Parkinson's disease subtype staging inference system based on corneal nerve images and su stain algorithm

Abstract

The application relates to the field of biomedical technology, and discloses a Parkinson disease subtype staging inference system based on a corneal nerve image and a SuStaIn algorithm, which comprises an image data acquisition module, a health control module, an abnormality division module, a SuStaIn model construction module and an output module. The image data acquisition module acquires an original image data set, and extracts core effective features after preprocessing the original image data set. The health control module calculates the Z scores of each feature in the corresponding core effective features of Parkinson disease patients through a standardization method. The abnormality division module divides abnormality grades based on the calculation results of the Z scores, and constructs a feature event matrix. The SuStaIn model construction module constructs a disease progression model based on the feature event matrix, inputs the patient feature Z score matrix, sets related parameters, completes model training, and then determines the optimal subtype number through multi-index evaluation. The output module outputs the model prediction results. The application realizes accurate identification of Parkinson disease patient subtypes and objective inference of disease stages, and provides strong support for early diagnosis, individualized treatment and drug research and development of diseases.

Description

Technical Field

[0001] This invention relates to the field of biomedical technology, specifically to a Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm. Background Technology

[0002] Against the backdrop of an accelerating global aging population, Parkinson's disease, a neurodegenerative movement disorder prevalent among middle-aged and elderly individuals, is becoming a significant challenge threatening human health and public health security. Epidemiological data shows that the number of people suffering from Parkinson's disease globally is increasing at a rate of approximately 10% annually. This not only causes tens of millions of patients to suffer from symptoms such as tremors, muscle rigidity, and bradykinesia, but also places a heavy burden on public health systems worldwide, and imposes long-term care and financial strain on countless families.

[0003] However, current technologies in the clinical assessment and research of Parkinson's disease still have many shortcomings that urgently need to be addressed. At the clinical assessment level, traditional symptom grading relies on the subjective judgment of physicians, and different physicians may yield inaccurate assessments of the same patient. Some invasive diagnostic methods, such as cerebrospinal fluid biopsy and deep brain stimulation electrode implantation, not only cause physical trauma and psychological burden to patients but also carry certain surgical risks. Furthermore, high-end imaging equipment and gene sequencing technology are difficult to popularize in primary healthcare institutions due to their high cost and complex procedures, thus limiting the coverage of disease screening and early diagnosis. At the research and monitoring level, current technologies lack the ability to conduct long-term dynamic monitoring of patients' motor function and neurophysiological indicators, failing to capture subtle changes in the disease progression process. Simultaneously, the statistical methods commonly used in data analysis have limitations, making it difficult to integrate multi-dimensional clinical, imaging, and molecular biological data, and thus unable to construct accurate disease progression models. These technical challenges are intertwined, severely hindering researchers' in-depth understanding of the pathogenesis of Parkinson's disease and restricting the transformation of clinical diagnosis and treatment from "empirical treatment" to "personalized precision treatment". Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a Parkinson's disease subtype staging inference system based on corneal neural imaging and the SuStaIn algorithm. It can accurately identify Parkinson's disease patient subtypes and objectively infer disease stages by combining non-invasive corneal neural imaging with advanced data-driven modeling methods, providing strong support for early diagnosis, personalized treatment and drug development.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm, comprising an image data acquisition module, a healthy control module, an abnormality classification module, a SuStaIn model construction module, and an output module; The image data acquisition module includes an image data acquisition unit and a feature extraction unit. The image data acquisition unit is used to acquire the original image dataset and preprocess the original image dataset to form the final image dataset. The feature extraction unit is used to extract core effective features based on the final image dataset. The healthy control module is used to calculate each feature in the core effective characteristics corresponding to Parkinson's disease patients using a standardized method. Fraction; The anomaly segmentation module is used for... The calculated scores are used to classify anomaly levels and construct a feature event matrix; The SuStaIn model building module is used to construct a disease progression model based on a feature event matrix and input patient features. The model training was completed by setting the score matrix and relevant parameters, and then the optimal number of subtypes was determined by multi-index evaluation. The output module is used to output the model prediction results.

