A multi-source parkinson's disease database integration method based on LEDD-UPDRS composite index alignment

Abstract

The application provides a multi-source Parkinson's disease database integration method based on LEDD-UPDRS composite index alignment, and is applied to the technical field of data processing. In view of the heterogeneous problems of multi-source Parkinson's disease databases, the application integrates heterogeneous data of PPMI, PDBP and local queues, disassembles characteristics according to four modes of heredity and clinic, introduces a standardized term library and a BERT-BiLSTM-CRF model unified term adjusted in combination with medical corpus, generates a cross-library feature table; a storage architecture is constructed based on FlashROM mode to realize data storage and retrieval; a composite observation system is built with LEDD and UPDRS-III, consistency is verified by multiple methods, data is standardized, missing data is filled in combination with multiple models to generate an integrated data set; a double prediction model is built relying on the data set, and the advantages of the data are realized by a Cohen's d value verification method; finally, cross-library joint analysis is carried out, phenotype-genotype correlation modeling and causal inference are completed, and data support is provided for Parkinson's disease research and clinical prediction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment. Background Technology

[0002] Different Parkinson's disease databases exhibit significant differences in their construction background, data collection standards, and terminology systems. PPMI, PDBP, and local cohort data differ in their descriptions of medical terms such as disease diagnosis, phenotypic characteristics, drug names, and examination items. Furthermore, the lack of unified standards across databases regarding the extraction dimensions, field definitions, and data types for modalities such as genetic, clinical, imaging, and sample characteristics hinders direct semantic integration and feature comparison across databases, creating "data silos" that severely restrict the fusion and analysis of multi-source data. Multi-source Parkinson's disease data includes structured clinical scale scores and genetic locus data, as well as unstructured image files and raw gene sequencing data. The storage characteristics and access requirements of different modalities vary significantly. Existing data storage methods often employ a single architecture, failing to adapt to the temporal correlations of clinical data, the network correlations of genetic data, and the large volume of unstructured data. This approach cannot guarantee the standardization and stability of various data storage methods, nor can it achieve accurate positioning and efficient cross-modal retrieval of multi-source heterogeneous data, increasing the operational costs of data use.

[0003] Differences in data collection equipment, batches, and population characteristics among different databases lead to significant biases in the cross-database distribution of core clinical biological indicators. Furthermore, the lack of a unified, multi-dimensional observation system for scientifically verifying and determining the consistency of cross-database data further exacerbates the problem. Existing data processing methods fail to accurately correct for batch effects in multi-source data or achieve standardization across different scales, resulting in low comparability of cross-database data. Directly using this data for analysis can introduce statistical bias, impacting the reliability of research results.

[0004] Parkinson's disease (Parkinson's disease) multi-source data commonly suffers from high-dimensional missing data due to factors such as follow-up dropouts and limitations in testing conditions. Furthermore, the missing mechanisms are complex. Traditional data imputation methods often employ single imputation techniques, failing to consider the biological characteristics and clinical logic of Parkinson's disease data. This makes it difficult to accurately characterize the missing data mechanisms, and the imputation results are prone to deviation from the true data distribution, leading to insufficient data completeness and reliability, and failing to meet the needs of subsequent in-depth analysis.

[0005] Existing methods for aligning multi-source medical data are mostly based on demographic variables and fail to incorporate the core clinical biological characteristics of Parkinson's disease. This results in inaccurate alignment at the medical level, making it difficult for the integrated data to effectively reflect the true biological patterns of the disease. Furthermore, the lack of quantitative validation indicators to scientifically evaluate the alignment effect makes it impossible to accurately determine the advantages of alignment methods in improving statistical power and reducing bias, thus compromising the quality of the integrated data.

[0006] Current technologies for integrating multi-source data on Parkinson's disease largely remain at the level of data stitching, failing to establish a systematic cross-database joint analysis framework tailored to the actual needs of clinical diagnosis and treatment, and also failing to fully explore the inherent correlations among multimodal data. This makes it impossible to achieve precise correlation modeling between phenotypic characteristics and genotype data, and even more difficult to conduct scientific causal inferences based on the integrated high-dimensional data. Consequently, the integrated data cannot be effectively transformed into practical support for disease mechanism exploration, personalized treatment, and clinical prediction, and its value is not fully realized.

[0007] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0008] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0009] According to one aspect of this application, a method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment is provided, comprising: receiving heterogeneous multi-source Parkinson's disease data including PPMI, PDBP, and local deep phenotypic cohorts; extracting core structured features according to genetic, clinical, imaging, and sample modality decomposition dimensions; introducing a standardized terminology database to construct semantic mapping; completing terminology unification by combining a BERT-BiLSTM-CRF model with pre-trained weights finely tuned from medical corpus; generating a standardized cross-database feature table; based on the multi-modal data storage characteristics adapted to FlashROM, constructing an entity-relationship model for clinical data, using graph database technology for genetic data, and establishing metadata indexes for unstructured data; completing structured storage and rapid retrieval of multi-source data according to modality characteristics; generating standardized multi-source Parkinson's disease database storage result information; and constructing a composite observation system with LEDD and UPDRS-III as the core. Cross-database consistency was verified through KS test, ANOVA, and coefficient of variation determination. ComBat batch effect correction and Z-score standardization were used to process the data, and a multimodal factor regression and dynamic weight allocation model was constructed. Missing data were filled using Bayesian joint modeling and variational autoencoder techniques, generating an integrated multimodal Parkinson's disease dataset. Common and specific fields were extracted from the dataset to construct an index. Based on the synthetic dataset, a Parkinson's disease drug efficacy and disease progression prediction model was built. The advantages of the Cohen's d-effects method were validated, achieving accurate alignment of clinical biological indicators from multiple sources, generating information on the integration and analysis of multi-source Parkinson's disease databases. Based on the integrated multimodal Parkinson's disease dataset, cross-database joint analysis was conducted to meet clinical diagnostic and treatment needs, achieving phenotypic-genotypic association modeling and causal inference for Parkinson's disease, providing data support for disease mechanism exploration and clinical prediction, and generating information on the application of integrated multi-source databases.

[0010] This application presents a method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment. This method integrates heterogeneous data from PPMI, PDBP, and local cohorts, decomposing features into four modalities: genetic, clinical, imaging, and sample. It combines a BERT-BiLSTM-CRF model fine-tuned from medical corpus with a standardized terminology database. A storage architecture is then constructed based on FlashROM modalities, with an entity-relationship model for clinical data, a graph database for genetic data, and metadata indexes for unstructured data. Subsequently, a composite observation system is built using LEDD and UPDRS-III, validated for consistency using multiple methods, and standardized. Missing data are filled in using multiple models to generate an integrated dataset. Next, a dual prediction model for drug efficacy and disease progression is built, validated using Cohen's d value to achieve accurate data alignment. Finally, cross-database joint analysis is conducted to complete phenotypic-genotypic association modeling and causal inference.

[0011] This research breaks down terminology and structural barriers in multi-source Parkinson's disease (Parkinson's disease) data, achieving cross-database semantic and feature unification and resolving the "data silo" problem. Furthermore, a modal storage architecture adapts to the characteristics of various data types, improving data retrieval and recall efficiency and ensuring data storage standardization. Through a composite observation system and multi-method data processing, batch effects are effectively eliminated, data scales are unified, missing values ​​are accurately filled, and the consistency and completeness of integrated data are improved. Based on precise alignment of clinical biological indicators, combined with quantitative validation and model optimization, the statistical power of Parkinson's disease prediction is enhanced, and analytical bias is reduced. Cross-database joint analysis achieves deep phenotypic-genotypic association and causal inference, providing high-quality data support for exploring the mechanisms of Parkinson's disease, personalized diagnosis and treatment, and clinical prediction, thus promoting the clinical translation of research findings.

