Intelligent algorithm system for accurate diagnosis and treatment and management of nephropathy in whole cycle

Abstract

The invention relates to the technical field of medical information, in particular to an intelligent algorithm system for accurate diagnosis, treatment and management of the whole cycle of nephropathy. Through deep coupling of a core algorithm and in combination with two exclusive quantification formulas, transformation of nephropathy diagnosis and treatment from'experience-driven 'to'algorithm-driven' is realized, the core pain points of high early-stage missed diagnosis rate, progress pre-judgment lag, treatment scheme homogenization, inaccurate follow-up management and the like in traditional diagnosis and treatment are solved, accurate decision support is provided for clinicians, and the clinical diagnosis and treatment efficiency is improved. The method provides whole-cycle individualized diagnosis and treatment service for patients, and is suitable for diagnosis and treatment scenes of various nephropathy such as chronic kidney disease (CKD), acute kidney injury (AKI), nephrotic syndrome and the like. The core innovation of the method lies in breakthrough of algorithm modeling logic and solving process, a scientific decision-making system is constructed through an exclusive quantification formula, the industry blank of nephropathy full-cycle algorithm collaborative innovation is filled, and intelligent upgrading of diagnosis and treatment is promoted.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of medical artificial intelligence and kidney disease diagnosis and treatment, specifically involving an intelligent algorithm system covering the entire process of early screening for kidney disease, prediction of disease progression, generation of individualized treatment plans, and prognosis follow-up management. Background Technology

[0002] The complexity of nephrology diagnosis and treatment stems from the nonlinear correlations and dynamic changes in multimodal clinical data. The limitations of existing algorithms have become a core bottleneck restricting the improvement of diagnostic and treatment effectiveness, urgently requiring targeted breakthroughs. Traditional nephrology diagnosis and treatment rely on individual physician clinical experience, lacking standardized and systematic algorithmic modeling support. It has weak capabilities for deep coupling analysis of multidimensional data such as serum creatinine, urine protein, pathological sections, and genetic markers, making it difficult to capture hidden abnormal signals in early-stage nephropathy. This results in a persistently high rate of missed diagnoses of asymptomatic nephropathy, with the disease often progressing to middle or late stages by the time of diagnosis. Existing risk prediction models generally adopt a single-indicator linear analysis framework, focusing only on the static changes of core indicators such as eGFR. They lack multi-factor collaborative prediction logic and cannot characterize the complex nonlinear correlation between pathological damage, comorbidities, medication interventions, and disease progression, resulting in significant prediction lag and hindering early intervention. Treatment plan generation is limited to standardized recommendations from clinical guidelines, lacking adaptive optimization algorithms based on individual differences such as patient genetic background, physiological characteristics, and treatment preferences. Homogeneous plans cannot adapt to complex clinical scenarios, easily leading to adverse reactions or treatment ineffectiveness. The current follow-up management system employs a fixed-cycle model, lacking dynamic risk stratification and frequency optimization algorithms. Data is fragmented from the initial diagnosis and treatment stages, making it impossible to adjust interventions in real-time based on changes in the patient's condition. Furthermore, existing technologies lack dedicated quantitative formulas to support core decision-making, relying on general medical algorithm frameworks leading to poor adaptability. Moreover, a closed-loop algorithm system encompassing the entire process of "screening-prediction-treatment-follow-up" has not been established; each stage operates independently with disconnected data, hindering overall diagnostic and treatment effectiveness. In conclusion, the industry urgently needs a full-cycle intelligent algorithm system based on original algorithms and supported by dedicated quantitative formulas to overcome the shortcomings of existing technologies and drive the transformation of kidney disease diagnosis and treatment towards precision and intelligence. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent algorithm system for precise diagnosis and management of kidney disease throughout its entire life cycle. It adopts a three-layer architecture of "data layer - algorithm layer - application layer". The algorithm layer integrates four original core algorithms, namely, a multimodal data fusion kidney disease early screening algorithm, a kidney disease progression nonlinear prediction algorithm, an individualized treatment plan adaptive optimization algorithm, and a prognostic follow-up intelligent optimization algorithm. It also embeds two original quantitative formulas as core decision-making basis to realize closed-loop empowerment of the entire process of kidney disease early screening, disease progression prediction, individualized treatment plan generation, and prognostic follow-up management. The data layer constructs a multimodal data acquisition and preprocessing module, which integrates clinical laboratory data, imaging data, pathological data, gene data, patient baseline data and follow-up data, and generates a standardized data matrix after normalization, noise filtering, missing value filling and structure extraction. The algorithm layer achieves deep coupling of four core algorithms through data interaction interface, parameter linkage mechanism and result feedback link. The two original quantitative formulas are the quantitative formula for early abnormality of kidney disease and the quantitative formula for risk of kidney disease progression, which provide quantitative support for the MDS-ES algorithm and NPP-PP algorithm, respectively. The application layer includes a clinical decision-making terminal, a patient management platform, and a data visualization module, which provide targeted functions for doctors, patients, and administrators, respectively, to realize the clinical translation and practical application of algorithm results.

[0004] Furthermore, the multimodal data fusion algorithm for early screening of kidney disease adopts a four-layer architecture of "data hierarchical parsing - feature fusion modeling - anomaly quantification - screening result calibration". The formula for quantifying early abnormalities in kidney disease is: γ = ω·(C Wc + I ·Wi + G ·Wg ) + (1-ω)·R·A; Where γ is the early abnormality coefficient of nephropathy, with a value range of [0,1]; ω is the modal weighting coefficient, with a value range of [0.6,0.75]; C Wc is the standardized value of the i-th clinical laboratory feature. Let Wc be the weight vector of clinical test features and satisfy ΣWc =1;I Wi is the standardized value of the j-th image feature. The image feature weight vector satisfies ΣWi =1; G Wg is the standardized value of the k-th gene characteristic. The gene feature weight vector satisfies ΣWg =1; R is the cross-modal feature correlation coefficient, with a value range of [0.8, 1.2]; A is the population fit coefficient, with a value range of [0.9, 1.1].

[0005] Furthermore, the nonlinear prediction algorithm for kidney disease progression employs an improved Long Short-Term Memory (LSTM) network combined with a dual attention mechanism to construct the model. The formula for quantifying the risk of kidney disease progression is: P = μ·(P ·W + P_c·W_c + P_i·W_i) + (1-μ)·T·(1 - P_b·W_b); Where P is the probability of disease progression, ranging from [0,1]; μ is the core factor weighting coefficient, ranging from [0.65,0.8]; P W is the comprehensive value of core pathological factors. The core pathological factor weight vector satisfies ΣW =1; P_c is the comprehensive value of clinical dynamic factors, W_c is the weight vector of clinical dynamic factors and satisfies ΣW_c=1; P_i is the comprehensive value of intervention impact factors, W_i is the weight vector of intervention impact factors and satisfies ΣW_i=1; T is the time series correction coefficient, with a value range of [0.85, 1.15]; P_b is the comprehensive value of baseline risk factors, W_b is the weight vector of baseline risk factors and satisfies ΣW_b=1.

[0006] Furthermore, the individualized treatment plan adaptive optimization algorithm adopts a four-layer architecture of "individual feature analysis - plan generation modeling - efficacy prediction and risk assessment - dynamic optimization calibration", and introduces a comprehensive adaptation quantification formula for individualized treatment plans, which is expressed as: S = α·F + β·(E - λ·R) + (1-α-β)·D; Wherein, S is the comprehensive suitability score of the treatment plan, with a value range of [0,10]; α is the individual characteristic suitability weight coefficient, with a value range of [0.35,0.45]; F is the individual characteristic suitability coefficient; β is the efficacy risk balance weight coefficient, with a value range of [0.4,0.5]; E is the efficacy prediction score; λ is the risk weight coefficient, with a value range of [1.2,1.5]; R is the adverse reaction risk score; and D is the dynamic feedback calibration coefficient.