[0006] Preferably, the original image dataset is obtained by connecting to the hospital system through an image data acquisition unit. Confocal microscopy image data of corneas from Parkinson's disease patients; The preprocessing includes: Corneal confocal microscopy images of Parkinson's disease patients and healthy controls were collected according to standard procedures, prioritizing the central corneal region and excluding images of peripheral or non-target areas. The validated ACCMetrics automated tool was used to objectively score the sharpness, contrast, signal-to-noise ratio, and nerve fiber edge sharpness of each image. A low-quality judgment threshold was set based on the objective scores, and images below the low-quality judgment threshold were automatically removed as invalid images. Images at or above the low-quality judgment threshold were sorted from highest to lowest score and selected. The final image dataset is constructed from the images with the highest scores. Preferably, the feature extraction unit includes: The proven and reliable automated tool ACCMetrics is used to perform batch analysis of images of the central corneal region of each eye in the final image dataset, automatically identify the morphology and distribution of corneal nerve fibers, perform quantitative calculations, and extract core effective features. The core effective feature includes: nerve fiber density. Nerve branch density Nerve fiber length Density of the main nerve trunk Nerve fiber area Fractal dimension of nerve fibers ; Preferably, the Fraction The calculation formula is: ; in, Representing the Among the core effective characteristics of Parkinson's disease patients, the first is... The original quantitative value of the feature; The first of the core effective characteristics representing healthy controls The mean of the features; The first of the core effective characteristics representing healthy controls The standard deviation of the feature; Preferably, the classification of abnormal levels , The number of exception levels is set here. ,in Indicates no obvious abnormalities. Indicates mild abnormality. This indicates a severe anomaly, and simultaneously sets the first core effective feature. Item Features The 95th percentile of the score is ; when ,represent That is, there are no obvious abnormalities; when ,represent This indicates a mild abnormality. when ,represent This indicates a severe abnormality. Preferably, the feature event matrix for: ; in, Represents the number of Parkinson's disease patients; Represents the number of core effective features; This represents the number of exception levels. The first Parkinson's patient The first feature corresponds to the first element The rules for selecting values ​​are as follows: ; Preferably, the SuStaIn model building module includes: S1, The feature event matrix and patient characteristics The score matrix serves as the model input data, and the patient features... The fraction matrix is of Fractional matrix; S2. Set core operating parameters, including: number of patients, number of features, number of maximum candidate subtypes, number of MCMC sampling iterations, number of starting points, and number of cross-validation folds; S3. Based on the maximum number of candidate subtypes set in S2, the number of candidate subtypes is trained independently. S4. Extract the cross-validation criteria and log-likelihood value of the model under the number of candidate subtypes; S5. When comparing the cross-validation information criteria and log-likelihood values, select the number of subtypes with the best model performance as the final result. Preferably, S3 includes: S3a. Randomly initialize the order of the occurrence of the level 3 anomalies of the 6 core features in the core effective features, and generate the initial event sequence; S3b: The event order of the feature anomaly level is adjusted iteratively using an optimization algorithm to maximize the probability of the SuStaIn model interpreting the input data. S3c, Perform multiple MCMC samplings on the event locations for each anomaly level, and calculate the mean, variance, and standard deviation of the event locations through multiple iterative samplings, including: S3c.1 Construct a sampling result set from multiple MCMC samplings. The expression is: ; in, Representing the The event location value obtained from the second MCMC sampling. This represents the total number of MCMC samplings. Represents the core effective feature index, Represents the anomaly level index; S3c.2 Calculate the mean of the sampling result event locations. The formula is: ; in, Represents a dynamic index; S3c.3 Calculate the location variance of the sampling result events. The formula is: ; S3c.4 Calculate the standard deviation of the sampling outcome events. The formula is: ; S3c.5. Based on the calculated mean, variance, and standard deviation of the event location, compare them with the preset thresholds for reasonableness of mean, convergence of variance, and convergence of standard deviation, respectively. When the calculated results of the three are all within the corresponding threshold range, the event location sampling results representing the abnormality level of the feature are converged and reasonable, and can be directly used for the posterior probability calculation of S3d; otherwise, it means that the sampling results are unreliable, and the MCMC sampling and the calculation steps of mean, variance, and standard deviation need to be repeated. S3d, Calculate the number of patients belonging to the largest candidate subtype. The posterior probability of each subtype and each disease stage within each subtype is used, and the maximum posterior probability is taken as the subtype classification and disease stage determination result for that patient, including: S3d1: Collect the valid statistics and basic data of the model that passed the S3c.5 test, and construct the input dataset. ; ; S3d2, for the first Patient No. 1, The first subtype, the Each disease stage, posterior probability The formula is: ; in, Represents the likelihood probability, indicating the patient's characteristic event state and Score data and current Subtype The degree of matching of the event sequence in each stage; Represents prior probability. , This represents the number of disease stages corresponding to a single subtype. S3d3, Traverse all patients All subtypes All stages of disease Substitute the values ​​into the S3d.2 formula to calculate the posterior probability and construct the probability set. : ; S3d4, for the first Patients, extract probability set The corresponding maximum posterior probability ,Right now: ; in, To determine the final subtype classification for this patient, This is the stage of disease progression in this patient; S3d5: Summarize the subtype-stage determination results of all patients and generate a results list. The expression is: ; Preferably, S4 includes: S4a, all corresponding to the maximum number of candidate subtypes set for S2. The value retrieves the corresponding model result file after S3 training is completed; S4b, from each From the model results corresponding to the values, extract the cross-validation information criterion values ​​respectively. Log-likelihood value ; S4c, Number of candidate subtypes With the corresponding , Perform correlation to construct a set of indicators The expression is: ; Preferably, S5 includes: S5a, From the set of indicators In the process, the number of candidate subtypes with the smallest cross-validation information criterion value is selected, denoted as . ,Right now: ; If only one exists Then one The number of candidate optimal subtypes is marked; if multiple exist... corresponding Then select the one with the largest log-likelihood value. As the number of candidate optimal subtypes.