[0012] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0013] Figure 1 This document illustrates a flowchart of a method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment, provided in an embodiment of this application. Figure 2 This illustration shows a schematic diagram of the structure of a multi-source Parkinson's disease database integration system based on LEDD-UPDRS composite index alignment, provided in an embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] The following is combined with Figure 1 This application describes a method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment, according to exemplary embodiments of this application. It should be noted that the application scenarios described below are merely illustrative for understanding the spirit and principles of this application, and the embodiments of this application are not limited in any way. Rather, the embodiments of this application are applicable to any suitable scenario.

[0016] In one implementation, Figure 1 A schematic flowchart of a multi-source Parkinson's disease database integration method based on LEDD-UPDRS composite index alignment according to an embodiment of this application is shown.

[0017] S101 receives heterogeneous Parkinson's disease data from multiple sources, including PPMI, PDBP, and local deep phenotypic cohorts. It extracts core structured features by decomposing them into genetic, clinical, imaging, and sample modal dimensions, introduces a standardized terminology database to construct semantic mapping, and completes terminology unification by combining a BERT-BiLSTM-CRF model with pre-trained weights finely tuned from medical corpus, generating a standardized cross-database feature table.

[0018] In one implementation, the received data sources are clearly defined and compliant, covering three core databases, as follows: Parkinson's Progression Biomarkers Initiative (PPMI): Obtained through official designated websites, containing multimodal longitudinal data including clinical assessments, neuroimaging, biosamples, and wearable devices, focusing on early diagnosis, disease progression monitoring, and biomarker discovery for Parkinson's disease. Parkinson's Disease Biomarkers Program (PDBP): Obtained compliantly from official open channels, with core data built around a biomarker-driven Parkinson's disease staging model, covering clinical and biosample data related to disease staging. Local / Institutional Deep Phenotypic Cohort Database: Integrates longitudinal treatment and periodic assessment data of Parkinson's disease patients accumulated within the institution, supplementing personalized treatment-related data not covered by the first two databases.

[0019] The received data types are categorized into four main modalities, all of which are core heterogeneous data for Parkinson's disease research: genetic data, such as single nucleotide polymorphism markers; clinical data, such as scale scores and medical history records; imaging data, such as CT, MRI, and PET images; and sample data, such as histopathological data and molecular marker data. A hierarchical feature extraction framework is adopted, decomposing the data into four modalities: genetic, clinical, imaging, and sample. Each modality is further extracted using a three-level standard of "core dimension - sub-dimension - feature item," with a screening threshold of "≥80% correlation with the diagnosis and treatment of neurological diseases." For the genetic modality, the core dimension includes static and dynamic genetic features. Static genetic features are extracted from sub-dimensions such as gene sequences and gene variations, while dynamic genetic features are extracted from sub-dimensions such as methylation, transcriptome, and gene expression levels. The extraction process employs a dual rule of "sequence alignment + variant annotation," performing sequence alignment based on the hg38 human genome reference sequence and using the ANNOVAR tool for variant annotation, ultimately extracting core features such as high-frequency pathogenic variants and drug sensitivity-related variants.

[0020] For the clinical modality, the core dimensions cover demographics, medical history and symptoms, laboratory tests, and treatment response. Demographics extracts basic demographic and social characteristics; medical history and symptoms extract disease history, symptoms, and signs; laboratory tests extract blood and urine tests; and treatment response extracts efficacy and safety indicators. Extraction follows the rule of "chronological order of diagnosis and treatment + clinical relevance," prioritizing diagnostic information conforming to the ICD-11 coding standards for neurological diseases, drug use information following clinical pharmacy guidelines, and prognostic information with a follow-up period of ≥6 months.

[0021] For imaging modalities, the core dimensions include structural images, functional images, and dynamic images. Structural images extract features from CT, MRI, and PET; functional images extract features from PET / CT and PET / MRI; and dynamic images extract features from ultrasound dynamic imaging and dynamic functional magnetic resonance imaging. During extraction, the selection criteria are "lesion identification + clinical diagnostic relevance," prioritizing the retention of image data with a signal-to-noise ratio ≥3 and lesion boundary clarity ≥85%. Quantitative and qualitative core features such as lesion size, density, and enhancement characteristics are extracted using a CNN convolutional neural network. For sample modalities, the core dimensions are histopathology and molecular markers. Histopathology extracts basic pathology and inflammation features; molecular markers extract DNA, RNA, and protein markers, all selected as core features closely related to the pathogenesis and treatment outcomes of Parkinson's disease. During the feature extraction process, Python software is used to scan the core data tables of each database to mine table structure information, clarify the field name, data type, value range, and logical relationship between fields for each feature item. At the same time, a dual mechanism of manual verification and machine verification is combined to ensure the accuracy of feature extraction results.

[0022] To address the inconsistencies in medical terminology across different databases, three internationally standardized terminology databases are introduced to construct a cross-database semantic mapping bridge: Experimental Factor Ontology (EFO), used to unify the terminology of experimentally related factors, such as the standardization of terms related to biomarker detection; Human Phenotype Ontology (HPO), focusing on the standardization of terminology related to Parkinson's disease phenotypic characteristics, such as the standardization of terms related to symptom presentation and sign description; and the Unified Medical Language System (UMLS), serving as the core terminology mapping standard, covering the standardized conversion of various types of terminology, including disease diagnosis, drug names, and clinical procedures.

[0023] During the semantic mapping construction process, the scattered terms in various databases (such as different expressions of "motor symptoms of Parkinson's disease" in different databases, different names of similar drugs, etc.) are associated with the corresponding terms in the above-mentioned standardized terminology database to eliminate terminological ambiguity and lay a semantic foundation for subsequent cross-database data fusion.

[0024] The model employs a joint architecture of "BERT encoding - BiLSTM feature enhancement - CRF decoding," with clearly defined module hierarchies and connections. Specifically, for the input layer: it receives raw medical terminology text data from various databases, including unstructured terminology information such as disease diagnosis names, phenotypic feature descriptions, drug names, and examination item names. For the BERT encoding layer: it uses BERT base weights finely tuned based on a large-scale medical corpus (covering specialized corpora such as neurological disease drug treatment and gene phenotypes) as pre-training weights. The model hierarchy is set according to the standard BERT architecture. Its core function is to convert the input terminology text into a vector representation containing semantic information, thereby enhancing the feature representation capability of medical terms.

[0025] For the BiLSTM layer: It is directly connected to the output of the BERT encoding layer. A bidirectional long short-term memory network captures the contextual dependencies of the terminology text, further enhancing the semantic relevance of the feature vectors. The layer structure is a standard two-layer structure, with the hidden layer dimensions optimized according to the length of the medical terminology text. For the CRF decoding layer: It is connected to the BiLSTM layer as the output layer. Its core function is to perform sequence labeling on the enhanced feature vectors, outputting standardized terminology results. The decoding process follows the sequence labeling rules for medical terms to ensure the accuracy of terminology conversion.