[0007] Furthermore, the intelligent optimization algorithm for prognosis follow-up adopts a four-layer architecture of "follow-up risk stratification - strategy generation and optimization - data collection and analysis - dynamic iterative adjustment", and introduces a comprehensive risk quantification formula for follow-up, which is expressed as: R_total =ω1·R_b + ω2·R_t + ω3·R_h + ε·K; Wherein, R_total is the follow-up comprehensive risk score, with a value range of [0,10]; ω1, ω2, and ω3 are the weight coefficients of the basic disease risk, treatment response risk, and home management risk, respectively, and satisfy ω1+ω2+ω3=1; R_b is the basic disease risk score, R_t is the treatment response risk score, and R_h is the home management risk score; ε is the dynamic correction coefficient, with a value range of [0.85,1.15]; and K is the strategy adaptation calibration coefficient, with a value range of [0.9,1.1].

[0008] Furthermore, the feature fusion modeling of the MDS-ES algorithm adopts a three-order strategy of "local feature enhancement - cross-modal fusion - feature dimensionality reduction optimization". Local feature enhancement determines the importance weight of each feature through the analytic hierarchy process, with core feature weights assigned to 0.6-0.8 and secondary feature weights assigned to 0.2-0.4. Cross-modal fusion constructs an attention mechanism fusion model and calculates the correlation coefficient between features of different modalities through a self-attention network. Feature dimensionality reduction optimization adopts a combination of principal component analysis (PCA) and sparse matrix factorization.

[0009] Furthermore, the dual attention mechanism of the NPP-PP algorithm includes a temporal attention mechanism and a factor association attention layer. The temporal attention mechanism assigns high attention weights to key time nodes of clinical dynamic factors. The factor association attention layer calculates the correlation coefficients between different categories of factors and generates a multi-factor nonlinear correlation matrix. The algorithm adopts a multi-task learning framework to simultaneously achieve three core tasks: prediction of progression risk probability, prediction of progression time nodes, and identification of key driving factors.

[0010] Furthermore, the ATO-TS algorithm's protocol generation modeling employs an improved collaborative filtering algorithm combined with clinical guideline constraints. A basic protocol library is constructed based on clinical guidelines, and effective protocol parameters are extracted by matching similar historical cases through the collaborative filtering algorithm. Protocol details are adjusted in conjunction with individual characteristics to generate 2-3 candidate protocols. The multi-task learning framework simultaneously outputs efficacy prediction results and adverse reaction risk assessment results, and uses the analytic hierarchy process (AHP) to determine efficacy and risk weights to select the optimal protocol.

[0011] Furthermore, the follow-up risk stratification of the PFO-FU algorithm is based on a three-factor system of "basic disease risk - treatment response risk - home management risk". The weights are assigned using the analytic hierarchy process, with basic disease risk weighted at 0.5, treatment response risk weighted at 0.3, and home management risk weighted at 0.2. The comprehensive follow-up risk level is calculated and divided into three levels: high risk, medium risk, and low risk. Personalized follow-up strategies are then customized based on the level.

[0012] Furthermore, the intelligent algorithm system's method for precise management of kidney disease throughout its entire lifecycle includes the following steps: Step 1: Collect multimodal data through multiple channels in the data layer, generate a standardized data matrix after preprocessing, and input it into the algorithm layer; Step 2: The MDS-ES algorithm performs hierarchical analysis, feature fusion, anomaly measurement, and result calibration on the data, and outputs early screening results; Step 3: The NPP-PP algorithm, based on the screening results, integrates multi-dimensional progression factors and outputs disease progression risk assessment results through nonlinear correlation modeling and quantitative formula calculation. Step 4: Based on the screening and prediction results, the ATO-TS algorithm integrates individual characteristics to generate candidate treatment plans, selects the optimal plan through efficacy and risk assessment, and pushes it to the application layer; Step 5: The PFO-FU algorithm, based on treatment plans and efficacy feedback, stratifies and customizes follow-up strategies, collects follow-up data, and dynamically iterates and optimizes to form a closed-loop management system throughout the entire cycle.

[0013] Compared with the prior art, the beneficial effects of the present invention are: By deeply coupling four original core algorithms and combining them with two exclusive quantitative formulas, the system transforms kidney disease diagnosis and treatment from "experience-driven" to "algorithm-driven," addressing core pain points in traditional diagnosis and treatment such as high early missed diagnosis rates, delayed progression prediction, homogenized treatment plans, and inaccurate follow-up management. It provides precise decision support for clinicians and offers patients personalized diagnosis and treatment services throughout the entire lifecycle. It is applicable to various kidney disease diagnosis and treatment scenarios, including chronic kidney disease (CKD), acute kidney injury (AKI), and nephrotic syndrome, promoting the intelligent upgrade of diagnosis and treatment. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the working principle of the system described in this invention; Figure 2 MDS-ES is a flowchart of the algorithm parsing process; Figure 3 The flowchart for parsing the ATO-TS algorithm. Detailed Implementation

[0015] 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.

[0016] 1.1 System Overall Architecture The intelligent algorithm system for precise diagnosis and management of kidney disease throughout the entire cycle provided by this invention adopts a three-layer architecture of "data layer - algorithm layer - application layer" to achieve multi-dimensional data integration and full-process algorithm empowerment. The core of the architecture design focuses on algorithm collaboration and clinical implementation adaptation to ensure data connectivity and functional closure at each stage. The data layer constructs a multimodal data acquisition and preprocessing module, integrating clinical laboratory data (serum creatinine, urine protein, cystatin C, electrolytes, uric acid, etc.), imaging data (morphological and density characteristics of renal ultrasound, CT, and MRI), pathological data (glomerular sclerosis degree, renal tubular injury grade, and interstitial fibrosis ratio in renal biopsy sections, etc.), gene data (expression level, mutation type, and regulatory relationship of differentially expressed genes), patient baseline data (age, gender, comorbidities, medication history, family history, body mass index, etc.), and follow-up data (home monitoring indicators, symptom feedback, and intervention implementation status). Through normalization processing to eliminate dimensional differences, noise filtering algorithms to remove abnormal fluctuations, missing value filling algorithms to fill data gaps, and structured extraction processing of non-text data, a standardized data matrix is ​​generated, providing high-quality input for the algorithm layer. The algorithm layer, the core of the system, integrates four original core algorithms to enable the entire process of early screening, progression prediction, treatment plan generation, and follow-up optimization. Deep coupling between algorithms is achieved through data interaction interfaces, parameter linkage mechanisms, and result feedback links. Two original quantitative formulas are embedded as core decision-making bases to ensure the accuracy and scientific rigor of the algorithm output. The application layer provides a clinical decision-making terminal, a patient management platform, and a data visualization module, offering targeted functions for doctors, patients, and administrators, respectively. The clinical decision-making terminal supports viewing algorithm results, fine-tuning treatment plans, and data entry; the patient management platform supports uploading data from home, medication reminders, and follow-up notifications; and the data visualization module supports full-process data monitoring and algorithm performance evaluation, enabling the clinical translation and practical application of algorithm results. The entire system constructs a closed-loop logic of "data acquisition - algorithm processing - decision output - feedback optimization," ensuring that algorithm performance dynamically adapts to clinical needs through continuous iteration, forming an intelligent empowerment system covering the entire cycle of kidney disease diagnosis and treatment.

[0017] 1.2 Core Invention Points and Algorithm Details 1.2.1 Invention Point 1: Multimodal Data Fusion Algorithm for Early Screening of Kidney Disease 1.2.1.1 Algorithm Modeling Logic The core modeling goal of the MDS-ES algorithm is to construct a multimodal data collaborative screening model, addressing the core problems of traditional early screening, such as reliance on single indicators, high false negative rates, poor specificity, and lack of quantitative evidence. This aims to achieve accurate capture and grading of early abnormal signals in kidney disease, thus promoting earlier diagnosis and treatment. The modeling adopts a four-layer architecture: "data hierarchical analysis - feature fusion modeling - anomaly quantification - screening result calibration." It integrates three core data categories: clinical laboratory data, imaging data, and genetic data. Through a full-link design of differentiated feature extraction, deep fusion modeling, quantitative formula support, and dual-verification optimization, it achieves multi-dimensional coverage of early kidney disease manifestations from molecular, structural, and functional perspectives, significantly improving the detection rate of early kidney disease (especially in the asymptomatic stage) while reducing the false positive rate, providing clear and quantifiable screening decision-making basis for clinicians. The algorithm's modeling logic breaks through the limitations of traditional data splicing and fusion, emphasizing the collaborative correlation and individualized adaptation of data from various modalities. Its core innovation lies in constructing an integrated logic of "feature hierarchical - fusion enhancement - quantitative calibration," ensuring the sensitivity, specificity, and clinical suitability of the screening results.