[0007] Compared with existing technologies, this invention provides a Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm, which has the following beneficial effects: This invention achieves accurate subtype identification and objective disease stage inference for Parkinson's disease patients by combining five modules: image data acquisition, healthy control, abnormal segmentation, SuStaIn model construction, and output. This is achieved through the synergistic collaboration of non-invasive corneal neural imaging and data-driven modeling. First, corneal neural images of the patient are acquired using a non-invasive corneal confocal microscope. After preprocessing and feature screening, a stable and effective core feature set is obtained. Then, through… Fraction standardization eliminates dimensional differences, based on An abnormality level is categorized using a score threshold, and a feature event matrix is ​​constructed to transform qualitative neural abnormalities into structured data that the model can parse. Subsequently, the feature event matrix and... Using a score matrix as input, the SuStaIn data-driven model is used to train, sample, validate, and calculate the posterior probability of multiple candidate subtypes. The optimal number of subtypes is determined by comprehensive model evaluation indicators. At the same time, the subtype classification and disease progression stage of each patient are output. Finally, through the system's output, the accurate identification of Parkinson's disease patient subtypes and the objective inference of disease stages are achieved, providing strong support for early diagnosis, personalized treatment, and drug development. Attached Figure Description

[0008] Figure 1 Effect size plots for each feature; Figure 2 For each feature Fractional box plot distribution; Figure 3 A diagram showing the selection of the optimal number of subtypes; Figure 4 This is a sequence diagram of the events' progression. Figure 5 This is a schematic diagram of the system of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It is worth noting that this application also relates to prior art. Since prior art is well known to those skilled in the art, it will not be described in detail in this application.