[0026] Medical terminology data was collected from clinical diagnosis and treatment databases of neurological diseases, domestic and international medical literature databases, and professional books to construct a training dataset containing multiple categories of terms, including disease diagnosis, phenotypic features, drug names, and examination items, ensuring that the data covers core terms related to Parkinson's disease research. Large-scale general BERT pre-trained weights were fine-tuned based on the aforementioned medical-specific corpus to adapt the model to the semantic understanding scenario of medical terms. The fine-tuning process used the gradient descent algorithm, with the learning rate set to 0.0001 according to the characteristics of the medical corpus, and 10 training epochs. The BERT encoding layer, BiLSTM layer, and CRF decoding layer were sequentially connected to form a complete model, which was then input into the training dataset for end-to-end training. The training objective was to minimize the terminology labeling error rate, using the Adam optimizer with a batch size of 32. After training, 30% of the automatically identified and matched terms were randomly sampled for verification. The sampled samples covered three categories: high-frequency terms, rare terms, and ambiguous terms. If labeling errors were found, the corrected terminology data was used as supplementary training data and re-inputted into the model for iterative optimization until the terminology labeling accuracy met the requirements for cross-database integration.

[0027] When the model encounters conflicting terms with unclear mappings between EFO / HPO / UMLS terms, a "hierarchical priority ranking + expert consensus adjudication" mechanism is adopted: the UMLS terminology mapping standard is followed first. If there is a clearly corresponding standardized term in UMLS, the terminology is unified directly according to that standard. If there is no clear mapping in UMLS, the cross-validation results of EFO and HPO terms are used as the standard, and the standardized expression that is consistently recognized in the two terminology databases is selected. If the above two levels still cannot resolve the conflict, it is submitted to an expert group composed of more than 3 clinical pharmacy experts and 2 neurology experts for consensus adjudication, and the standardized terminology expression is finally determined.

[0028] After standardizing terminology, the core structured features extracted from each modality were integrated to construct a standardized cross-database feature table. The feature table includes common and specific fields. The common fields are core features covered by all three major databases (such as patient basic information, core scale scores, and key biomarker test results), while the specific fields are unique features specific to each database (such as wearable device data from PPMI and personalized treatment plan data from the local database). All fields are clearly labeled. The data in the feature table has been semantically standardized, and field names, terminology, and data type definitions all follow a unified standard, ensuring direct comparability of data from different database sources and providing a structured and standardized data foundation for subsequent cross-database data integration and analysis.

[0029] S102, based on the multimodal data storage characteristics of FlashROM, constructs an entity-relationship model for clinical data, uses graph database technology for genetic data, and establishes metadata indexes for unstructured data. It completes the structured storage and rapid retrieval of multi-source data according to modal characteristics, and generates standardized multi-source Parkinson's disease database storage result information.

[0030] In one implementation, combining the storage characteristics of FlashROM (such as read / write speed, storage capacity adaptability, and data stability) with the heterogeneous characteristics of Parkinson's disease multimodal data (different modalities such as genetic, clinical, imaging, and sample data differ significantly in terms of data format, structural complexity, and intended use), with "modal adaptation, efficient storage, and fast retrieval" as the core objectives, a hierarchical and categorized modal structured storage architecture is built to specifically match the exclusive storage needs of clinical, genetic, and unstructured data, generating multi-source data storage architecture adaptation results.

[0031] Clinical data, centered on patients' longitudinal medical records, encompasses assessment indicators, treatment plans, and follow-up results at different time points. The core requirement is to support rapid association of multi-time-series data by patient ID. Therefore, the storage architecture emphasizes data correlation and temporal adaptation to ensure complete traceability of patient disease progression and treatment interventions. Genetic data focuses on the correlation of gene sequences, variant sites, and gene functions. The core requirement is to present the complex network relationships between genetic factors. Therefore, the architecture focuses on optimizing the storage of nodes (such as single nucleotide polymorphism sites, gene functions, and phenotypic characteristics) and biologically related edges to ensure efficient storage and retrieval of network relationships. Unstructured data primarily consists of large-volume image files (such as CT and MRI scans) and raw gene sequencing data. The core requirement is to balance storage capacity and retrieval efficiency. Therefore, the architecture emphasizes the separate storage design of metadata and raw data, using metadata indexes to quickly locate raw data and avoiding the efficiency loss caused by direct retrieval of large-volume data.

[0032] This modal structured storage architecture clearly defines the dedicated storage paths, unified storage format specifications, and standardized data interaction interfaces for clinical, genetic, and unstructured data. This ensures that all types of heterogeneous data can obtain storage solutions adapted to their characteristics, laying the foundation for rapid retrieval of multi-source data and cross-modal joint analysis. It also aligns with the core objective of "achieving efficient structured storage based on modal characteristics" in the construction of multi-source databases.

[0033] To address the temporal correlation and continuity of clinical data, an entity-relationship model is constructed using the patient ID as the core unique identifier. This model systematically links and stores various assessment indicators of the patient at different time points, generating structured storage results for clinical data. The core association logic of the entity-relationship model uses "patient" as the core entity, vertically linking four core sub-entities: "demographic information," "medical history and symptom records," "laboratory test results," and "treatment plan and response assessment." Each sub-entity corresponds to data at different time points in the patient's entire treatment process, forming a complete treatment trajectory chain: The "demographic information" sub-entity links to the patient's basic information at baseline, such as age, gender, and social characteristics, serving as a basic reference for subsequent diagnostic and treatment analysis. The "medical history and symptom records" sub-entity links to data such as the patient's medical history, symptom presentation, and changes in signs at the time of diagnosis and during follow-up, capturing the initial state and dynamic evolution of the disease. The "laboratory test results" sub-entity links to data such as blood tests, urine tests, and tissue and cell tests at various time points, reflecting the temporal changes in the patient's physiological and biochemical indicators. The "Treatment Plan and Response Assessment" sub-entity links data such as drug dosage, treatment plan adjustments, and UPDRS scale scores at different stages of patient treatment, reflecting the correspondence between treatment intervention and efficacy feedback.

[0034] This model clearly presents the complete disease progression and dynamic process of treatment intervention for patients from baseline assessment to subsequent treatment and follow-up, ensuring the relevance, temporality, and traceability of clinical data. It not only meets the core requirement of "structured storage of clinical data" in the construction of multi-source databases, but also lays the foundation for quickly retrieving complete diagnosis and treatment trajectory data by patient ID or time node and conducting longitudinal efficacy analysis.

[0035] This paper employs Neo4j graph database technology to construct a complex relational network for genetic data. Genetic loci, gene functions, and phenotypic characteristics are used as nodes, and biological associations are used as edges to generate a networked storage result for genetic data.

[0036] The node structure encompasses three core categories of genetic information: single nucleotide polymorphism (SNP) sites (such as identifiers of variant sites on specific genes), gene functions (such as the function of proteins encoded by genes and the biological pathways they participate in), and phenotypic characteristics (such as bradykinesia and tremor associated with Parkinson's disease). Edges are configured based on actual biological associations. For example, connecting "specific SNP sites" with "corresponding gene functions" through an "influence" relationship demonstrates the regulatory role of variant sites on gene function; connecting "gene functions" with "Parkinson's disease-related phenotypic characteristics" through an "association" relationship demonstrates the biological link between gene function and disease phenotype. This network structure visually presents the interactions between genetic factors, supports tracing associated genetic information from any node, and provides efficient support for subsequent correlation analysis of genetic and phenotypic data.

[0037] For unstructured data, including imaging and gene sequencing data, a metadata indexing system is established based on key information such as scan parameters and gene locations. Dedicated index tags are constructed to generate indexed storage results for unstructured data. For unstructured data such as image files (e.g., CT, MRI, PET) and raw gene sequencing data, a detailed metadata indexing system is constructed based on its core key information. Dedicated index tags are generated to achieve indexed storage of unstructured data and produce indexed storage results for unstructured data.