[0018] Data hierarchical analysis and modeling focuses on extracting differentiated features from multimodal data. Dedicated analysis logic is constructed for different types of data based on their structural characteristics and clinical value, ensuring that core information from each modality is not lost. The clinical laboratory data layer employs a three-tiered analysis strategy: "core indicator extraction - derived indicator calculation - dynamic feature quantification." Core biochemical indicators include serum creatinine, urine albumin / creatinine ratio (ACR), cystatin C, electrolytes, and uric acid—key indicators directly related to the early onset of kidney disease—precisely capturing the baseline state of renal function. Derived indicators are calculated based on core indicators using standardized formulas, including estimated glomerular filtration rate (eGFR), kidney injury index, and dynamic change rate of urine protein, enhancing the clinical relevance of the indicators. Dynamic change features are quantified by comparing multiple test data within the past three months using a linear regression model, capturing potential latent renal function abnormalities and avoiding the limitations of single-point-of-time data. The image data layer employs an improved convolutional neural network (CNN) to construct a feature extraction model. The network structure is optimized for the morphological characteristics of kidney images, with the addition of edge detection modules and texture feature extraction units. This accurately extracts core features such as kidney volume, cortical thickness, cortical echo intensity, aggregate system morphology, space-occupying lesions, and renal parenchymal blood flow signals. Simultaneously, through grayscale quantification and standardization, image features are transformed into computable numerical vectors, quantifying the degree of abnormality and addressing the issues of reliance on manual interpretation and high subjectivity in traditional image analysis. The gene data layer targets differentially expressed genes related to kidney disease (such as DUSP1, GADD45B, IFI44L, and CXCL10). It utilizes gene expression profiling and feature selection algorithms to extract features such as gene expression levels, mutation types, intergenetic regulatory relationships, and pathway enrichment. Through correlation analysis and importance ranking, a subset of core genes highly correlated with early-stage kidney disease is selected, redundant gene data is eliminated, reducing algorithm computational complexity while ensuring accurate capture of abnormal signals at the gene level.

[0019] Feature fusion modeling employs a three-tiered strategy of "local feature enhancement - cross-modal fusion - feature dimensionality reduction optimization," overcoming the limitations of traditional single-dimensional fusion and achieving deep coupling and collaborative characterization of multimodal features. Local feature enhancement uses the analytic hierarchy process (AHP) to determine the importance weight of each feature, assigning high weights (0.6-0.8) to core features (such as ACR, eGFR, renal cortical thickness, and key gene expression levels) and low weights (0.2-0.4) to secondary features (such as uric acid, echo homogeneity, and non-core gene expression levels). This ensures that core information plays a dominant role in screening results, while a dynamic weight adjustment mechanism adapts to the needs of different populations and screening scenarios. A cross-modal fusion model employs an attention mechanism to calculate correlation coefficients between features from different modalities. This model accurately captures the nonlinear relationships between clinical indicators and imaging features, as well as between genetic features and pathological manifestations. Examples include the synergistic relationship between abnormal gene expression and renal cortical thinning, and the correlation between dynamically elevated urinary protein and abnormal echo intensity. This generates a cross-modal fusion feature matrix, achieving complementarity and enhancement of multi-dimensional information and avoiding misjudgments caused by isolated analysis of individual modal data. Feature dimensionality reduction optimization utilizes a combination of principal component analysis (PCA) and sparse matrix factorization to eliminate redundant information and noise interference in the fused features, retaining core discriminative features. The high-dimensional fusion feature vector is transformed into a low-dimensional core feature vector, reducing computational complexity, improving screening efficiency, and ensuring that the discriminative value of the features is not lost, providing accurate input for subsequent anomaly quantification.

[0020] The abnormality quantification modeling is based on an original formula for quantifying early-stage renal abnormalities. It transforms the fused feature matrix into a calculable abnormality coefficient, enabling quantitative screening and grading of early-stage renal disease, overcoming the challenges of traditional screening methods lacking quantitative basis and having vague judgment standards. The formula design follows the logic of "core feature-driven + correlation correction auxiliary + population adaptation calibration," integrating quantitative information from abnormal features across various modalities. Through nonlinear mapping, feature values ​​are transformed into standardized abnormality scores. Simultaneously, a population adaptation factor is introduced to adjust the quantification standards for different age groups, genders, and underlying diseases, avoiding screening biases caused by universal standards and ensuring the formula's applicability across different populations. Based on clinical experience and expert consensus, an abnormality grading threshold is constructed, dividing screening results into three levels: "normal," "suspected," and "abnormal," with corresponding clinical management recommendations for each level: the normal level indicates no obvious early-stage renal disease signals, recommending routine physical examinations and follow-up; the suspected level indicates a potential risk of early-stage renal disease, recommending further specialized examinations (such as 24-hour urine protein quantification, repeated renal biopsy, and gene depth testing); and the abnormal level indicates a high probability of early-stage renal disease, recommending immediate initiation of the clinical diagnostic process, providing doctors with clear and actionable decision-making basis.

[0021] The screening result calibration model constructs a dual-validation mechanism of "internal data validation + external clinical validation" to ensure the accuracy and reliability of screening results, forming a closed loop for algorithm iterative optimization. Internal validation uses a large-scale historical case dataset for 10-fold cross-validation of screening results. Based on confusion matrix analysis, it assesses screening accuracy, sensitivity, and specificity, optimizing formula parameters and grading thresholds for false positives and false negatives, adjusting feature weight allocation, and improving the algorithm's generalization ability. External validation incorporates feedback from experienced clinicians, assembling a review team composed of senior nephrologists to manually review suspected and borderline abnormal cases. Combining clinical diagnoses and pathology reports, the review opinions are transformed into algorithm adjustment factors, inversely optimizing feature extraction rules, weight allocation, and quantification formula parameters. Simultaneously, a dynamic calibration interface is built to support continuous optimization of model parameters and judgment criteria based on new clinical data, research findings, and treatment guidelines, ensuring the adaptability and accuracy of the screening algorithm and achieving continuous improvement in algorithm performance.

[0022] 1.2.1.2 Algorithm Solution Process The MDS-ES algorithm employs a comprehensive solution strategy encompassing "data acquisition - hierarchical analysis - feature fusion - anomaly measurement - result calibration." Each step is closely linked and interconnected, forming a standardized and replicable screening process to ensure the stability of the algorithm and the consistency of the results. The specific process is as follows: The first step is multimodal data acquisition: Through the hospital information system (HIS), laboratory information system (LIS), image archiving and communication system (PACS), gene testing platform and patient self-reporting data channels, we comprehensively collect four major categories of data: clinical laboratory data, kidney imaging data, gene data and patient baseline data, and build a multi-source data input system. Clinical laboratory data collection includes serum and urine test results from the past three months, containing at least three dynamic test data points to ensure the capture of dynamic trends in indicators. Core indicators must be fully covered, including creatinine, ACR, cystatin C, electrolytes, and uric acid. Imaging data prioritizes renal ultrasound images (due to their convenience and applicability), supplementing with CT and MRI images for suspected cases. Imaging reports and physician interpretations are obtained simultaneously to ensure the completeness of the imaging data. Genetic data collection includes results from differentially expressed genes, including gene expression levels and mutation types. For primary healthcare institutions lacking gene testing capabilities, the gene data module can be flexibly disabled, with the algorithm automatically adjusting feature weights to accommodate missing data. Baseline data includes age, gender, comorbidities (hypertension, diabetes, obesity, etc.), medication history, family history, body mass index, and lifestyle habits, comprehensively covering the fundamental factors influencing kidney disease incidence. A data verification mechanism is implemented during data collection. Invalid data (such as abnormal testing times, indicator values ​​exceeding reasonable ranges, or incorrect data formats) is removed through rule-based verification. Repeated data collection and comparison ensure data accuracy, laying the foundation for subsequent algorithm processing.

