning, and the method comprises the steps: integrating the multi-dimensional data of a patient, including demographic information, motor and non-motor symptoms, iconography and biomarkers, and utilizing a technical means of combining unsupervised learning and supervised learning, and carrying out the subtype intelligent typing of the Parkinson's disease patient. The Parkinson's disease classification method has the advantages that fine layering of Parkinson's disease patients is achieved, the limitation problems of roughness and single dimension of a traditional classification method are solved, disease heterogeneity can be revealed more accurately, and a scientific basis is provided for formulating an individualized treatment strategy; the disease subtypes with different clinical features and prognosis differences are recognized through clustering analysis, then the supervised learning model is used for conducting rapid and accurate subtype classification on new patients, the classification result is integrated into a clinical decision support system, comprehensive and personalized patient management suggestions are provided for doctors, and the life quality of the patients is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of Parkinson's disease, and in particular to a Parkinson's disease patient subtype intelligent classification method based on machine learning. BACKGROUND

[0002] Parkinson's disease has high heterogeneity in clinical manifestations and progression. The age of onset, main symptoms, disease progression rate and complications of different patients are different. Traditionally, clinicians roughly divide Parkinson's disease into tremor-dominant type (tremor type) and bradykinesia / gait instability-dominant type (rigidity type or PIGD type) according to the main motor symptoms; there are also early-onset and late-onset according to the age of onset, or benign (slow progression) and malignant (fast progression) according to the progression rate. However, these classification methods are relatively rough, often only considering a single dimension, and cannot fully reflect the differences between patients in multiple aspects. For example, patients with the same tremor type may have different cognitive functions and progression rates. In addition, in a group of patients with "benign" progression, there may also be subgroups with different pathological mechanisms. Therefore, there is an urgent need for a data-driven method to automatically identify more detailed and scientific subtypes based on the multi-modal data of patients to support accurate prognosis evaluation and the development of individualized treatment strategies.

[0003] Large-scale cohort studies (such as PPMI) have accumulated longitudinal clinical scores, neuropsychological assessments, imaging and biomarker data of Parkinson's patients, providing an important basis for data-driven subtype division. In recent years, researchers have attempted to apply unsupervised machine learning (such as K-means clustering, hierarchical clustering, Gaussian Mixture Models, Self-Organizing Maps, etc.) to cluster analysis of Parkinson's patients, and found that more detailed subtypes can be divided than traditional classification methods. For example, a study based on 294 patients combined with years of follow-up data identified three disease subtypes with different progression rates: slow progression, moderate progression and rapid progression. Patients in the rapid progression group deteriorate the fastest in motor and cognitive aspects. Other studies focus on baseline data and divide subtypes such as "tremor is dominant but progression is slow" and "non-tremor but fast progression with cognitive impairment". These subtypes are closely related to prognosis and complication risk. For example, tremor-dominant type usually has a relatively good prognosis, while patients with postural instability and gait difficulty (PIGD) tend to progress faster and are more prone to cognitive impairment.

[0004] However, different studies have different numbers and characteristics of subtypes due to differences in data and methods. Currently, there is no universally accepted standard for the classification of Parkinson's disease. It still faces challenges to use the clustering results of machine learning in clinical decision-making, such as ensuring the explainability of these groups (ensuring that doctors understand the characteristics of each subtype), stability (ensuring that similar classification results can be obtained on different data sets), and predictability for individual patients (new patients can be reliably classified into a certain subtype). With the development of artificial intelligence technology, especially the combination of clustering and classification methods (first find subtypes through unsupervised learning, then train a classifier to determine new cases), it is expected to realize intelligent stratified management of Parkinson's patients in the clinic. This will provide the basis for individualized diagnosis and treatment, such as customizing treatment plans for different subtypes of patients, predicting disease progression, and selecting appropriate clinical trials. SUMMARY

[0005] In order to solve the problems of rough classification of Parkinson's patients in the prior art and inability to accurately guide individualized treatment, the present application provides an intelligent classification method for Parkinson's disease patient subtypes based on machine learning, which uses multi-dimensional data of patients and combines unsupervised and supervised learning techniques to achieve objective stratification of Parkinson's disease patients, thereby revealing disease heterogeneity and predicting the prognosis of different stratified patients.