[0038] The core information extraction of the metadata indexing system follows the characteristics of data types: metadata for image data includes scan parameters (such as scan slice thickness, scan sequence, scan time), patient-related information (patient ID, examination time), and image lesion-related annotations (such as lesion location, size description), etc.; metadata for gene sequencing data includes sequencing platform information, sequencing depth, gene location (such as chromosome number, gene segment), sequencing sample number, etc. Based on this metadata, dedicated index tags are constructed. For example, the index tag for an MRI image might include key information such as "Patient ID: P001, Examination Time: 2023-01-15, Scan Type: Brain MRI, Scan Sequence: T1-weighted, Lesion Location: Basal Ganglia Region," etc. By associating index tags with the original unstructured data, data location can be achieved without directly reading large amounts of raw data, significantly improving the retrieval efficiency of unstructured data.

[0039] A unified, rapid retrieval mechanism is built based on the modal storage results to achieve accurate location and efficient retrieval of multi-source heterogeneous data, generating standardized multi-source Parkinson's disease database storage and retrieval results.

[0040] The core design of the retrieval mechanism is a multi-dimensional retrieval entry point and cross-modal data association retrieval function. The retrieval entry point covers key dimensions such as patient ID, data type, time range, and core indicators. For example, one can retrieve all modal data for a patient at a given time point using "patient ID + examination time," or retrieve associated genetic and clinical data using "genetic variant site + phenotypic characteristics." Cross-modal association retrieval achieves association through common fields of data from different modalities (such as patient ID and timestamp). For example, when retrieving data on a patient's "specific gene variant site," corresponding cross-modal data such as clinical symptom scores and imaging examination results can be retrieved simultaneously. The retrieval mechanism adapts to the interface of the multi-modal storage architecture, enabling rapid distribution of retrieval requests and efficient data return. This ensures accurate location and rapid retrieval of multi-source heterogeneous data under a unified retrieval entry point, meeting the data usage needs of cross-modal joint analysis.

[0041] S103 uses LEDD and UPDRS-III as the core to construct a composite observation system. Cross-database consistency is verified by KS test, ANOVA and coefficient of variation determination. ComBat batch effect correction and Z-score standardization are used to process the data. A multimodal factor regression and dynamic weight allocation model is constructed. Bayesian joint modeling and variational autoencoder technology are combined to fill in the missing data and generate an integrated multimodal Parkinson's disease dataset.

[0042] In one implementation, a composite observation system is constructed using LEDD (Levodopa Equivalent Daily Dose) and UPDRS-III as core indicators. KS test and ANOVA statistical tests are used to analyze the cross-database distribution characteristics and correlations of the indicators. LEDD cross-database consistency is determined by KS test P>0.05, ANOVA P>0.05, and coefficient of variation CV ≤15%. If consistency is insufficient, principal component / factor analysis is used to fuse the indicators, extracting principal component factors with a variance contribution rate ≥85% and assigning weights to generate validity verification results for the composite observation system. A cross-database aligned composite observation system is constructed using levodopa equivalent daily dose (LEDD) and the Unified Parkinson's Disease Rating Scale-III (UPDRS-III) as core indicators. Multi-dimensional testing and optimization ensure the system's effectiveness, generating validity verification results for the composite observation system.

[0043] Two types of statistical methods were used to conduct systematic analysis: first, the Kolmogorov-Smirnov test (KS test) was used to analyze whether the distribution characteristics of LEDD and UPDRS-III were consistent across PPMI, PDBP and local databases, and to determine the homogeneity of the indicators across databases; second, analysis of variance was used to test whether the differences between groups of the same indicator in different databases were significant; at the same time, Spearman correlation analysis was used to assess the association between LEDD and UPDRS-III and to clarify their synergistic role in the assessment of disease progression.

[0044] A triple criterion was set: KS test result P>0.05 (indicating no significant difference in LEDD distribution across different databases), ANOVA result P>0.05 (indicating no statistically significant difference in LEDD among different databases), and coefficient of variation (CV) of LEDD index across different databases ≤15%. If all three criteria are met, LEDD is considered to have good cross-database consistency and can be directly used as a common observation variable across databases.

[0045] If the cross-library consistency of LEDD does not meet the above standards, principal component analysis or factor analysis techniques will be used to integrate LEDD with key clinical indicators such as UPDRS-III to construct a more representative composite co-observation variable. Weight allocation will be determined based on the variance contribution ratio of the principal component analysis: principal component factors with a variance contribution rate ≥85% will be extracted, and the weights of each core indicator (such as LEDD and UPDRS-III) will be allocated according to their loading coefficients in the corresponding principal component factors, ensuring that the weight allocation aligns with the biological significance and clinical relevance of the indicators.

[0046] To address the characteristics of multi-source heterogeneous data, ComBat batch effect correction is employed. This involves optimizing accuracy through non-parametric Bayesian estimation to eliminate batch interference, and Z-score standardization to unify the data scale, thus completing the multimodal data standardization process and generating standardized multi-source Parkinson's disease baseline data. Specifically, to address the two core issues of scale differences and batch effects in multi-source Parkinson's disease data, a dual standardization strategy of "ComBat batch effect correction + Z-score standardization" is adopted to systematically eliminate data interference factors, completing the multimodal data standardization process and generating standardized multi-source Parkinson's disease baseline data.

[0047] The core objective of ComBat batch effect correction is to eliminate systematic biases caused by differences in data acquisition equipment, acquisition time, and population characteristics across different databases, while fully preserving the original biological signals inherent in the data to ensure its clinical relevance. The correction model explicitly incorporates gender, age, and disease duration as core covariates, employing nonparametric Bayesian estimation methods to optimize correction accuracy. By specifically adjusting batch-affected variables in genetic, imaging, and clinical data, the distribution trends of data from different sources, such as PPMI, PDBP, and local cohorts, are made more consistent, significantly improving the comparability of cross-database data and laying the foundation for subsequent multi-source data integration.

[0048] Z-score standardization is applicable to continuous variables in multimodal data that conform to an approximately normal distribution. Specifically, it includes gene expression levels in genetic data, quantitative parameters of lesions in imaging data (such as lesion size and density), and continuous physiological indicators in clinical data (such as UPDRS-III scores and UPDRS total scores). Through a standardization transformation formula, the above data from different scales and batches are mapped to the same quantitative dimension (mean of 0 and standard deviation of 1), completely eliminating the interference of data magnitude differences on subsequent integration analysis. This ensures that various continuous variables are correlated and fused under the same benchmark, improving the accuracy and reliability of multimodal data integration.

[0049] This study aims to uncover shared latent factors from multi-source data and achieve efficient dimensionality reduction. It combines Wald hypothesis testing with dynamic allocation of modal weights, determining the optimal number of factors based on scree plot test and a cumulative variance contribution rate ≥90%. A multimodal factor regression and dynamic weight allocation model is then constructed. Modal weights are calculated to generate the multi-source data fusion model construction results. By mining shared latent factors from multi-source data and dynamically allocating modal weights, a fusion model is constructed, generating the multi-source data fusion model construction results. Utilizing the low-rank characteristics of the data, shared latent factors from multi-modal data such as PPMI, PDBP, and local databases (genetic, clinical, imaging, and sample data) are deeply mined to achieve efficient dimensionality reduction of high-dimensional data, reduce redundant information interference, and focus on core correlation features. A dual standard is adopted, combining scree plot test and cumulative variance contribution rate: factors with a cumulative variance contribution rate ≥90% and corresponding to inflection points in the scree plot are extracted as the optimal number of factors, ensuring coverage of core data information while avoiding excessive model complexity due to an excessive number of factors.