[0010] Please see Figure 1 A Parkinson's disease subtype staging inference system based on corneal neural images and SuStaIn algorithm includes an image data acquisition module, a healthy control module, an abnormality classification module, a SuStaIn model construction module, and an output module. The image data acquisition module includes an image data acquisition unit and a feature extraction unit. The image data acquisition unit is used to connect to the hospital system to acquire image data. Corneal confocal microscopy images of Parkinson's disease patients were collected to form a raw image dataset. Corneal confocal microscopy images of Parkinson's disease patients and healthy controls were also collected according to standard procedures, prioritizing the central corneal region and excluding images of peripheral or non-target areas. The validated ACCMetrics automated tool was used to objectively score the sharpness, contrast, signal-to-noise ratio, and nerve fiber edge sharpness of each image. A low-quality judgment threshold was set based on the objective scores, and images below the low-quality judgment threshold were automatically removed as invalid images. Images at or above the low-quality judgment threshold were sorted from highest to lowest score and selected. The final image dataset is constructed from the images with the highest scores. The feature extraction unit is used to call the proven and reliable automated tool ACCMetrics based on the final image dataset to perform batch analysis of images of the central corneal region of each eye in the final image dataset, automatically identify the morphology and distribution of corneal nerve fibers, complete quantitative calculations, and extract quantitative indicators of corneal nerve structures. Quantitative indicators of corneal neural structure include: nerve fiber density. Nerve branch density Nerve fiber length Density of the main nerve trunk Nerve fiber area Nerve fiber width Fractal dimension of nerve fibers Intergroup trend analysis was performed on the extracted quantitative indicators of corneal nerve structures to compare the changes in various characteristics between the Parkinson's disease patient group and the healthy control group, revealing the changes in nerve fiber width. In the Parkinson's disease patient group, the fiber width exhibited a trend opposite to the other six characteristics. To ensure consistency in the direction of disease events during subsequent SuStaIn modeling, fiber widths with abnormal trends were removed. Final screening of nerve fiber density Nerve branch density Nerve fiber length Density of the main nerve trunk Nerve fiber area Fractal dimension of nerve fibers A total of 6 effective features constitute the core effective features; The image data acquisition module acquires raw corneal confocal microscope image data of Parkinson's disease patients through the hospital system, then completes data preprocessing through a standardized process, and finally extracts quantitative indicators related to corneal neural structures through automated tools. Combined with inter-group trend analysis, abnormal trend indicators are eliminated, and the core effective feature set is finally determined to provide high-quality input features for subsequent modeling. The healthy control module is used to calculate the scores of each feature in the core effective features corresponding to Parkinson's disease patients using standardized methods. Fraction; The health control module is passed through The score calculation formula uses the mean and standard deviation of each core effective characteristic in healthy controls as a benchmark. It then standardizes the raw quantitative values ​​of the corresponding characteristics for each Parkinson's disease patient to obtain the core effective characteristics for each patient. Scores eliminate the differences in units between different features, making feature data comparable and laying the foundation for subsequent anomaly classification and model training. The exception partitioning module is used for... The calculated scores are used to classify anomaly levels and construct a feature event matrix; The anomaly segmentation module first uses each core effective feature Using the 95th percentile of the score as the baseline threshold, the patient's various characteristics were then analyzed. The score results are divided into three abnormal levels; then a binary feature event matrix is ​​constructed based on the division results. The abnormal level corresponding to each feature of the patient is identified by the value of the matrix element, and the qualitative abnormal level information is transformed into structured data that the model can recognize. The SuStaIn model building module is used to construct disease progression models based on feature event matrices and input patient features. The model training was completed by setting the score matrix and relevant parameters, and then the optimal number of subtypes was determined by multi-index evaluation. The SuStaIn model building module first combines the feature event matrix and the patient... The score matrix is ​​used as input data, and several core parameters are set. Then, independent training is carried out for each candidate subtype number. By randomly initializing the event sequence, optimizing and adjusting the event order, MCMC sampling and statistical calculation, and convergence judgment, the model training is completed and the subtype classification and disease stage judgment results of each patient are obtained. Finally, the cross-validation information criterion and log-likelihood value of the model under each candidate subtype number are extracted, and the optimal number of subtypes with the best model performance is selected by comprehensive comparison. The output module is used to output the model prediction results, including the subtype at the individual patient level, the disease progression stage and the corresponding posterior probability, as well as the number of the best subtypes at the group level, the distribution ratio of patients in each subtype and the progression order of the characteristic abnormalities corresponding to each subtype; the auxiliary validation information includes the model evaluation index of the number of candidate subtypes and the statistics of the location of characteristic abnormality level events, comprehensively presenting the basis for the model prediction effect and reliability.