[0006] To achieve the above object, the present application provides an intelligent classification method for Parkinson's disease patient subtypes based on machine learning, comprising the following steps: S1, establishing a multi-dimensional data set of Parkinson's disease patients, the multi-dimensional data set comprising demographic information, motor symptom scores, non-motor symptom indicators, imaging parameters and biomarkers, and containing longitudinal follow-up information; S2, preprocessing and standardizing the multi-dimensional data, including feature standardization, categorical feature encoding, missing value processing and feature extraction; S3, using an unsupervised clustering algorithm to perform clustering analysis on the preprocessed data, dividing the patients into multiple disease subtypes, and determining the optimal number of clusters; S4, characterizing each clustering subtype and describing its clinical characteristics; S5, using longitudinal follow-up data to verify the prognosis differences between subtypes; S6, constructing a supervised learning classification model based on the clustering results, the supervised learning classification model being used for subtype classification of new patients; S7, integrating the supervised classification model into a clinical decision support system, inputting new patient data, and outputting subtype classification and prognosis information.

[0007] As an improved scheme of the present application, the multi-dimensional data comprises: Basic demographic information includes: age, gender, disease duration; Motor symptom scores include: tremor score, rigidity score, gait score in UPDRS-III; Non-motor symptom indicators include: olfactory test score, MoCA cognitive score, depression scale score; Imaging parameters include: quantitative value of brain dopamine transporter SPECT, MRI brain atrophy pattern; Biomarkers include: concentration of alpha-synuclein, concentration of neurofilament light chain in cerebrospinal fluid or blood.

[0008] As an improved scheme of the application, the preprocessing step comprises: Mean-variance standardization is performed on the quantitative features; One-hot encoding or binning is performed on the classification features; Principal component analysis (PCA) or autoencoder is used for feature dimension reduction.

[0009] As an improved scheme of the application, the unsupervised clustering algorithm comprises at least one of K-means, hierarchical clustering, and Gaussian mixture model, and the optimal number of clusters is determined by the silhouette coefficient or the Calinski-Harabasz index.

[0010] As an improved scheme of the application, the clustering analysis adopts a consensus clustering method, comprising multiple random sampling clustering of data, and taking the classification with the highest consistency as the final division.

[0011] As an improved scheme of the application, the subtype feature characterization step comprises using statistical tests to identify significantly different variables in each subtype, thereby defining the clinical feature combination of the subtype.

[0012] As an improved scheme of the application, the prognosis verification step comprises comparing at least one of the UPDRS score growth rate, the fall incidence, and the cognitive impairment incidence of patients in each subtype during follow-up, and using statistical tests to verify the significance of the difference.

[0013] As an improved scheme of the application, the supervised learning classification model comprises at least one of decision tree, support vector machine, random forest, or neural network, and the model parameters are adjusted through cross-validation.

[0014] As an improved scheme of the application, the output content of the clinical decision support system comprises: The subtype to which the patient belongs; The clinical feature description of the subtype; Prognosis risk assessment; Individualized treatment recommendation.

[0015] As an improved scheme of the present application, the multi-dimensional data set adopts a large-scale Parkinson's disease cohort study database, including a PPMI Parkinson's disease progression marker plan database.

[0016] The present application has the following advantages: compared with the prior art, the present application provides a machine learning-based intelligent subtyping method for Parkinson's disease patients, which integrates multi-dimensional data of patients, including demographic information, motor and non-motor symptoms, imaging and biomarkers, and uses a combination of unsupervised and supervised learning techniques to achieve fine stratification of Parkinson's disease patients. The present application not only improves the limitations of traditional subtyping methods, such as roughness and single dimension, but also more accurately reveals the heterogeneity of the disease, providing a scientific basis for the development of individualized treatment strategies. The present application identifies disease subtypes with different clinical characteristics and prognosis differences through cluster analysis, and then uses a supervised learning model to quickly and accurately classify new patients into subtypes, and integrates the classification results into a clinical decision support system to provide comprehensive and personalized patient management recommendations for doctors, improving the quality of life of patients. BRIEF DESCRIPTION OF DRAWINGS

[0017] Fig. 1 A flowchart of the method of the present application; Fig. 2 A schematic diagram of the Parkinson's disease patient subtypes obtained by cluster analysis of the present application; Fig. 3 A schematic diagram of the characteristic profiles of each subtype of the present application. DETAILED DESCRIPTION

[0018] In order to more clearly illustrate the present application, the present application will be further described below in conjunction with the accompanying drawings.

[0019] In the following description, the example details are given in order to provide a more in-depth understanding of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. It should be understood that the described specific embodiments are only used to explain the present application, and are not used to limit the present application.

[0020] It should be understood that when the terms "comprise" and / or "include" are used in the present specification, it means that the features, integers, steps, operations, elements or components exist, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components or combinations thereof.