[0050] Wald hypothesis testing was used to assess the significance of each data modality (genetic, clinical, imaging, and sample) to the latent factors, and weights were dynamically assigned based on the magnitude of their contribution. The weights were calculated using a composite quantification formula: , where ωi represents the weight of the i-th modality; λi is the normalized value of the factor loading coefficient of the i-th modality; γi is the clinical relevance score of the i-th modality; λj is the normalized value of the factor loading coefficient of the j-th modality; γj is the clinical relevance score of the j-th modality; α is the weight balance coefficient (taken as 0.6, taking into account both data characteristics and clinical significance), and the denominator is summed over all modalities to achieve weight normalization.

[0051] A Bayesian network model was used with the K2 algorithm for structure learning. The missing data mechanism was validated using the Little test. Multiple imputation and a joint model were used for iterative imputation of missing values. The expectation-maximization algorithm was used to jointly estimate factor regression coefficients and missing data. A variational autoencoder technique was introduced to build an encoder-decoder symmetric network. The latent space mapping was optimized using weighted nearest neighbor analysis based on Euclidean distance and feature importance. The model was trained to generate synthetic samples that conform to the real data distribution, achieving accurate imputation of high-dimensional missing data. Leave-one-out cross-validation and external validation using a Chinese Parkinson's disease cohort were employed. Standardized mean difference and sensitivity analysis were used to control imputation bias, generating complete imputation results. For non-random and modally correlated missing data in multi-source datasets, a multi-method fusion imputation strategy was adopted to generate complete imputation results.

[0052] The K2 algorithm was used for Bayesian network structure learning (adapting to the sparsity characteristics of high-dimensional medical data), with prior knowledge based on clinical guidelines and expert consensus for neurological diseases. The missing mechanism was validated using the Little test, and the goodness-of-fit (AIC) values ​​were compared under three hypotheses: completely random missing (MCAR), random missing (MAR), and non-random missing (MNAR) to determine the optimal missing mechanism hypothesis, providing a theoretical basis for accurate incompleteness.

[0053] By combining multiple imputation with a joint model, missing values ​​are filled in iteratively. The expectation-maximization algorithm is used to jointly estimate factor regression coefficients and missing data, ensuring that the imputation results are consistent with the latent factor structure. This avoids statistical inference bias caused by purely random imputation, thus improving data integrity and reliability.

[0054] A variational autoencoder technique is introduced to construct a symmetric encoder-decoder network structure: both the encoder and decoder have three fully connected layers. The input layer dimension adapts to the total dimension of the multimodal features, the intermediate hidden layer dimension is set to 128 dimensions (determined through grid search to balance feature compression efficiency and information retention), and the output layer dimension is consistent with the input layer. The latent space mapping is optimized using weighted nearest neighbor analysis, and the weights are calculated using a weighted method based on Euclidean distance and feature importance. ,in, Let σ be the Euclidean distance between sample i and sample j, and σ be the distance adjustment parameter. The feature importance weights for sample j are calculated using the random forest algorithm, addressing the alignment bias caused by scale and noise differences across different modalities. The trained and optimized model generates synthetic samples that conform to the distribution characteristics of the real data, effectively filling in missing data blocks and enhancing the representativeness of small sample modalities.

[0055] The reasonableness of the imputation results was double-checked using leave-one-out cross-validation (internal validation) and external validation with independent data from a Chinese Parkinson's disease cohort. Standardized mean difference and sensitivity analysis were used to strictly control imputation bias, ensuring that the imputed data were consistent with the true data in terms of distribution characteristics and biological associations, and that there was no systematic bias.

[0056] This process integrates data standardization, multi-source model fusion, and missing data completion to achieve deep integration of multi-source Parkinson's disease data, generating an integrated multimodal Parkinson's disease dataset. The integration process adheres to three core principles: structural uniformity, information integrity, and quality reliability. Structural uniformity relies on standardized terminology databases (EFO, HPO, UMLS) and data format specifications to ensure complete consistency in field definitions, terminology, and quantification dimensions for similar data from different database sources. Information integrity utilizes Bayesian joint modeling and variational autoencoder techniques to fill in non-random missing data and generate synthetic samples to enhance the representativeness of small sample modalities, ensuring no critical information is omitted. Quality reliability employs ComBat batch effect correction to eliminate systematic bias, Z-score standardization to unify data scales, and leave-one-out cross-validation and external cohort validation to control imputation bias, ensuring the data conforms to clinical biological principles.

[0057] The resulting integrated dataset can directly support downstream analysis scenarios such as drug efficacy prediction (e.g., LEDD response rate model) and disease progression prediction (e.g., UPDRS-III score change model), laying a high-quality data foundation for in-depth mining and clinical translation of multi-source data.

[0058] S104 extracts common and specific fields from the dataset to build an index, builds a Parkinson's disease drug efficacy and disease progression prediction model based on the synthetic dataset, and verifies the advantages of the effect size Cohen's d value method to achieve accurate alignment of clinical biological indicators of multi-source data, generating information on the integration and analysis of multi-source databases for Parkinson's disease.

[0059] In one implementation, common and specific fields are extracted from the integrated multimodal Parkinson's disease dataset to build a dedicated data indexing system, enabling rapid data location and feature retrieval, and generating multi-source data index construction results. Common fields are core features covered by PPMI, PDBP, and the local deep phenotypic cohort database, including basic patient identity information, core clinical assessment indicators (such as UPDRS scores), key genetic markers, and basic imaging examination types. These serve as the core link ensuring cross-database data association and guaranteeing the consistency of basic associations among data from different sources. Specific fields are unique features of each database, such as wearable device monitoring data from PPMI and personalized treatment plan records from the local database, preserving the unique value of each database while enriching the information dimensions of the integrated dataset.

[0060] The indexing system is built around the core objectives of "rapid retrieval" and "cross-modal association," and is designed in layers according to "data modality + core purpose." The data modality layer covers four major categories: genetic, clinical, imaging, and sample. The core purpose layer focuses on three major scenarios: disease diagnosis, efficacy evaluation, and progression monitoring, forming a multi-dimensional and cross-cutting index structure. Through this indexing system, target data can be quickly located based on multiple dimensions such as patient ID, data type, time range, and core indicators. It also supports cross-modal data association and retrieval. For example, by using the genetic variant site index of a patient, the corresponding clinical symptom scores, imaging examination results, and sample test data can be retrieved simultaneously, significantly improving the efficiency of using multi-source heterogeneous data and providing efficient data support for subsequent cross-database joint analysis and predictive modeling.

[0061] Based on synthetic datasets and Parkinson's disease risk prediction models, drug efficacy prediction models and disease progression prediction models were built respectively, forming a two-dimensional prediction model system for Parkinson's disease, and generating prediction model construction results. The core objective of the drug efficacy prediction model is to predict the patient's response to drug treatment, with the LEDD response rate as the key predictive indicator. The input data is integrated multimodal data, including patient genetic data (such as drug sensitivity-related variants), clinical data (such as baseline UPDRS scores and medical history information), and sample data (such as molecular marker detection results). During the model construction process, the clinical logic of drug treatment is combined, focusing on the influence of genetic factors on drug metabolism, the correlation between clinical phenotypes and drug efficacy, and learning the mapping relationship between multidimensional features and LEDD response rate through algorithms, outputting the predicted results of patient drug efficacy (such as effective, moderately effective, ineffective).