[0023] The second step is data layered analysis and feature extraction: For the collected multimodal data, analysis and feature extraction are carried out separately for three layers: clinical laboratory, imaging, and genetics, generating standardized feature vectors. In the clinical laboratory data layer, derived indicators such as eGFR and renal injury index are calculated using standardized formulas. A linear regression model is used to quantify the dynamic trends and fluctuations of these indicators. Combining core indicators, derived indicators, and dynamic features, a 20-dimensional clinical feature vector is generated. In the imaging data layer, an improved CNN model is used to segment, detect edges, and extract features from images, quantifying 15-dimensional imaging features such as kidney morphology, echo intensity, and blood flow signals. Min-max normalization is used to map feature values ​​to the [0,1] interval to eliminate dimensional differences. In the genetic data layer, 10 core differentially expressed genes in the early stages of kidney disease are selected. Gene expression levels and mutation type features are extracted to generate a 10-dimensional gene feature vector. Gene expression levels are standardized, with mutated genes marked as high feature values ​​and non-mutated genes marked as low feature values. Consistency checks are performed on the feature vectors of each layer, and abnormal feature values ​​are removed to ensure the reliability of the feature vectors and provide high-quality input for the feature fusion process.

[0024] The third step is cross-modal feature fusion: an attention mechanism fusion model is used to achieve deep coupling of clinical, imaging, and genetic features, constructing a collaborative feature system. First, a local feature enhancement matrix is ​​constructed, and the feature vectors of each layer are weighted based on feature importance scores. The weight of core features is set to 0.6-0.8, and the weight of secondary features is set to 0.2-0.4, enhancing the discriminative value of core features. Second, a self-attention network is used to calculate the correlation coefficients between different modal features, capturing the nonlinear correlation between clinical indicators and imaging and genetic features, generating a cross-modal correlation matrix to quantify the degree of collaborative anomalies among features of each modality. Finally, the weighted feature vectors and the correlation matrix are multiplied to generate a 30-dimensional core fusion feature vector, achieving a collaborative characterization of multi-dimensional information. This retains the core features of each modality while highlighting the correlation value between features, providing accurate input for anomaly measurement.

[0025] The fourth step is abnormality quantification and graded screening: The core fusion feature vector is input into the original early kidney disease abnormality quantification formula to calculate the abnormality coefficient γ (value range [0,1]). Based on preset thresholds, the screening results are divided into three levels: γ < 0.3 is normal, indicating no obvious early kidney disease signals, and it is recommended to follow up at the routine physical examination frequency, with core indicator testing at least once a year; 0.3 ≤ γ < 0.6 is suspected, indicating a potential risk of early kidney disease, and it is recommended to further improve specific examinations, including 24-hour urine protein quantification, repeated kidney biopsy, and gene depth testing, to clarify the diagnosis; γ ≥ 0.6 is abnormal, indicating a high probability of early kidney disease, and it is recommended to immediately initiate the clinical diagnosis and treatment process, and carry out comprehensive assessment and targeted intervention. A standardized screening report is generated, which clarifies each abnormality characteristic, corresponding risk level, quantification coefficient, and clinical treatment recommendations, and is pushed to the clinical decision terminal to provide doctors with clear decision support.

[0026] The fifth step, result calibration and iterative optimization, employs a dual-validation mechanism to calibrate the screening results, ensuring screening accuracy. Internally, 10-fold cross-validation is used to evaluate the screening accuracy, sensitivity, and specificity based on historical case datasets. Formula parameters (such as modality weight coefficient ω and feature weight vector W) and grading thresholds are adjusted for false positives and false negatives, optimizing feature extraction rules. Externally, at least two senior nephrologists are invited to manually review suspected and abnormal cases. Combining clinical diagnostic results and pathology reports, the review opinions are transformed into adjustment factors to optimize grading thresholds and feature extraction rules. The calibrated results are fed back to the data hierarchical analysis and feature fusion stages, forming an iterative optimization loop to continuously improve screening accuracy. Simultaneously, the calibration data is stored in a database to accumulate experience for subsequent algorithm iterations.

[0027] 1.2.1.3 Derivation of Innovative Formulas and Efficiency Enhancement Principles To achieve precise quantification of early abnormal signals in kidney disease and overcome the limitations of traditional single-indicator screening, lack of quantitative basis, and poor population adaptability, this invention derives a quantitative formula for early abnormality measurement in kidney disease, serving as the core quantitative basis for early screening. This original formula, unreported in existing technology, integrates clinical, imaging, and genetic modalities. Through a three-order coupling logic of "feature weighting + association correction + population adaptation," it achieves precise quantification of abnormality, ensuring the discriminative value of single-modal features while also considering the synergistic correlation of multi-modal features and individual population differences. This significantly improves the sensitivity and specificity of early screening, providing scientific and quantifiable support for algorithmic decision-making. The formula design fully incorporates the pathophysiological laws of early kidney disease onset, adapting to different screening scenarios and population characteristics, exhibiting strong flexibility and practicality.

[0028] Formula expression: γ = ω·(C Wc + I ·Wi + G ·Wg ) + (1-ω)·R·A Formula parameter description: γ: Early abnormality coefficient of kidney disease, ranging from [0,1]. The larger the value, the higher the probability of early kidney disease. It is used for screening result grading and clinical decision-making. The coefficient is calculated through third-order coupling, which fully integrates multimodal features, cross-modal correlations and population adaptability to achieve quantitative characterization of early abnormal signals and avoid the subjective judgment bias of traditional screening.

[0029] ω: Modality weight coefficient, ranging from [0.6, 0.75], is used to balance the weight ratio of core feature items (weighted sum of clinical, imaging, and genetic features) and association correction items (cross-modal association + population adaptation), highlighting the discriminative value of the three modalities. This parameter can be dynamically adjusted according to the screening scenario: in scenarios with a high proportion of clinical data and missing genetic data (such as screening in primary healthcare institutions), a value of 0.7-0.75 is used to strengthen the weight of core feature items and adapt to data loss; in scenarios with complete genetic data and suspected case screening, a value of 0.6-0.65 is used to moderately increase the weight of association correction items, improve screening specificity, and reduce false positives. The value of ω can be iteratively optimized through clinical feedback data to continuously adapt to the needs of different screening scenarios.

[0030] C The standardized value of the i-th clinical laboratory feature, ranging from [0,1], covers core indicators such as eGFR, ACR, cystatin C, electrolytes, and uric acid. It is obtained through min-max normalization to eliminate dimensional differences. The higher the degree of abnormality of the clinical indicator, the higher the C value. The higher the value, the more specific the corresponding relationship is based on clinical practice guidelines: for example, eGFR < 60 ml / min / 1.73m 2 At that time, C Values ​​range from 0.6 to 1.0, with eGFR between 60 and 90 ml / min / 1.73 m. 2 Values ​​between 0.3 and 0.6 are considered within this range, and eGFR > 90 ml / min / 1.73m. 2 The values ​​are set to 0.1-0.3 for ACR > 30 mg / g, 0.6-1.0 for ACR between 30-300 mg / g, 0.3-0.6 for ACR between 30-300 mg / g, and 0.1-0.3 for ACR < 30 mg / g, to ensure that the characteristic values ​​accurately correspond to the degree of clinical abnormality.

[0031] Wc Clinical laboratory feature weight vector, satisfying ΣWc =1, with the core indicators (eGFR, ACR) weighted at 0.3-0.4 and the secondary indicators (electrolytes, uric acid) weighted at 0.05-0.15. The weights were determined using analytic hierarchy process (AHP) combined with clinical expert opinions to ensure the core clinical indicators play a dominant role in the screening results. The weight vector can be adjusted according to the type of kidney disease. For example, for screening diabetic nephropathy, the weight of ACR is appropriately increased; for screening hypertensive nephropathy, the weight of eGFR is appropriately increased to enhance the clinical suitability of the formula.