[0021] Please refer to Figs. 1-3 The machine learning-based intelligent subtyping method for Parkinson's disease patients of the present application comprises the following steps: S1, a Parkinson's disease patient multidimensional dataset is established, the multidimensional dataset includes demographic information, motor symptom scores, non-motor symptom indicators, imaging parameters and biomarkers, and contains longitudinal follow-up information; S2, the multidimensional data is preprocessed and standardized, including feature standardization, categorical feature coding, missing value processing and feature extraction; S3, an unsupervised clustering algorithm is used to perform clustering analysis on the preprocessed data, and the patients are divided into multiple disease subtypes, and the optimal cluster number is determined; S4, the characteristics of each cluster subtype are described; S5, the prognosis difference between each subtype is verified by using longitudinal follow-up data; S6, a supervised learning classification model is constructed based on the clustering results, and the supervised learning classification model is used for subtype classification of new patients; S7, the supervised classification model is integrated into a clinical decision support system, new patient data is input, and subtype classification and prognosis information are output.

[0022] Firstly, the Parkinson's disease patient multidimensional dataset is collected, including demographic information, motor symptom scores, non-motor symptom indicators, imaging parameters and biomarkers, and contains longitudinal follow-up information, which can more accurately reflect the disease state and progression of Parkinson's disease patients based on the above multidimensional dataset, and provides a rich data basis for subtype division, then the collected multidimensional dataset is preprocessed and standardized to ensure the data quality and integrity of the multidimensional dataset, which can be processed by feature standardization, categorical feature coding, missing value processing and feature extraction, then the preprocessed data is clustered by an unsupervised clustering algorithm, and the patients are divided into multiple disease subtypes, and the optimal cluster number is determined, then the mean and distribution of the characteristics of each cluster subtype are calculated, and a clinically interpretable description is given, such as the characteristics of subtype 1 may be “young onset, tremor dominant, slow progression, and good cognitive function”, and the characteristics of subtype 2 may be “old onset, prominent limb rigidity and balance disorder, accompanied by significant non-motor symptoms, and fast progression”. Statistical tests can be used to find variables that are significantly different from other subtypes, thereby defining the subtype characteristic combination, further, a supervised learning classification model is constructed based on the clustering results, which is used for subtype classification of new patients, the supervised learning classification model learns the data pattern of the classified patients, and can accurately classify new patients into the corresponding subtype, finally, the trained supervised classification model is integrated into a clinical decision support system, when new patient data is input, the system can quickly output the subtype to which the new patient belongs, the clinical feature description of the subtype, the prognosis risk assessment and individualized treatment suggestion, providing comprehensive and scientific decision-making basis for doctors.

[0023] In the embodiment, the multi-dimensional data set comprises: The basic demographic information comprises: age, gender, disease duration; The motor symptom score comprises: tremor score, rigidity score, gait score in UPDRS-III; The non-motor symptom index comprises: olfactory test score, MoCA cognitive score, depression scale score; The imaging parameters comprise: brain dopamine transporter SPECT quantitative value, MRI brain atrophy pattern; The biomarkers comprise: alpha-synuclein concentration, neurofilament light chain concentration in cerebrospinal fluid or blood; The above multi-dimensional data is closely related to the treatment effect and disease progression of Parkinson's disease patients, and can comprehensively reflect the disease state of the patients from multiple dimensions, wherein the age, gender and disease duration in the basic demographic information are important basic factors affecting the occurrence and development of Parkinson's disease, and patients of different ages, genders and disease durations may have differences in disease performance and progression; the tremor score, rigidity score and gait score in the motor symptom score can directly reflect the degree of impairment of the motor function of the patients, and are important indicators for evaluating the severity of Parkinson's disease; the non-motor symptom indicators such as olfactory test score, MoCA cognitive score and depression scale score can detect abnormalities in the olfaction, cognition and emotion of the patients, and these non-motor symptoms are also common in Parkinson's disease patients and have an important impact on the quality of life of the patients; the brain dopamine transporter SPECT quantitative value and MRI brain atrophy pattern in the imaging parameters can directly show the structural and functional changes of the brain of the patients, and provide important imaging basis for the diagnosis and typing of the disease. The alpha-synuclein concentration and neurofilament light chain concentration in cerebrospinal fluid or blood in the biomarkers can reflect the damage and metabolism of the neural cells in the brain of the patients, and play an important role in early diagnosis and prognosis evaluation of the disease; through comprehensive analysis of the above multi-dimensional data, the subtypes of Parkinson's disease patients can be more accurately divided, which provides strong support for individualized treatment, and the multi-dimensional data set contains longitudinal follow-up information to evaluate the progression rate and ensure data quality and integrity.