[0062] The core objective of the disease progression prediction model is to predict the disease progression trend in Parkinson's disease patients, using changes in the UPDRS-III score as the core predictive indicator. Input data includes longitudinal follow-up data (such as clinical assessment results at different time points), genetic data (such as high-frequency pathogenic variants), and imaging data (such as lesion change characteristics). The model focuses on key drivers of disease progression, learning the correlation between patient baseline characteristics and subsequent disease progression trajectories to output predicted changes in the UPDRS-III score over a future period (such as 6 months or 1 year), providing data support for disease progression monitoring.

[0063] This study uses the Cohen's d's effect size as a performance indicator. The performance of models based on clinical biological indicator alignment is compared with that of traditional demographic variable alignment methods. This quantitatively validates the advantages of the clinical biological indicator alignment method in improving statistical power and reducing bias, generating validation results for the method's effectiveness. The Cohen's d's effect size is used as the core performance indicator. By comparing the model performance of models based on clinical biological indicator alignment with those of traditional demographic variable alignment methods, the advantages of the clinical biological indicator alignment method are quantitatively validated, generating validation results for the method's effectiveness.

[0064] The validation logic revolves around two core dimensions: "statistical power" and "bias control." Regarding statistical power, the prediction accuracy of the two alignment methods for drug efficacy and disease progression is compared. The Cohen'sd value quantifies the difference in prediction results between the two methods; a higher Cohen'sd value indicates that the prediction effect of the clinical biological indicator-based alignment method is superior to the traditional method. Regarding bias control, the reasonableness of the data distribution after integration and the accuracy of missing data imputation are analyzed under both methods. Auxiliary indicators such as the standardized mean difference are used to verify the advantages of the clinical biological indicator-based alignment method in reducing data bias and improving data consistency. Through quantitative comparison, the effectiveness of the LEDD-UPDRS composite clinical biological indicator alignment method in cross-database data integration is clarified, providing data support for the method's wider application.

[0065] Based on the validation results, the model parameters and data alignment strategies were optimized to achieve accurate alignment of clinical biological indicators of multi-source Parkinson's disease data, generating accurate alignment results for multi-source data. Model parameter optimization focused on the core parameters of the prediction model, such as adjusting the weight balance coefficients of the multimodal factor regression model and optimizing the latent space dimension of the variational autoencoder, making the model more consistent with clinical biological laws. Data alignment strategy optimization mainly targeted aspects such as semantic mapping and scale unification of cross-database data, such as supplementing and improving the medical terminology conflict adjudication case library, optimizing the covariate settings of ComBat batch effect correction, and adjusting the applicability of Z-score standardization, to solve alignment bias problems found during validation.

[0066] The optimized data alignment results must meet three requirements: First, the terminology must be completely consistent and conform to the standardized terminology databases such as EFO, HPO, and UMLS; second, the data scale must be consistent, with similar indicators from different databases on the same quantitative dimension; and third, the biological correlations must be coherent, with the aligned data conforming to clinical biological pathways such as "drug dosage-microbiota metabolism-clinical outcome" to avoid data logical breaks.

[0067] This project integrates data indexing, predictive models, method validation, and data alignment across the entire workflow to achieve the consolidation and analysis of multi-source Parkinson's disease databases, generating integrated analysis results. The consolidation results are centered on structured datasets, including complete multimodal data records, a hierarchical indexing system, optimized predictive models, and method validation reports. The analysis results focus on core needs of Parkinson's disease research, such as cross-database correlation analysis conclusions, the correlation between drug efficacy and genetic factors, and key influencing factors of disease progression. The final output of the integrated analysis results ensures the integrity, consistency, and usability of the multi-source data, and provides a high-quality data foundation and methodological support for subsequent Parkinson's disease phenotype-genotype association modeling, causal inference, and clinical prediction.

[0068] S105, based on the integrated multimodal Parkinson's disease dataset, conducts cross-database joint analysis in conjunction with clinical diagnosis and treatment needs, realizes phenotype-genotype association modeling and causal inference for Parkinson's disease, provides data support for disease mechanism exploration and clinical prediction, and generates multi-source database integration application results information.

[0069] In one implementation, based on the integrated multimodal Parkinson's disease dataset (including genetic, clinical, imaging, and sample data across all dimensions), and combined with actual clinical needs (such as disease mechanism exploration, early diagnosis, drug efficacy evaluation, and disease progression monitoring), a general cross-database joint analysis framework is built using multi-source data collaborative mining methods. This framework is adapted to various analysis scenarios in clinical research by inputting multidimensional Parkinson's disease feature data.

[0070] The analytical framework is designed with a "data-driven + clinically oriented" core logic, encompassing four core modules: data input, feature selection, association analysis, and results output. The data input module supports batch import and format adaptation of multimodal data, compatible with integrated data formats from PPMI, PDBP, and local cohorts. The feature selection module automatically filters core features relevant to the analysis scenario based on a screening threshold of "≥80% correlation with the diagnosis and treatment of neurological diseases." The association analysis module incorporates various analytical algorithms, including phenotypic-genotypic association, disease progression trajectory analysis, and drug efficacy association. The results output module supports structured reports and visualized results presentation.

[0071] This framework is adaptable to diverse clinical research scenarios. For example, in the discovery of Parkinson's disease biomarkers, it can input genetic data, imaging data, and sample data to focus on the correlation analysis between gene variations and lesion characteristics and molecular markers; in the evaluation of drug efficacy, it can input clinical treatment data and genetic data to analyze the correlation between drug sensitivity-related variations and treatment response rates; and in the prediction of disease progression, it can input longitudinal follow-up clinical data and imaging data to uncover key drivers of disease progression, enabling flexible reuse of the analytical framework.

[0072] When conducting in-depth data mining during cross-library joint analysis, clinical biological indicator alignment results and multimodal data fusion features are introduced. In the in-depth mining process of cross-library joint analysis, clinical biological indicator alignment results (such as standardized data after LEDD-UPDRS composite indicator alignment) and multimodal data fusion features (such as comprehensive features after fusion of genetic-clinical-imaging features) are introduced to improve the clinical relevance and data reliability of the mining results.

[0073] The introduction of clinical biological indicator alignment results focuses on the dynamic association between LEDD and UPDRS-III (Spearman correlation coefficient |r|≥0.6), using core clinical information such as drug dosage adjustment and changes in disease symptoms as core clues to link relevant features in multimodal data. The introduction of multimodal data fusion features is based on the dynamic weight allocation model (ωi=(α×λi+(1-α)×γi) / ∑j(α×λj+(1-α)×γj), α=0.6) to generate fusion features, where ωi represents the weight of the i-th modality; α represents the weight balance coefficient; λi is the normalized value of the factor loading coefficient of the i-th modality; γi is the clinical correlation score of the i-th modality; λj is the normalized value of the factor loading coefficient of the j-th modality; and γj is the clinical correlation score of the j-th modality. This integrates the core information of different modalities and avoids the limitations of single-modality data.

[0074] During the data mining process, algorithms automatically connect the inherent logic of multi-dimensional data. For example, high-frequency pathogenic variants in genetic data are associated with symptom manifestations in clinical data and lesion characteristics in imaging data to explore the influence of genetic factors on clinical phenotypes and imaging features. Drug treatment data are associated with LEDD response rate and UPDRS-III score changes to analyze the differences in drug efficacy under different genetic backgrounds, providing comprehensive data support for in-depth analysis.

[0075] When analyzing clinical research data on Parkinson's disease, the analysis involves linking phenotypic features with genotype data and the dynamic changes in disease progression. Causal inference analysis incorporates high-dimensional data association chains obtained through cross-database integration. The core logic for analyzing Parkinson's disease clinical research data is "phenotype-genotype-disease progression," establishing a dynamic relationship among these three elements. Phenotypic features encompass multi-dimensional content such as demographic information, medical history and symptoms, laboratory test results, and treatment response. Genotype features focus on core genetic information such as single nucleotide polymorphism sites, gene function, and high-frequency pathogenic variants. The dynamic changes in disease progression are derived from longitudinal follow-up data (such as UPDRS-III scores at different time points, changes in imaging lesions, and drug dosage adjustment records).