[0032] I The j-th image feature is a standardized value, ranging from [0,1], encompassing features such as renal cortical thickness, echo intensity, volume index, collecting system morphology, and blood flow signal. It is obtained through extraction and standardization using a modified CNN. The more significant the image abnormality, the higher the value of I. The larger the value, the more specific the corresponding relationship: for example, when the cortical thickness is <1cm, the value is 0.7-1.0; when the cortical thickness is between 1-1.5cm, the value is 0.3-0.7; when the cortical thickness is >1.5cm, the value is 0.1-0.3. When the echo intensity is significantly abnormal, the value is 0.7-1.0; when it is slightly abnormal, the value is 0.3-0.7; and when it is normal, the value is 0.1-0.3, so as to achieve precise quantification of the degree of image abnormality.

[0033] Wi Image feature weight vector, satisfying ΣWi =1, with a weight of 0.25-0.35 for core morphological features (cortical thickness, kidney volume) and 0.1-0.2 for secondary features (echoic homogeneity, blood flow signal). This weighting is tailored to the core features of early renal disease imaging, ensuring the full discriminative value of the imaging data is fully realized. The weight vector can be adjusted according to the type of imaging equipment; for example, the weight of echo intensity can be appropriately increased for ultrasound imaging, and the weight of volume index can be appropriately increased for CT imaging.

[0034] G : Standardized value of the k-th gene feature, ranging from [0,1]. Values ​​are 0.6-1.0 for abnormal expression of core differentially expressed genes, 0.1-0.3 for normal expression, 1.0 for pathogenic mutant genes, and 0.1 for genes without mutations. This value is obtained through gene expression profiling and standardization. For genes highly associated with early kidney disease (such as DUSP1 and GADD45B), the feature value is assigned higher discriminative power to ensure accurate capture of abnormal signals at the molecular level.

[0035] Wg Gene feature weight vector, satisfying ΣWg =1, with the weight of core genes highly correlated with early-stage kidney disease ranging from 0.2 to 0.3, and the weight of other related genes ranging from 0.05 to 0.1. This weighting is determined through gene correlation analysis and clinical validation, enhancing the discriminative value of core genes while avoiding interference from redundant gene data. In scenarios with missing gene data, this weight vector automatically becomes invalid, and the algorithm adapts to data changes by adjusting the ω value.

[0036] R: Cross-modal feature correlation coefficient, ranging from [0.8, 1.2], quantifies the strength of abnormal correlations between different modal features, compensating for the limitations of single-modal feature analysis. When multiple modal features are simultaneously abnormal (e.g., elevated clinical ACR + thinning of the cortex on imaging + abnormal gene expression), the value is set to 1.0-1.2, enhancing the discriminative value of synergistic abnormal signals; when a single modality is abnormal and other modalities are normal, the value is set to 0.8-0.9, weakening the influence of a single abnormal signal and avoiding misjudgment; when there are no obvious abnormal features, the value is set to 1.0, ensuring the stability of the formula calculation. This coefficient is calculated using a self-attention network, accurately capturing the correlation patterns between features.

[0037] A: Population fit coefficient, ranging from [0.9, 1.1], is adjusted based on age, gender, and comorbidities to accommodate the differences in baseline nephropathy incidence among different populations and avoid bias caused by homogenization of screening standards. For elderly patients (≥65 years old) and patients with comorbidities such as diabetes / hypertension / obesity, a value of 1.05-1.1 is used to increase the sensitivity of the abnormality coefficient and reflect the high incidence of nephropathy in these populations. For young healthy individuals (<40 years old, without underlying diseases), a value of 0.9-0.95 is used to decrease the sensitivity of the abnormality coefficient and reduce false positives. For the general middle-aged population, a value of 1.0 is used to maintain the neutrality of the screening standard. This coefficient can be iteratively optimized based on population epidemiological data to improve the population fit of the formula.

[0038] Formula Derivation Logic: A three-tiered modeling approach—"core feature-driven + correlation correction + population adaptation calibration"—is employed to construct a multi-dimensional collaborative quantification system, fully aligning with the complex characteristics of early-stage kidney disease. The first tier comprises the core feature term, which uses ω to weight the sum of clinical, imaging, and genetic modalities, focusing on core abnormal signals in each modality to ensure screening sensitivity. Simultaneously, feature weight allocation highlights the discriminative value of core indicators and avoids interference from secondary features. The second tier is the correlation correction term, introducing a cross-modal correlation coefficient R (1-ω) to capture collaborative abnormal signals from multiple modalities, avoiding false positives caused by single-modal feature abnormalities, improving screening specificity, and quantifying the correlation patterns between features to achieve synergistic empowerment of multi-dimensional information. The third tier uses a population adaptation coefficient A to calibrate baseline differences in the population, ensuring the formula's applicability across different populations, avoiding screening bias caused by universal standards, and improving the clinical reliability of screening results. The formula transforms multi-dimensional features into standardized abnormality coefficients through nonlinear mapping, completely abandoning traditional experience-based screening methods and achieving precise quantification of abnormal signals in early-stage kidney disease. All parameters support dynamic adjustment and can be iteratively optimized through clinical data and research results to adapt to different screening scenarios and population characteristics, demonstrating strong flexibility and practicality.

[0039] The core principle of this synergistic effect formula fundamentally addresses the limitations of traditional early screening methods, such as "single indicator limitations, lack of quantitative basis, and poor population fit," through an integrated design of multimodal collaborative quantification, correlation correction, and population adaptation. Its synergistic value is reflected in three aspects: First, it significantly improves screening sensitivity. Multimodal features comprehensively cover the molecular, structural, and functional abnormalities in early-stage kidney disease, avoiding missed diagnoses due to single indicators. In particular, it greatly enhances the ability to capture abnormal signals in asymptomatic kidney disease, accurately identifying occult kidney diseases that are difficult to detect with traditional screening, thus advancing the diagnostic and treatment process. Second, it optimizes and upgrades screening specificity. Cross-modal correlation coefficients effectively correct misjudgments caused by single-modal abnormalities, significantly reducing false positive rates, minimizing unnecessary further examinations and patient psychological burden, while enhancing the clinical reference value of screening results and providing doctors with precise decision-making guidance. Third, it significantly strengthens population fit. By calibrating individual baseline differences through population fit coefficients, it ensures that people of different ages and underlying disease states receive reasonable screening standards, avoiding insufficient screening for high-risk groups and over-screening for low-risk groups, thus improving the scientific rigor and fairness of the screening program. The formula is deeply integrated with other modules of the MDS-ES algorithm to form a closed loop of "feature extraction - fusion enhancement - quantitative calibration", which promotes the transformation of early screening from "experience-driven" to "algorithm-driven and quantitatively supported", and greatly improves screening efficiency.

[0040] 1.2.1.4 Qualitative Analysis of Unreported Examples and Synergistic Mechanisms Example: Early kidney disease screening in a primary healthcare institution, covering 5,000 high-risk individuals aged 40 and above (with hypertension, diabetes, and obesity). This scenario is characterized by limited medical resources, insufficient physician experience, limited testing equipment (no gene testing equipment), and a large population base. Traditional screening, using a single eGFR test combined with physician experience, suffers from high early-stage missed diagnoses (>60% in the asymptomatic phase), high false-positive rates (>25%), lack of quantitative basis for screening results, and subjective judgment criteria. This leads to delayed diagnosis and treatment for some early-stage kidney disease patients, increasing the difficulty of treatment and worsening the prognosis as the disease progresses to the middle and late stages. Furthermore, many false-positive patients bear unnecessary psychological burdens and testing costs, wasting medical resources. After deploying the MDS-ES algorithm, based on an original anomaly quantification formula, multimodal collaborative screening is conducted by integrating clinical laboratory and renal ultrasound data (the gene data module is disabled, and the algorithm automatically adjusts the ω value to 0.75, strengthening the weight of clinical and imaging features), addressing the pain points of traditional screening and adapting to the needs of primary healthcare scenarios.