[0024] As an improved scheme of the present application, the preprocessing step comprises: The quantitative features are subjected to mean-variance standardization; The classification features are subjected to one-hot encoding or binning processing; Principal component analysis (PCA) or autoencoder is used for feature dimension reduction; In the preprocessing step, mean-variance standardization of quantitative features can eliminate the dimensional differences between different features, making each feature comparable in subsequent analysis. For example, for features with different magnitudes such as age and tremor score, after standardization, they will be in a similar numerical range, thus avoiding the bias of clustering results caused by different dimensions. Second, one-hot encoding or binning is performed on the classification features. One-hot encoding can convert categorical variables into binary vectors, making classification features better handled by machine learning algorithms. For example, the gender classification feature can be converted into a vector with only 0 and 1 values after one-hot encoding. Binning is a process of dividing continuous classification features into intervals, converting them into discrete classification features, which helps to simplify the data structure and improve the stability of the model. Finally, principal component analysis (PCA) or autoencoder is used for feature dimension reduction. PCA projects the original high-dimensional data into a low-dimensional space through linear transformation, retaining the most important feature components of the data. It reduces the data dimension while preserving the main information of the data as much as possible. For example, when dealing with high-dimensional data containing numerous imaging parameters and biomarkers, PCA can extract the principal components that best represent the data features, reducing the computational complexity. Autoencoder is a dimension reduction method based on neural networks. Through the encoding-decoding process, it learns the low-dimensional representation of the data, which can capture the nonlinear relationships in the data, further optimize the data structure, and provide a more suitable data format for subsequent clustering analysis, improving the clustering effect and model performance.

[0025] In this embodiment, the unsupervised clustering algorithm includes at least one of K-means, hierarchical clustering, and Gaussian mixture model, and the optimal number of clusters is determined by the silhouette coefficient or Calinski-Harabasz index; the unsupervised clustering algorithm plays a key role in the subtyping of Parkinson's disease patients, and the K-means algorithm can quickly divide the patient data into K clusters due to its simple and efficient characteristics, and through iterative optimization of intra-cluster distance and inter-cluster distance, patients in the same cluster have similar characteristics, and patients in different clusters have significant differences in characteristics; hierarchical clustering can reveal the hierarchical relationship between patients by constructing a tree diagram to show the hierarchical structure of the data, and can gradually merge or split patients into different clusters according to different distance measurement methods such as Euclidean distance and Manhattan distance, thereby revealing the hierarchical relationship between patients; the Gaussian mixture model assumes that the data is mixed by multiple Gaussian distributions, and by estimating the parameters of each Gaussian distribution, the patients are assigned to the cluster corresponding to the most likely Gaussian distribution, which can handle more complex data distribution situations; in actual application, in order to determine the optimal number of clusters, evaluation indicators such as the silhouette coefficient or Calinski-Harabasz index can be used. The silhouette coefficient comprehensively considers the tightness within the cluster and the separation between the clusters, and the value range is between -1 and 1, and the larger the value, the better the clustering effect. The Calinski-Harabasz index evaluates the clustering effect by calculating the ratio of inter-cluster dispersion to intra-cluster dispersion, and the larger the ratio, the better the clustering effect, and by comparing the evaluation indicator values under different numbers of clusters, the optimal number of clusters can be selected to more accurately subdivide the subtypes of Parkinson's disease patients; and the above unsupervised clustering algorithms can also be combined for use, for example, the K-means algorithm and hierarchical clustering can be combined, the K-means algorithm is used for preliminary clustering to obtain several relatively compact clusters, and then hierarchical clustering is used to further merge or split these clusters to discover more detailed patient subtype structures, fully exerting the advantages of different algorithms and improving the accuracy and stability of clustering; in addition, the Gaussian mixture model and the K-means algorithm can be combined, the Gaussian mixture model is used to probabilistically model the data and estimate the distribution of the data, and then the K-means algorithm is used for clustering based on the estimated distribution, thereby better handling complex data distribution and improving clustering effect, and different unsupervised clustering algorithms are flexibly selected and combined according to the characteristics of the data and actual needs to achieve the most correct subtyping.