[0076] During the analysis process, the algorithm learns the association patterns between phenotypic and genotypic features to identify which genetic variations are highly correlated with specific clinical phenotypes (such as bradykinesia and tremor). At the same time, it associates the dynamic relationship between the two and disease progression, analyzes the differences in disease progression rate and response to drug treatment among patients with different genotypes, and provides a basis for subsequent causal inference.

[0077] In causal inference analysis, a structured causal inference method is used to combine high-dimensional data association chains after cross-database integration (such as a complete association chain of "genetic variation - gene function change - molecular marker change - clinical symptom progression - drug treatment response"), eliminate confounding factors, and clarify the causal relationship between each factor.

[0078] The construction of association chains is based on the biological logic of multimodal data. For example, using LEDD as the core link, a clinical association chain is constructed for "drug dosage adjustment - LEDD change - UPDRS-III score fluctuation - disease progression stage"; starting with genetic variation, a biological association chain is constructed for "genetic variation - gene expression change - protein biomarker change - clinical phenotype presentation". Causal inference algorithms are used to quantify the strength of causal effects at each stage, clarify the causal relationship between core driving factors and downstream outcomes, and avoid false associations and misjudgments caused by pure data correlation.

[0079] After the analysis module outputs relevant research conclusions exploring disease mechanisms, it is optimized and improved through validation using actual clinical data. This generates a multi-source database integration application method and implementation details that combine phenotypic-genotypic association modeling and causal inference with clinical application guidance. The analysis module outputs relevant research conclusions exploring disease mechanisms (such as the association between specific genetic variations and the risk of Parkinson's disease, the therapeutic advantages of certain drugs for patients with specific genotypes, and key predictive indicators of disease progression), which are then double-validated and optimized using actual clinical data to ensure the reliability and clinical applicability of the conclusions.

[0080] The validation process was divided into internal and external validation: internal validation employed cross-database validation, using independent data subsets from PPMI, PDBP, and local cohorts to verify the consistency of the conclusions; external validation used independent external data, such as the Chinese Parkinson's Disease cohort, to examine the generalizability of the conclusions across different populations. For any biases identified during validation (such as conclusions not holding true in some populations or significant differences in effect strength), the process was retrospectively analyzed to adjust feature selection criteria or causal inference model parameters, optimizing and refining the research conclusions to ensure they align with clinical realities.

[0081] Based on the validated and optimized research conclusions, a multi-source database integration application method and implementation details combining phenotype-genotype association modeling and causal inference with clinical application orientation are generated to provide standardized operation guidelines for clinical research and diagnosis and treatment practice.

[0082] The application methodology clearly defines the core process, including multi-source data integration standards, specific steps in phenotype-genotype association modeling (such as feature selection, model construction, and parameter setting), key aspects of causal inference, and methods for controlling confounding factors. The implementation guidelines cover data quality control standards (such as missing value handling thresholds and outlier identification rules), specific requirements for model validation (such as the sample size ratio of the validation dataset and validation indicators), and clinical standards for interpreting results (such as determining the clinical significance of causal effect strength). This method and its guidelines can be directly applied to clinical research on Parkinson's disease (such as biomarker screening and patient stratification in drug clinical trials) and personalized treatment (such as genotype-based drug selection and disease progression risk assessment), achieving the clinical translation of multi-source database integration results.

[0083] In one implementation, such as Figure 2 As shown, this application also provides a multi-source Parkinson's disease database integration system based on LEDD-UPDRS composite index alignment, comprising: The multi-source heterogeneous data receiving and feature extraction module 201 is used to acquire heterogeneous Parkinson's disease data from PPMI, PDBP, and local deep phenotypic cohorts. It extracts core structured features according to the dimensions of genetic, clinical, imaging, and sample modality decomposition, introduces a standardized terminology library to construct semantic mapping, and combines the BERT-BiLSTM-CRF model with pre-trained weights finely tuned by medical corpus to unify terminology and generate a standardized cross-database feature table. The multimodal data structured storage module 202 is used to adapt to the multimodal data storage characteristics based on FlashROM, construct an entity-relationship model for clinical data, use graph database technology for genetic data, and establish metadata indexes for unstructured data. It completes the structured storage and rapid retrieval of multi-source data according to modal characteristics, and generates standardized multi-source Parkinson's disease database storage result information. The composite observation system construction and data integration module 203 is used to construct a composite observation system with LEDD and UPDRS-III as the core. The consistency across databases is verified by KS test, analysis of variance and coefficient of variation determination. ComBat batch effect correction and Z-score standardization are used to process the data. A multimodal factor regression and dynamic weight allocation model is constructed. The missing data is filled by Bayesian joint modeling and variational autoencoder technology to generate an integrated multimodal Parkinson's disease dataset. The data alignment verification and analysis module 204 is used to extract common and specific fields from the integrated dataset to build an index, build a Parkinson's disease drug efficacy and disease progression prediction model based on the synthetic dataset, and achieve accurate alignment of clinical biological indicators of multi-source data through the effect size Cohen'sd value verification method, and generate Parkinson's disease multi-source database integration and analysis results information. The cross-database joint analysis and application module 205 is used to conduct cross-database joint analysis based on the integrated multimodal Parkinson's disease dataset and in combination with clinical diagnosis and treatment needs. It realizes the modeling of Parkinson's disease phenotype-genotype association and causal inference, provides data support for disease mechanism exploration and clinical prediction, and generates multi-source database integration application results information.

[0084] The computer-readable storage medium provided in the above embodiments of this application and the multi-source Parkinson's disease database integration method based on LEDD-UPDRS composite index alignment provided in the embodiments of this application are based on the same inventive concept and have the same beneficial effects as the methods adopted, run or implemented by the applications stored therein.

[0085] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the integration method, system, electronic device, and readable storage medium for multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment are relatively simple to describe because they are substantially similar to the embodiments of the multi-source Parkinson's disease database integration method based on LEDD-UPDRS composite index alignment described above. Relevant parts can be referred to in the descriptions of the embodiments of the multi-source Parkinson's disease database integration method based on LEDD-UPDRS composite index alignment described above.

Claims

1. A method for integrating multi-source Parkinson's disease databases based on LEDD-UPDRS composite index alignment, characterized in that, include: It receives heterogeneous Parkinson's disease data from multiple sources, including PPMI, PDBP, and local deep phenotypic cohorts. It extracts core structured features by decomposing them into genetic, clinical, imaging, and sample modal dimensions, introduces a standardized terminology database to construct semantic mapping, and combines a BERT-BiLSTM-CRF model with pre-trained weights finely tuned by medical corpus to complete terminology unification and generate a standardized cross-database feature table. Based on the multimodal data storage characteristics of FlashROM, an entity-relationship model is constructed for clinical data, graph database technology is used for genetic data, and metadata indexes are established for unstructured data. Multi-source data is structured and quickly retrieved according to modal characteristics, generating standardized multi-source Parkinson's disease database storage result information. A composite observation system was constructed with LEDD and UPDRS-III as the core. Cross-database consistency was verified by KS test, analysis of variance, and coefficient of variation determination. ComBat batch effect correction and Z-score standardization were used to process the data. A multimodal factor regression and dynamic weight allocation model was constructed. The missing data were filled by Bayesian joint modeling and variational autoencoder technology to generate an integrated multimodal Parkinson's disease dataset. We extract common and specific fields from the dataset to build an index, build a prediction model for the efficacy of Parkinson's disease drugs and the progression of the disease based on the synthetic dataset, and verify the advantages of the Cohen's d effect size method to achieve accurate alignment of clinical biological indicators of multi-source data, and generate information on the integration and analysis of multi-source databases of Parkinson's disease. Based on the integrated multimodal Parkinson's disease dataset, cross-database joint analysis is carried out in combination with clinical diagnosis and treatment needs to realize the modeling of the association between Parkinson's disease phenotype and genotype and causal inference, provide data support for disease mechanism exploration and clinical prediction, and generate multi-source database integrated application results information.