[0041] Qualitative Analysis of the Enhancement Principle: The MDS-ES algorithm, through multidimensional modeling and precise quantitative design, achieves a comprehensive improvement in early screening efficacy. Its enhancement value is reflected in four core dimensions, requiring no data support; its clinical value can be verified through qualitative analysis. Firstly, multimodal data hierarchical analysis enables comprehensive capture of abnormal features, breaking through the limitations of traditional single-indicator screening. The algorithm extracts core indicators and dynamic changes such as eGFR and ACR at the clinical level, accurately capturing subtle early abnormalities in kidney function, such as the occult abnormality of a persistently mildly elevated ACR despite a normal eGFR range; at the imaging level, it extracts morphological features such as renal cortex thickness and echo intensity, discovering early structural lesions, such as thinning of the cortex and abnormal echo intensity—signals easily overlooked by traditional manual interpretation. The combination of these two approaches achieves multi-dimensional coverage of early kidney disease manifestations from both functional and structural perspectives, avoiding missed diagnoses based on a single dimension. For example, some patients with normal eGFR but dynamically elevated urinary protein and thinning of the cortex, who would be considered normal under traditional screening, can be accurately identified as suspected cases by the MDS-ES algorithm, providing opportunities for early intervention.

[0042] Secondly, the original quantitative formula constructs a multi-dimensional collaborative quantitative system to achieve scientific quantification and classification of screening results, abandoning the subjectivity of traditional experience-based judgment. The formula, through the coupling of core feature terms and associated correction terms, transforms abstract multimodal abnormal features into calculable abnormality coefficients, providing clear quantitative basis for screening results. Simultaneously, it achieves classification based on thresholds, ensuring the consistency and objectivity of screening standards. For high-risk groups with diabetes, the population fit coefficient A is set to 1.08, enhancing the screening sensitivity for this high-risk group; for those with normal eGFR but slightly elevated ACR (C... =0.5), slightly increased kidney echogenicity (I For patients with a clinical characteristic weight Wc = 0.45, the clinical characteristic weight is 0.45. =0.35, image feature weight Wi =0.3, the core feature calculation result is 0.75×(0.5×0.35+0.45×0.3)=0.75×0.31=0.2325; the cross-modal association coefficient R is 1.05 (abnormal association between two modalities), the population fit coefficient A=1.08, the association correction term calculation result is 0.25×1.05×1.08≈0.2835, and the final abnormality coefficient γ≈0.516, which is considered a suspected case. It is recommended to complete the 24-hour urine protein quantification test, and finally confirm the patient as an early diabetic nephropathy patient, accurately identifying cases missed by traditional screening. For a single mild decrease in eGFR (C =0.4) but imaging features are normal (I For patients with a cross-modal correlation coefficient R of 0.85 (γ = 0.2), the correlation correction term was weakened, and the final result was γ ≈ 0.28, which was considered normal. This effectively reduced the false positive rate and avoided unnecessary examinations.

[0043] Third, the attention mechanism feature fusion enables the synergistic correlation of multimodal features, improving screening accuracy and avoiding misjudgments caused by isolated analysis of individual modal data. The algorithm captures nonlinear correlations between different modal features through a self-attention network, such as the correlation between dynamically elevated urinary protein and abnormal echo intensity, and the correlation between abnormal ACR and the duration of diabetes, synergistically integrating multi-dimensional information to form a more comprehensive basis for abnormality determination. For example, a patient with a slightly elevated ACR (C... =0.4), but the kidney ultrasound showed no obvious abnormalities (I =0.2), traditional screening would easily judge it as normal, but the algorithm captured the strong association between elevated ACR and the patient's 10-year diabetes course through a self-attention network. The cross-modal association coefficient R was 0.95, combined with the population fit coefficient A=1.08, and finally γ≈0.32, which was judged as suspected. It was recommended to complete the 24-hour urine protein quantification. The final diagnosis was early diabetic nephropathy, which reflects the core value of association fusion and avoids missed diagnosis caused by isolated data analysis.

[0044] Fourth, the dual-verification iterative optimization mechanism continuously improves screening efficiency, forming a closed-loop optimization logic that adapts to the actual needs of primary healthcare scenarios. The algorithm continuously optimizes formula parameters and feature weights through internal cross-validation and external physician review. For scenarios with missing gene data in primary healthcare screening, the ω value is adjusted to 0.75, strengthening the weights of clinical and imaging features to ensure screening accuracy is not affected by missing data. Simultaneously, it incorporates clinical feedback from primary healthcare physicians to optimize feature extraction rules and grading thresholds, making the algorithm more aligned with the realities of primary healthcare. Compared to traditional screening models, the MDS-ES algorithm transforms early kidney disease screening from a "single indicator" to "multimodal collaboration" and from "experience-based judgment" to "precise quantification." It significantly reduces the missed diagnosis rate of asymptomatic kidney disease, drastically decreases the false positive rate, and significantly enhances the clinical reference value of screening results. This provides efficient algorithmic support for early kidney disease prevention and control in primary healthcare institutions, effectively promoting earlier diagnosis and treatment, reducing the incidence of late-stage kidney disease, optimizing medical resource allocation, and lowering overall healthcare costs.

[0045] 1.2.2 Nonlinear prediction algorithm for kidney disease progression 1.2.2.1 Algorithm Modeling Logic The core modeling goal of the NPP-PP algorithm is to construct a multi-factor nonlinear collaborative predictive model, addressing the core problems of existing disease prediction methods that rely on single indicators, linear models, predictive lag, and lack of quantitative support. This aims to achieve accurate prediction and dynamic assessment of the risk of kidney disease progression. The modeling adopts a four-layer architecture: "progression factor analysis - nonlinear correlation modeling - predictive model construction - result calibration and optimization." Based on the screening results of the MDS-ES algorithm, it integrates baseline disease data, treatment intervention history, and dynamic monitoring data to accurately capture the nonlinear and dynamic correlation characteristics of disease progression. This enables probability prediction, time-point prediction, and identification of key driving factors for the risk of kidney disease progression (such as CKD stage progression, AKI onset, and progression to end-stage renal disease) within the next 3-12 months, providing forward-looking decision support for clinical intervention strategy formulation. This algorithm breaks through the limitations of the traditional linear predictive framework. Its core innovation lies in constructing a predictive system of "multi-factor collaboration - nonlinear characterization - dynamic iteration," which not only covers the complex influencing factors of disease progression but also accurately captures the dynamic correlation patterns between factors, filling the industry gap in nonlinear quantitative prediction of kidney disease progression.

[0046] The progression factor analysis and modeling focuses on the systematic review of multi-dimensional influencing factors and the screening of core factors, constructing a four-category factor system: "core pathological factors, clinical dynamic factors, intervention influencing factors, and baseline risk factors," ensuring the comprehensiveness and specificity of the predictive basis. Core pathological factors are extracted from renal biopsy pathology reports and gene testing results, covering the proportion of glomerular sclerosis, the grade of renal tubular damage, the degree of interstitial fibrosis, and the expression level and mutation status of core pathogenic genes. These factors directly determine the baseline level of renal pathological damage and are the core driving factors for disease progression. For example, when the proportion of interstitial fibrosis is >20%, the risk of disease progression increases significantly and should be given high predictive weight. Clinical dynamic factors are extracted from dynamic monitoring data over the past 3-6 months, including the rate of change of eGFR, the amplitude of ACR fluctuations, the frequency of electrolyte disturbances, and the rate of blood pressure / glycemic control. This accurately captures the dynamic trends of renal function and comorbidities, avoiding the limitations of single-point-of-time indicators, such as a monthly eGFR decline rate >1 ml / min / 1.73m. 2 When the disease is in a rapid progression phase, it indicates that the condition is progressing rapidly. Intervention influencing factors encompass previous treatment regimens (drug type, dosage, duration of treatment), treatment effect feedback (degree of improvement in indicators, occurrence of adverse reactions), and patient treatment adherence. This quantifies the regulatory effect of interventions on disease progression; for example, patients who respond well to hormone combined with immunosuppressant therapy have a significantly reduced risk of progression, while those with poor adherence leading to treatment interruption have a significantly increased risk of progression. Baseline risk factors include age, gender, type and duration of comorbidities, family history, and body mass index. These factors constitute the basic risk background for disease progression. For example, elderly patients (≥65 years old) with diabetic nephropathy and a disease duration exceeding 10 years have a significantly higher baseline risk of progression than younger patients with simple kidney disease. Through correlation analysis and random forest feature importance ranking, redundant factors are eliminated, and 20-30 core progression factors are selected, reducing the computational complexity of the algorithm while ensuring the core discriminative value of the predictive factors.