[0026] In this embodiment, the clustering analysis adopts the consensus clustering method, which includes multiple random sampling clustering of the data, and taking the classification with the highest consistency as the final division; the consensus clustering method has significant advantages, which can effectively avoid the deviation of single clustering due to data randomness or local characteristics. In specific operation, first, the multidimensional data set is randomly sampled multiple times, and after each sampling, the selected unsupervised clustering algorithm is used for clustering analysis. Since the data of each sampling is not exactly the same, the clustering results obtained will also have certain differences. Subsequently, through a specific algorithm, the different clustering results are comprehensively evaluated, and the consistency degree between them is calculated. The classification with the highest consistency means that in different sampling conditions, the stability of patients being divided into the corresponding subtypes is the strongest, so it is taken as the final division result, which improves the accuracy and reliability of the subtype division, so that the divided subtypes can better reflect the true disease characteristics and internal rules of Parkinson's disease patients; please refer to Fig. 2 , the title of the coordinate axis is used to represent the relative position and similarity of the patient sample in the overall feature space, and does not directly correspond to any specific clinical feature, each point represents a patient sample, and the color is distinguished according to the clustering allocation type.

[0027] In this embodiment, the subtype feature description step includes using statistical tests to identify significantly different variables in each subtype, thereby defining the clinical feature combination of the subtype. In the process of subtype feature description, statistical tests play a crucial role. Common statistical test methods include t-test, analysis of variance, chi-square test, etc. The specific choice depends on the type and distribution characteristics of the variable. For example, for continuous variables and normal distribution, t-test can be used to compare the mean difference between different subtypes; if the variable does not meet the normal distribution, non-parametric test methods such as Mann-Whitney U test can be selected. For categorical variables, chi-square test can effectively test the differences in category distribution among subtypes. Through these statistical tests, variables with significant differences among subtypes can be identified, and these variables constitute the clinical feature combination of the subtype. For example, after statistical testing, it is found that subtype A has significant differences with other subtypes in tremor score, olfactory test score, and alpha-synuclein concentration, so "tremor as the main, olfactory decline, and alpha-synuclein concentration increase" can be taken as the clinical feature combination of subtype A. Through statistical testing, the defined clinical feature combination is ensured to be scientific and objective, providing a reliable basis for prognosis verification and new patient classification; please refer to Fig. 3, Radar chart showing the average scores of three different subtypes of Parkinson's disease on four key clinical features: Tremor, Bradykinesia, Gait, and Cognition. Each vertex in the chart represents a feature, and all vertices are connected to form a closed figure to visually demonstrate the distribution of each subtype on different features. Each color represents a subtype, and the intensity of the fill color in each area reflects the concentration or average level of the subtype on the corresponding feature.

[0028] In this embodiment, the prognosis verification step includes comparing at least one of the UPDRS score growth rate, the fall incidence, and the cognitive impairment incidence of patients in each subtype during follow-up, and verifying the significance of the differences using statistical tests. During the prognosis verification process, key indicators such as UPDRS scores, fall incidence, and cognitive impairment assessment results of patients in each subtype are collected through regular follow-up. Taking the UPDRS score growth rate as an example, the change in the score of each patient from baseline to the end of follow-up is divided by the follow-up time to obtain the annualized growth rate. For the fall incidence, the number of falls of each patient during follow-up is divided by the number of follow-up months to convert it into the average monthly fall frequency. Cognitive impairment incidence is determined by MoCA score changes combined with clinical diagnosis. Then, repeated measures ANOVA is used to compare the differences in UPDRS growth rates between different subtypes, negative binomial regression is used to analyze the group differences in fall frequency of each subtype, and Cox proportional hazards model is used to evaluate the subtype differences in the time of cognitive impairment occurrence. All statistical tests set the two-sided test level to 0.05 and correct the false positive risk brought by multiple comparisons. The results show that the average annual UPDRS score growth rate of subtype B patients is significantly increased by 2.3 points compared to subtype A (p=0.004), and the average monthly fall frequency of subtype B is 1.8 times that of subtype C (IRR=1.8, 95% CI: 1.3-2.5), while the cumulative incidence of cognitive impairment in subtype C during the 3-year follow-up period is 42%, which is significantly higher than that of other subtypes (HR=2.1, p=0.01). These differences are visualized through Kaplan-Meier survival curves and forest plots, confirming that different subtypes have clinically significant differences in disease progression speed and complication risk, providing quantitative basis for subsequent development of differentiated intervention strategies.