2. The method as described in claim 1, characterized in that, Based on the multimodal data storage characteristics of FlashROM, an entity-relationship model is constructed for clinical data, graph database technology is used for genetic data, and metadata indexes are established for unstructured data. Structured storage and rapid retrieval of multi-source data are completed according to modal characteristics, generating standardized multi-source Parkinson's disease database storage results information, including: Combining the storage characteristics of FlashROM with the heterogeneous features of multimodal Parkinson's disease data, we adapt to the storage needs of clinical, genetic, and unstructured data, build a multimodal structured storage architecture, and generate multi-source data storage architecture adaptation results. Using patient ID as the core identifier, an entity-relationship model is constructed to associate evaluation indicators at different time points for clinical data, generating structured storage results for clinical data; Neo4j graph database technology is used for genetic data. Genetic loci, gene functions, and phenotypic characteristics are used as nodes, and biological associations are used as edges to construct a complex association network, generating a networked storage result for genetic data. For unstructured data including images and gene sequencing, a metadata indexing system is established based on key information including scanning parameters and gene locations, exclusive index tags are constructed, and unstructured data indexed storage results are generated; A unified and rapid retrieval mechanism is built based on the results of the modal storage, which enables accurate positioning and efficient retrieval of multi-source heterogeneous data, and generates standardized multi-source Parkinson's disease database storage and retrieval results information.

3. The method as described in claim 1, characterized in that, A composite observation system was constructed using LEDD and UPDRS-III as the core. Cross-database consistency was verified through statistical testing. ComBat batch effect correction and Z-score standardization were applied to process the data. A multimodal factor regression and dynamic weight allocation model was constructed. Bayesian joint modeling and variational autoencoder techniques were used to fill in missing data, generating an integrated multimodal Parkinson's disease dataset, including: A composite observation system was constructed using LEDD and UPDRS-III as core indicators. KS test and ANOVA statistical test methods were used to analyze the cross-database distribution characteristics and correlation of the indicators. The consistency of LEDD across databases was determined by KS test P>0.05, ANOVA P>0.05, and coefficient of variation CV≤15%. If the consistency was insufficient, the indicators were integrated through principal component / factor analysis, and the principal component factors with a variance contribution rate ≥85% were extracted and weighted to generate the validity verification results of the composite observation system. To address the characteristics of multi-source heterogeneous data, ComBat batch effect correction is employed, and batch interference is eliminated by optimizing accuracy through nonparametric Bayesian estimation and unifying the data scale through Z-score standardization. This completes the standardization processing of multimodal data and generates standardized multi-source Parkinson's disease baseline data. This study aims to uncover shared latent factors from multi-source data and achieve efficient dimensionality reduction. It combines Wald hypothesis testing with dynamic allocation of modal weights, determining the optimal number of factors based on scree plot test and a cumulative variance contribution rate ≥90%. A multimodal factor regression and dynamic weight allocation model is then constructed. Calculate the modal weights, where, Let i be the weight of the i-th mode. This is the normalized value of the modal factor loading coefficient. The clinical correlation score between this modality and the diagnosis and treatment of Parkinson's disease is calculated, with α being the weight balancing coefficient, and the results of the multi-source data fusion model construction are generated. The K2 algorithm was used to learn the structure of a Bayesian network model. The missing data mechanism was verified by the Little test. The missing values ​​were filled by iterative multiple imputation and joint model. The factor regression coefficients and missing data were estimated by the expectation-maximization algorithm. The variational autoencoder technology was introduced to build an encoder-decoder symmetric network. The latent space mapping was optimized by weighted nearest neighbor analysis based on Euclidean distance and feature importance. The model was trained to generate synthetic samples that conform to the distribution of real data, and high-dimensional missing data was accurately filled. The leave-one-out cross-validation method and external validation by a Chinese Parkinson's disease cohort were used. The imputation bias was controlled by standardized mean difference and sensitivity analysis to generate the missing data completion results. By integrating the results of standardized data processing, multi-source model fusion, and missing data completion, a deep integration of multi-source Parkinson's disease data is achieved, generating an integrated multimodal Parkinson's disease dataset.

4. The method as described in claim 3, characterized in that, A common and specific field was extracted from the dataset to construct an index. Based on the synthetic dataset, a predictive model for the efficacy of Parkinson's disease drugs and disease progression was built. The advantages of the Cohen's d' effect size method were validated to achieve accurate alignment of clinical biological indicators from multiple data sources. This resulted in the generation of integrated and analyzed results from a multi-source Parkinson's disease database, including: Common and specific fields are extracted from the integrated multimodal Parkinson's disease dataset to build a dedicated data indexing system, enabling rapid data location and feature retrieval, and generating multi-source data index construction results; Based on synthetic datasets and Parkinson's disease risk prediction models, drug efficacy prediction models and disease progression prediction models were built respectively, forming a two-dimensional prediction model system for Parkinson's disease and generating prediction model construction results. The Cohen's d value of the effect size was used to detect the efficacy index. The model performance based on the clinical biological indicator alignment method and the traditional demographic variable alignment method were compared. The advantages of the clinical biological indicator alignment method in improving statistical power and reducing bias were quantitatively verified, and the effectiveness of the method was verified. Based on the validation results, the model parameters and data alignment strategy were optimized to achieve accurate alignment of clinical biological indicators of multi-source Parkinson's disease data and generate accurate alignment results of multi-source data. By integrating the results of the entire process of data indexing, predictive modeling, method validation, and data alignment, the system completes the integration and analysis of multi-source Parkinson's disease databases and generates information on the integration and analysis results of multi-source Parkinson's disease databases.

5. The method as described in claim 1, characterized in that, Based on the integrated multimodal Parkinson's disease dataset, cross-database joint analysis was conducted in conjunction with clinical diagnosis and treatment needs to achieve phenotypic-genotypic association modeling and causal inference for Parkinson's disease. This provides data support for disease mechanism exploration and clinical prediction, and generates multi-source database integration application results, including: Based on the integrated multimodal Parkinson's disease dataset and combined with the actual needs of clinical diagnosis and treatment, cross-database joint analysis was carried out. An analysis framework was built through multi-source data collaborative mining methods, and multi-dimensional feature data of Parkinson's disease were input to adapt to various analysis scenarios in clinical research. When conducting in-depth data mining during cross-library joint analysis, the alignment results of clinical biological indicators and the fusion features of multimodal data are introduced; When analyzing clinical research data on Parkinson's disease, we correlate phenotypic characteristics with genotypic data and dynamic changes in disease progression. When conducting causal inference analysis, we combine high-dimensional data association chains after cross-database integration. After the analysis module outputs relevant research conclusions that explore disease mechanisms, it is optimized and improved through verification with actual clinical diagnosis and treatment data, generating a multi-source database integration application method and implementation details that combine phenotype-genotype association modeling and causal inference with clinical application orientation.

Images