[0011] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm, characterized in that, It includes an image data acquisition module, a healthy control module, an anomaly segmentation module, a SuStaIn model construction module, and an output module; The image data acquisition module includes an image data acquisition unit and a feature extraction unit. The image data acquisition unit is used to acquire the original image dataset and preprocess the original image dataset to form the final image dataset. The feature extraction unit is used to extract core effective features based on the final image dataset. The healthy control module is used to calculate each feature among the core effective characteristics corresponding to Parkinson's disease patients using statistical standardization methods. Fraction; The anomaly segmentation module is used to base on The calculated scores are used to classify anomaly levels and construct a feature event matrix; The SuStaIn model building module is used to construct a disease progression model based on a feature event matrix and input patient features. The model training was completed by setting the score matrix and relevant parameters, and then the optimal number of subtypes was determined by multi-index evaluation. The output module is used to output the model prediction results.

2. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 1, characterized in that, The original image dataset was obtained by connecting to the hospital system via an image data acquisition unit. Confocal microscopy image data of corneas from Parkinson's disease patients; The preprocessing includes: Corneal confocal microscopy images of Parkinson's disease patients and healthy controls were collected according to standard procedures, prioritizing the central corneal region and excluding images of peripheral or non-target areas. The validated ACCMetrics automated tool was used to objectively score the sharpness, contrast, signal-to-noise ratio, and nerve fiber edge sharpness of each image. A low-quality judgment threshold was set based on the objective scores, and images below the low-quality judgment threshold were automatically removed as invalid images. Images at or above the low-quality judgment threshold were sorted from highest to lowest score and selected. The final image dataset is constructed from the images with the highest scores.

3. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 2, characterized in that, The feature extraction unit includes: The proven and reliable automated tool ACCMetrics is used to perform batch analysis of images of the central corneal region of each eye in the final image dataset, automatically identify the morphology and distribution of corneal nerve fibers, perform quantitative calculations, and extract core effective features. The core effective feature includes: nerve fiber density. Nerve branch density Nerve fiber length Density of the main nerve trunk Nerve fiber area Fractal dimension of nerve fibers .

4. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 3, characterized in that, The Fraction The calculation formula is: ； in, Representing the Among the core effective characteristics of Parkinson's disease patients, the first is... The original quantitative value of the feature; The first of the core effective characteristics representing healthy controls The mean of the features; The first of the core effective characteristics representing healthy controls The standard deviation of the feature.

5. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 4, characterized in that, The classification of abnormal levels , The number of abnormal levels is set here. ,in Indicates no obvious abnormalities. Indicates mild abnormality. This indicates a severe anomaly, and simultaneously sets the first core effective feature. Item Features The 95th percentile of the score is ; when ,represent That is, there are no obvious abnormalities; when ,represent This indicates a mild abnormality. when ,represent This indicates a severe abnormality.

6. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 1, characterized in that, The feature event matrix for: ； in, Represents the number of Parkinson's disease patients; Represents the number of core effective features; This represents the number of exception levels. The first Parkinson's patient The first feature corresponds to the first element The rules for selecting values ​​are as follows: 。 7. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 6, characterized in that, The SuStaIn model building module includes: S1, The feature event matrix and patient characteristics The score matrix serves as the model input data, and the patient features... The fraction matrix is of Fractional matrix; S2. Set core operating parameters, including: number of patients, number of features, number of maximum candidate subtypes, number of MCMC sampling iterations, number of starting points, and number of cross-validation folds; S3. Based on the maximum number of candidate subtypes set in S2, the number of candidate subtypes is trained independently. S4. Extract the cross-validation criteria and log-likelihood value of the model under the number of candidate subtypes; S5. By comprehensively comparing the cross-validation information criterion and the log-likelihood value, the number of subtypes with the best model performance is selected as the final result.

8. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 7, characterized in that, S3 includes: S3a. Randomly initialize the order of the occurrence of the level 3 anomalies of the 6 core features in the core effective features, and generate the initial event sequence; S3b: The event order of the feature anomaly level is adjusted iteratively using an optimization algorithm to maximize the probability of the SuStaIn model interpreting the input data. S3c, Perform multiple MCMC samplings on the event locations for each anomaly level, and calculate the mean, variance, and standard deviation of the event locations through multiple iterative samplings, including: S3c.1 Construct a sampling result set from multiple MCMC samplings. The expression is: ； in, Representing the The event location value obtained from the second MCMC sampling. This represents the total number of MCMC samplings. Represents the core effective feature index, Represents the anomaly level index; S3c.2 Calculate the mean of the sampling result event locations. The formula is: ； in, Represents a dynamic index; S3c.3 Calculate the location variance of the sampling result events. The formula is: ； S3c.4 Calculate the standard deviation of the sampling outcome events. The formula is: ； S3c.

5. Based on the calculated mean, variance, and standard deviation of the event location, compare them with the preset thresholds for reasonableness of mean, convergence of variance, and convergence of standard deviation, respectively. When the calculated results of the three are all within the corresponding threshold range, the event location sampling results representing the abnormality level of the feature are converged and reasonable, and can be directly used for the posterior probability calculation of S3d; otherwise, it means that the sampling results are unreliable, and the MCMC sampling and the calculation steps of mean, variance, and standard deviation need to be repeated. S3d, Calculate the number of patients belonging to the largest candidate subtype. The posterior probability of each subtype and each disease stage within each subtype is used, and the maximum posterior probability is taken as the subtype classification and disease stage determination result for that patient, including: S3d1: Collect the valid statistics and basic data of the model that passed the S3c.5 test, and construct the input dataset. ; ； S3d2, for the first Patient No. 1, The first subtype, the Each disease stage, posterior probability The formula is: ； in, Represents the likelihood probability, indicating the patient's characteristic event state and Z-score data relative to the current situation. Subtype The degree of matching of the event sequence in each stage; Represents prior probability. , This represents the number of disease stages corresponding to a single subtype. S3d3, Traverse all patients All subtypes All stages of disease Substitute the values ​​into the S3d.2 formula to calculate the posterior probability and construct the probability set. : ； S3d4, for the first Patients, extract probability set The maximum posterior probability corresponding to the middle ,Right now: ； in, To determine the final subtype classification for this patient, This is the stage of disease progression in this patient; S3d5: Summarize the subtype-stage determination results of all patients and generate a results list. The expression is: 。 9. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 8, characterized in that, S4 includes: S4a, all corresponding to the maximum number of candidate subtypes set for S2. The value retrieves the corresponding model result file after S3 training is completed; S4b, from each From the model results corresponding to the values, extract the cross-validation information criterion values ​​respectively. Log-likelihood value ; S4c, Number of candidate subtypes With the corresponding , Perform correlation to construct a set of indicators The expression is: 。 10. The Parkinson's disease subtype staging inference system based on corneal neural images and the SuStaIn algorithm according to claim 8, characterized in that, S5 includes: S5a, From the set of indicators In the process, the number of candidate subtypes with the smallest cross-validation information criterion value is selected, denoted as . ,Right now: ； If only one exists Then one The number of candidate optimal subtypes is marked; if multiple exist... corresponding Then select the one with the largest log-likelihood value. As the number of candidate optimal subtypes.

Images