[0029] In this embodiment, the supervised learning classification model includes at least one of a decision tree, a support vector machine, a random forest, or a neural network, and the model parameters are adjusted through cross-validation. The selection of the supervised learning classification model is based on its ability to process complex data patterns and make accurate classifications. The decision tree model divides patients step by step based on feature thresholds by constructing a tree structure, has the advantages of intuitive understanding and strong interpretability, and can clearly show the classification decision process. The support vector machine separates patient data of different subtypes in the feature space by finding the optimal hyperplane, and is particularly good at handling high-dimensional data, and can maintain good classification performance even with small sample sizes. The random forest is composed of multiple decision trees, which integrates the classification results of multiple decision trees through ensemble learning, effectively reduces the risk of overfitting, and improves the stability and accuracy of the model. The neural network has strong non-linear mapping ability and can automatically learn complex features and patterns in the data, and has excellent classification performance for large-scale and high-dimensional Parkinson's disease patient data. In practical applications, the data set is divided into training set and validation set through cross-validation method, and the training and validation process is repeated multiple times. According to the performance of the model on the validation set, such as accuracy, recall rate, F1 value and other indicators, the parameters of the model are adjusted, such as the depth of the decision tree, the kernel function parameters of the support vector machine, the number of trees in the random forest, the number of layers and the number of neurons in the neural network, etc., to optimize the performance of the model and ensure that the supervised learning classification model can accurately classify new patients into the corresponding subtype.

[0030] In this embodiment, the output content of the clinical decision support system includes: the subtype to which the patient belongs; clinical feature description of the subtype; prognostic risk assessment; individualized treatment recommendation; wherein the subtype to which the patient belongs is obtained by analyzing the new patient data through the supervised learning classification model, and the specific subtype category to which the patient belongs is determined; the clinical feature description of the subtype is summarized based on the significantly different variables identified by statistical testing in the subtype feature description step, and the unique performance of the subtype in tremor, olfaction, cognition, biomarkers, etc. is clearly presented; the prognostic risk assessment is based on the comparison results of the UPDRS score growth rate, the incidence of falls, the incidence of cognitive impairment, etc. of patients in each subtype during follow-up in the prognosis verification step, combined with the significance of the difference of statistical testing, to provide quantitative assessment of the patient's disease progression speed and complication risk; the individualized treatment recommendation considers the patient's subtype, clinical feature description, and prognostic risk assessment, and refers to existing clinical treatment guidelines and research evidence to provide scientific basis for doctors to develop personalized treatment plans for the patient, covering drug treatment plan selection and adjustment, rehabilitation training focus direction, specific measures of lifestyle intervention, etc., aiming to improve treatment effect and improve patient's quality of life.

[0031] In this embodiment, the multi-dimensional dataset adopts the large-scale Parkinson's Disease Cohort Study database, including the PPMI Parkinson's Disease Progression Markers Initiative database, which is an internationally recognized core resource for Parkinson's disease research, integrating multi-modal data of more than 1500 patients from 33 clinical centers around the world, covering the whole course tracking from early non-drug patients to middle and late stage patients. Its data dimensions include: clinical evaluation scales collected every 6 months (UPDRS, H&Y staging), multi-modal neuroimaging (DaTSCAN SPECT, 3T MRI), biological sample library (cerebrospinal fluid alpha-synuclein, serum neurofilament light chain), wearable device monitored motor function data (gait speed, tremor frequency) and genomics information (LRRK2, GBA and other pathogenic gene mutation detection). The database adopts standardized data collection process, and through the audit of the central quality control committee to ensure the reliability of the data, and sets strict data access rights to protect the privacy of patients. In addition, the PPMI database realizes data sharing with international platforms such as EuroPark, China Parkinson's Disease Research Consortium (CPDRC), etc., forming the largest Parkinson's disease dataset covering different races and different treatment stages in the world, providing a solid data foundation for the multi-dimensional data-driven subtype analysis in this embodiment.