[0047] Nonlinear association modeling uses an improved long short-term memory network (LSTM) combined with an attention mechanism to build the model, which breaks through the limitation of traditional linear models that cannot characterize complex associations and achieves accurate capture of nonlinear and dynamic associations between multiple factors. While traditional LSTM models can process time-series data, they lack sufficient focus on key time-point features and struggle to quantify the correlation strength between different factors. This algorithm improves modeling efficiency through two major optimizations: First, it introduces a time-series attention mechanism into the LSTM hidden layer, assigning high attention weights to key time points of clinical dynamic factors (such as indicator mutations and intervention adjustment points) to accurately capture core change signals in time-series data, such as the month of a sudden drop in eGFR and the indicator response period 1-2 months after drug adjustment, thus strengthening the impact of key time-series features on predictive results. Second, it constructs a factor association attention layer to calculate the association coefficients between different categories of factors, quantifying the nonlinear association patterns between pathological and clinical factors, and between intervention and dynamic factors, such as the positive correlation between the proportion of interstitial fibrosis and the rate of eGFR decline, and the dose-dependent association between SGLT-2 inhibitor intervention and ACR improvement, generating a multi-factor nonlinear association matrix and achieving a transformation from "isolated factor analysis" to "synergistic association characterization." Through the fusion of dual attention mechanisms, the model can simultaneously and accurately capture both time-series dynamic features and inter-factor association features, providing precise association modeling support for subsequent predictions.

[0048] The predictive model employs a multi-task learning framework, simultaneously implementing three core predictive tasks to form a three-in-one predictive result system of "risk probability - time node - driving factor". The first task is the prediction of progression risk probability. A fully connected layer maps the nonlinear correlation matrix to a progression risk probability P within the interval [0,1]. A higher value indicates a higher probability of disease progression within a specified future period. Based on clinical data, thresholds are set: P ≥ 0.7 for high progression risk, 0.3 ≤ P < 0.7 for medium progression risk, and P < 0.3 for low progression risk, thus clarifying intervention priorities. The second task is the prediction of progression time nodes. A time-series prediction branch captures the dynamic patterns of disease progression, combining risk probability to output high-probability progression time nodes (e.g., 3-6 months for high-risk patients, 6-9 months for medium-risk patients), providing precise evidence for selecting the timing of clinical intervention and avoiding overtreatment due to premature intervention or delayed treatment due to excessive intervention. The third task is to identify key driving factors. Through the factor contribution analysis module, the influence weight of each core factor on the risk of progression is quantified, and the top 3-5 key driving factors (such as interstitial fibrosis, eGFR decline rate, and poor glycemic control) are identified. The core causes of disease progression are clarified, providing direction for the formulation of targeted intervention strategies and realizing a closed-loop connection of "precise prediction - targeted intervention".

[0049] The calibration and optimization modeling employs a dual mechanism of "time-series iterative calibration + clinical feedback calibration" to ensure the accuracy and dynamic adaptability of the prediction results. Time-series iterative calibration, based on newly added dynamic monitoring data of patients, updates the prediction model input every 1-2 months, adjusting the progression risk probability, time points, and driver factor ranking to adapt to dynamic changes in the disease. For example, if eGFR stabilizes and ACR decreases after treatment, the algorithm automatically lowers the progression risk probability and postpones the progression time point; if electrolyte disturbances or worsening pathological damage occur, the risk probability is immediately raised to advance the intervention window. Clinical feedback calibration incorporates physician experience and clinical diagnostic results. A team of senior nephrology experts manually reviews high-risk and borderline risk cases, combining pathology reports and treatment effect feedback. The review opinions are transformed into model adjustment factors, optimizing factor weights, correlation coefficients, and prediction thresholds. Simultaneously, new case data is included in the model training set to improve the model's generalization ability through incremental learning. This dual calibration mechanism ensures that the prediction model remains consistent with the dynamics of the patient's condition and the needs of clinical practice, continuously improving prediction efficiency.

[0050] 1.2.2.2 Algorithm Solution Process The NPP-PP algorithm employs a full-process solution strategy of "factor collection and screening - nonlinear correlation modeling - multi-task prediction - result calibration - dynamic iteration". Using the screening results from the MDS-ES algorithm as the initial input, each step is progressive and dynamically closed-loop, ensuring the accuracy and timeliness of the prediction results. The specific process is as follows: The first step, core progression factor collection and screening: Based on the screening results (abnormality coefficient, risk level) of the MDS-ES algorithm, data on four types of core progression factors were collected. Core pathological factors were obtained from renal biopsy reports and gene testing platforms, including key indicators such as the proportion of glomerulosclerosis and the degree of interstitial fibrosis; clinical dynamic factors were extracted from the LIS system using dynamic monitoring data such as eGFR and ACR over the past 3-6 months, and derived characteristics such as rate of change and fluctuation amplitude were calculated; intervention impact factors were collected from the HIS system and physician terminals, covering previous treatment plans, effect feedback, and compliance scores; baseline risk factors were extracted from the patient baseline database, including information such as age, comorbidities, and family history. A random forest algorithm was used to rank the features by importance, eliminating redundant factors with importance scores below 0.1, and selecting 25 core progression factors. Min-max normalization was used to eliminate dimensional differences, generating a standardized factor matrix to provide high-quality input for nonlinear correlation modeling.

[0051] The second step is multi-factor nonlinear association modeling: The standardized factor matrix is ​​input into an improved LSTM+dual attention mechanism model, and association modeling is completed in three stages. First, the LSTM temporal layer processes clinical dynamic factors to capture the dynamic patterns of indicator changes over time and outputs a temporal feature vector. Second, the temporal attention layer strengthens the feature weights of key time nodes, focusing on core temporal signals such as indicator mutations and intervention adjustments. Finally, the factor association attention layer calculates the association coefficients between different categories of factors, generating a multi-factor nonlinear association matrix to quantify the complex association patterns between pathological, clinical, intervention, and baseline factors, providing association support for multi-task prediction. Dropout regularization is used during the modeling process to prevent overfitting and ensure the model's generalization ability.

[0052] The third step is multi-task collaborative prediction: Based on a nonlinear correlation matrix, three prediction tasks are completed simultaneously through a multi-task learning framework. The progression risk probability prediction outputs a probability value P in the [0,1] interval through a fully connected layer, and classifies it into high, medium, and low risk levels based on preset thresholds. The progression time point prediction uses a time-series prediction branch to output the most probable progression time point within the next 3-12 months based on risk probability and dynamic factor change trends. The key driving factor identification uses factor contribution analysis to calculate the influence weight of each core factor on the prediction result, and outputs the top 5 key driving factors after ranking. A standardized prediction report is generated, clearly defining the risk level, probability value, progression time, core causes, and preliminary intervention suggestions, and is pushed to the clinical decision-making terminal.

[0053] The fourth step involves dual calibration of the prediction results: a dual mechanism of temporal iteration and clinical feedback is employed to calibrate the results. Internal temporal calibration uses 10-fold cross-validation to assess prediction accuracy and time-point deviation based on historical case datasets. For cases with significant prediction deviations, the model's attention weights and correlation coefficients are adjusted to optimize the prediction threshold. External clinical calibration involves at least two senior experts manually reviewing high-risk and borderline-risk cases. Combining pathology reports and treatment outcomes, the review opinions are translated into adjustment parameters to optimize factor weights and prediction logic. The calibrated results are simultaneously fed back to the factor selection and correlation modeling stages, forming a preliminary closed loop.

[0054] The fifth step is dynamic iterative updates: Based on newly added dynamic monitoring data of patients and the effects of treatment interventions, the predictive model is iteratively updated every 1-2 months. This involves supplementing new factor data, adjusting the ranking and weights of core factors, updating the nonlinear correlation matrix, and regenerating predictive results to achieve real-time adaptation of the predictive model to the dynamics of the patient's condition. Simultaneously, new case data and calibration results are incorporated into the model training set, and incremental learning is used to continuously optimize the model structure and parameters, improving predictive accuracy and generalization ability, forming a dynamic closed loop of "prediction-intervention-monitoring-iteration".