[0032] Embodiment 1: Subtype clustering and verification The subjects were selected from 421 Parkinson's disease patients in the PPMI database, with 3-year follow-up data. The features at baseline of each case were extracted, including: age, disease duration, UPDRS-III sub-item scores (tremor, rigidity, bradykinesia, posture balance, etc.), non-motor and daily life scores of UPDRS-I and II, MoCA cognitive score, Geriatric Depression Scale score, UPSIT olfactory score, and hippocampal volume measured by brain MRI, radioligand uptake ratio of DAT scan, etc., a total of about 50-dimensional features. First, principal component analysis (PCA) was performed on these features, and the first 10 principal components (explaining ~80% of the variance) were taken as the clustering input to reduce the dimension and noise effect. Then hierarchical agglomerative clustering (Ward minimum variance method) was used, and the patients were divided into 3 clusters by observing the dendrogram and selecting the silhouette coefficient. The results are: Cluster1 has 150 people, Cluster2 has 180 people, and Cluster3 has 91 people. Analyze the mean values of the characteristics of each cluster: Cluster1 patients are younger (average age 55), have high tremor scores and low posture instability scores, normal cognitive function, and the smallest 3-year UPDRS-III deterioration rate, which can be defined as the "tremor dominant / slow progression" subtype; Cluster2 patients are of medium age (62 years old), have moderate motor symptoms, and have obvious non-motor symptoms (such as constipation and reduced olfaction), with moderate progression in 3 years, which can be defined as "intermediate type"; Cluster3 patients are the oldest (68 years old), have the highest scores for posture instability and gait difficulty, have a low baseline cognitive MoCA, and are accompanied by significant depression and RBD symptoms, with significant deterioration in UPDRS-III and cognition in 3 years, which can be defined as "rapid progression type", similar to the fast progression subtype reported in the literature. Statistical verification shows that the proportion of falls and the incidence of cognitive impairment in Cluster3 at 3 years are significantly higher than those in Cluster1 and 2 (p<0.01), while the motor scores of Cluster1 have almost no change and no one develops dementia. This proves that these clustering subtypes have clear differences in prognosis. Then, to facilitate the classification of new patients, the clustering results are used as labels to train a random forest classifier, and the classification accuracy is about 85% under 10-fold cross-validation. This example shows that the method of the present application can effectively distinguish Parkinson's disease subtypes with different clinical characteristics and outcomes, providing a basis for personalized management.

[0033] Example 2: Clinical application scenarios A newly diagnosed Parkinson's disease patient, male, 65 years old. The main manifestations are right hand resting tremor and writing difficulty, but walking is still stable, mild hyposmia, and no obvious depression. MoCA cognitive score is 26 / 30. The doctor inputs his data into the hierarchical system of the application. After preprocessing, the built-in subtype classification model is used for judgment, and the result shows that the patient belongs to the "tremor dominant / progression slow" subtype (corresponding to Cluster1 of Example 1). The system interface prompts the characteristics of this subtype: usually mild onset, tremor is prominent, motor symptoms progress slowly within 5 years, and cognition remains good; at the same time, it suggests: "routine treatment is enough, pay more attention to tremor control and quality of life, and consider delaying invasive treatment such as deep brain stimulation". The doctor discusses with the patient to formulate a plan mainly based on drug treatment and regular review. For another patient (72 years old, with balance disorder and history of falling at onset, mild cognitive impairment), the system classifies him into the "malignant progression" subtype. The prompt information includes: "patients in this subtype progress faster, pay attention to early gait and cognitive problems. Suggest close follow-up, timely adjustment of drugs and consideration of rehabilitation and cognitive training intervention. If conditions permit, explore early DBS surgery to improve quality of life." The doctor therefore formulates a more aggressive intervention plan for him, including 3-month follow-up, early posture balance training, etc. Through this example, it can be seen that the subtype intelligent typing method of the application can provide differentiated management suggestions for different patients in the clinic, making medical decision-making more accurate. Doctors feedback that this auxiliary system is intuitive and easy to use, the output subtype is consistent with their experience, and the prognosis information and suggestions contained have reference value, improving the clinical confidence.

[0034] Example 3: Research application scenarios In research, the method of the application is also useful. Researchers analyzed the differences in genes between the three subtypes obtained in Example 1. It was found that the carrying rate of LRRK2 gene variation in the "malignant progression" subtype was significantly higher than that in the other two types (20% vs<5%, p<0.05), while the mutation rate of GBA gene in the "tremor dominant" subtype was slightly higher. This suggests that different genetic backgrounds may partially drive the performance of different subtypes. Further, researchers use the hierarchical results of the application to select patients with fast progression subtype for cerebrospinal fluid proteome analysis, and identify some protein markers that may be related to rapid progression, such as increased amyloid precursor protein concentration. As can be seen, the application not only helps in the clinic, but also can be used as a research tool to promote the mechanism research of Parkinson's disease heterogeneity.

[0035] The advantages of the application are: 1. Refine patient stratification and improve prediction accuracy: Through machine learning, patient subgroups that are difficult to identify by traditional experience can be discovered. Differences in the natural history of different subtypes of patients are quantified, and the future trajectory of individual patients can be more accurately predicted. This is much more scientific than giving an average prognosis to all Parkinson's patients with "one size fits all", thus making clinical decision-making more personalized and precise.

[0036] 2. After identifying subtypes, clinicians can develop targeted management plans based on subtype characteristics. For example, for patients with fast progression, more aggressive drug dose adjustment and more frequent rehabilitation training are needed; for patients with slow progression and mainly tremor, tremor control should be focused on and over-treatment should be avoided. Subtype information can also help choose the most appropriate surgical timing or intervention to improve efficacy 3. Optimize clinical trial design: In drug clinical trials, different subtypes of patients may respond differently to new therapies. If pre-stratification is performed, and subgroups are evenly grouped or analyzed for efficacy in the trial, the interpretation and success rate of the trial results can be improved, ensuring the effectiveness of the drug for different subtypes. In particular, for disease-modifying therapies, it may be necessary to focus on fast-progressing subtypes to see drug effects more quickly.

[0037] 4. Assist genetic and biological research: Subtypes divided by machine learning can also be analyzed in combination with genetic data or biomarkers, which may find that certain subtypes are enriched with specific genetic variations or pathological patterns. This will help to understand the differences in pathogenesis of different patient groups. For example, if the plasma NfL level of a certain subtype of patients is significantly high, it suggests that this subtype may have more severe neuronal damage, and the corresponding pathways can be further studied biologically.

[0038] 5. Intelligence and objectivity: The present invention reduces human subjective bias and automatically divides based on data, making the typing more objective. Doctors can get subtype judgments and recommendations by inputting patient data through a friendly software interface, which is simple and easy to use. At the same time, since the model is trained on a large amount of data, its judgment criteria are uniform in the background, which can improve consistency in multi-center applications.

[0039] The above disclosure is only a few specific embodiments of the present application, but the present application is not limited thereto, and any changes that can be thought of by those skilled in the art shall fall within the scope of the present application.

Claims

1.A method for intelligent subtyping of Parkinson's disease patients based on machine learning, characterized in that, The method comprises the following steps: S1, establishing a Parkinson's disease patient multidimensional dataset, the multidimensional dataset comprising demographic information, motor symptom scores, non-motor symptom indicators, imaging parameters and biomarkers, and containing longitudinal follow-up information; S2, pre-processing and standardizing the multidimensional data, comprising feature standardization, categorical feature encoding, missing value processing and feature extraction; S3, using an unsupervised clustering algorithm to perform clustering analysis on the pre-processed data, dividing the patients into multiple disease subtypes, and determining the optimal number of clusters; S4, characterizing the features of each clustering subtype, and describing the clinical characteristics thereof; S5, verifying the prognosis differences between the subtypes by using longitudinal follow-up data; S6, constructing a supervised learning classification model based on the clustering results, the supervised learning classification model being used for subtype classification of new patients; S7, integrating the supervised classification model into a clinical decision support system, inputting new patient data, and outputting subtype classification and prognosis information. 2.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, characterized in that The multidimensional data comprises: Basic demographic information including age, gender, disease duration; Motor symptom scores including tremor score, rigidity score and gait score in UPDRS-III; Non-motor symptom indicators including olfactory test score, MoCA cognitive score and depression scale score; Imaging parameters including brain dopamine transporter SPECT quantitative value and MRI brain atrophy pattern; Biomarkers including α-synuclein concentration and neurofilament light chain concentration in cerebrospinal fluid or blood. 3.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, characterized in that The pre-processing step comprises: Mean-variance standardization of quantitative features; One-hot encoding or binning processing of categorical features; Using principal component analysis (PCA) or autoencoder for feature dimension reduction. 4.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The unsupervised clustering algorithm comprises at least one of K-means, hierarchical clustering and Gaussian mixture model, and the optimal number of clusters is determined by the silhouette coefficient or Calinski-Harabasz index. 5.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The clustering analysis adopts a consensus clustering method, comprising multiple random sampling clustering of the data, and taking the classification with the highest consistency as the final division. 6.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The subtype feature characterization step comprises using statistical tests to identify significantly different variables in each subtype, thereby defining the clinical feature combination of the subtype. 7.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The prognosis verification step comprises comparing at least one of the UPDRS score growth rate, the fall incidence and the cognitive impairment incidence of the patients in each subtype during the follow-up period, and using statistical tests to verify the significance of the differences. 8.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The supervised learning classification model comprises at least one of decision tree, support vector machine, random forest or neural network, and the model parameters are adjusted by cross-validation. 9.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The output content of the clinical decision support system comprises: The subtype to which the patient belongs; Clinical feature description of the subtype; Prognosis risk assessment; Individualized treatment recommendation. 10.The machine learning-based intelligent subtyping method for Parkinson’s disease patients according to claim 1, wherein, The multidimensional dataset adopts a large-scale Parkinson's disease cohort study database, including the PPMI Parkinson's Disease Progression Markers Initiative database.

Images