[0055] 1.2.2.3 Derivation of Innovative Formulas and Efficiency Enhancement Principles To achieve accurate quantitative prediction of kidney disease progression and overcome the limitations of traditional linear models, such as predictive lag, lack of specific quantitative formulas, and insufficient characterization of factor correlations, this invention derives a kidney disease progression risk quantification formula, which serves as the core quantitative basis for the NPP-PP algorithm. This original formula, unreported in existing technology, integrates four core progression factors and achieves accurate quantification of progression risk through a three-order coupling logic of "core factor weighting + nonlinear correlation correction + time-series dynamic calibration." It also considers the correlation patterns between factors and the dynamic changes in the disease, providing scientific quantitative support for multi-task prediction and significantly improving the accuracy and foresight of predictions. The formula design fully aligns with the pathophysiological laws of kidney disease progression, adapts to different types and stages of kidney disease, and possesses strong clinical applicability.

[0056] Formula expression: P = μ·(P ·W + P_c·W_c + P_i·W_i) + (1-μ)·T·(1 - P_b·W_b) Formula parameter description: P: Probability of disease progression risk, ranging from [0,1]. A higher value indicates a higher probability of disease progression (stage advancement, AKI onset, or progression to end-stage renal disease) within the next 3-12 months. It is used for risk level classification and intervention priority determination. The probability is calculated through third-order coupling, comprehensively integrating core progression factors, nonlinear correlations, time-series dynamics, and baseline risk to achieve accurate quantification of progression risk and avoid the subjective bias of traditional linear prediction.

[0057] μ: Core factor weight coefficient, ranging from [0.65, 0.8], is used to balance the weight ratio of core factor items (weighted sum of pathology, clinical dynamics, and intervention factors) and time-series correction items (time-series coefficient + baseline risk calibration), highlighting the dominant role of the three core factors in progression risk. This parameter can be dynamically adjusted according to the type of kidney disease: for kidney diseases with significant pathological damage (such as focal segmental glomerulosclerosis), a value of 0.75-0.8 is used to strengthen the weight of core factor items; for kidney diseases dominated by clinical dynamic changes (such as the recovery phase of AKI), a value of 0.65-0.7 is used to moderately increase the weight of the time-series correction item, adapting to different disease characteristics. The value of μ is iteratively optimized through clinical feedback data and cross-validation to continuously improve the formula's adaptability.

[0058] P : Core pathological factor comprehensive value, ranging from [0,1], is obtained by weighted summation of indicators such as the proportion of glomerular sclerosis, the degree of interstitial fibrosis, the grade of renal tubular damage, and the expression level of core genes. The more severe the pathological damage, the higher the value of P. The higher the value, the better. Specific calculation rules: When the proportion of glomerular sclerosis is >30% and the proportion of interstitial fibrosis is >25%, the value for a single indicator is 0.8-1.0; when the proportion is between 10%-30% and 15%-25%, the value is 0.4-0.8; when the proportion is <10% and <15%, the value is 0.1-0.4; when the core pathogenic gene is highly expressed or has pathogenic mutations, add an additional 0.2 to ensure the core driving value of the pathological factor.

[0059] W : Core pathological factor weight vector, satisfying ΣW =1, with interstitial fibrosis degree and glomerular sclerosis ratio weighted at 0.3-0.35, and renal tubular damage grade and gene characteristics weighted at 0.15-0.2. These values ​​are determined through analytic hierarchy process combined with expert pathology opinions, highlighting the pathological indicators that have the most significant impact on disease progression and adapting to the pathological characteristics of different kidney diseases. For example, the weight of interstitial fibrosis is appropriately increased in diabetic nephropathy, and the weight of glomerular sclerosis is appropriately increased in IgA nephropathy.

[0060] P_c: A comprehensive value of dynamic clinical factors, ranging from [0,1]. It integrates dynamic indicators such as the rate of change of eGFR, the amplitude of ACR fluctuations, the rate of blood pressure / blood glucose control achievement, and the frequency of electrolyte disturbances. The more significant the dynamic deterioration of the condition, the larger the P_c value. Calculation rule: monthly eGFR decline rate > 1 ml / min / 1.73m 2 When ACR continues to rise and fluctuates by more than 50%, the value for a single item should be 0.7-1.0; the rate should be 0.5-1 ml / min / 1.73m. 2 When the fluctuation range is between 20% and 50%, the value should be 0.3-0.7; the rate should be <0.5 ml / min / 1.73m. 2 When the fluctuation range is <20%, the value is 0.1-0.3; when the blood pressure / blood sugar control target achievement rate is <60%, add an additional 0.15 to accurately capture clinical dynamic deterioration signals.

[0061] W_c: Clinical dynamic factor weight vector, satisfying ΣW_c=1, with eGFR change rate and ACR fluctuation amplitude weighted at 0.3-0.35, and blood pressure / blood glucose control target achievement rate and electrolyte disturbance frequency weighted at 0.15-0.2. It is adapted to the characteristics of different comorbidities, such as increasing the weight of blood pressure control target achievement rate in cases of hypertensive nephropathy, and increasing the weight of blood glucose control target achievement rate in cases of diabetic nephropathy, to ensure the targeting of dynamic factors.

[0062] P_i: Comprehensive value of intervention impact factors, ranging from [0,1], covering the effectiveness of the treatment plan, treatment compliance, and the occurrence of adverse reactions. The worse the intervention effect and the lower the compliance, the larger the P_i value. Calculation rules: When there is no improvement or worsening of indicators after treatment and the compliance score is <60 points (out of 100), the value is 0.7-1.0; when the indicators partially improve and the compliance score is between 60 and 80 points, the value is 0.3-0.7; when the indicators significantly improve and the compliance score is >80 points, the value is 0.1-0.3; when serious adverse reactions occur and treatment is interrupted, an additional 0.2 is added to quantify the regulatory effect of the intervention on the risk of progression.

[0063] W_i: The weight vector of intervention influencing factors, satisfying ΣW_i=1. The weight of treatment effectiveness is 0.4-0.5, and the weight of treatment compliance and adverse reactions is 0.25-0.3, highlighting the core influence of treatment effect, while taking into account the regulatory value of compliance and safety on disease progression.

[0064] T: Time-series correction coefficient, ranging from [0.85, 1.15], quantifies the impact of the time-series changes in clinical dynamic factors on the risk of disease progression, compensating for the limitations of static factor analysis. A value of 1.05-1.15 is used when the condition shows a continuous worsening trend (e.g., a continuous decline in eGFR for 3 months), reinforcing the risk of progression; a value of 0.85-0.95 is used when the condition tends to stabilize or improve, weakening the risk of progression; and a value of 1.0 is used when the condition fluctuates without a clear trend, ensuring the stability of the formula calculation. This coefficient is calculated using an improved LSTM model, accurately capturing the time-series dynamic patterns.

[0065] P_b: The composite value of baseline risk factors, ranging from [0,1], is composed of indicators such as age, type and duration of comorbidities, and family history. The lower the baseline risk, the smaller the P_b value (reverse calibration of progression risk). Calculation rules: For ages ≥65 years, comorbidity duration exceeding 10 years, and a family history of kidney disease, the value is 0.1-0.3 (high baseline risk, reverse calibration strength is reduced); for ages 40-65 years, comorbidity duration 5-10 years, and no family history, the value is 0.4-0.6; for ages <40 years, no comorbidities, and no family history, the value is 0.7-0.9 (low baseline risk, reverse calibration strength is enhanced), achieving accurate calibration of baseline risk to progression probability.

[0066] W_b: Baseline risk factor weight vector, satisfying ΣW_b=1, with age and duration of comorbidities weighted at 0.3-0.35, and family history and body mass index weighted at 0.15-0.2, to adapt to the baseline risk characteristics of different populations and ensure the scientific nature of baseline calibration.

[0067] Formula Derivation Logic: A three-tiered modeling approach—"core factor-driven + time-series correlation correction + baseline reverse calibration"—is adopted to align with the complex and dynamic characteristics of kidney disease progression, achieving multi-dimensional quantitative prediction. The first tier comprises the core factor term, which uses a μ-weighted sum of pathological, clinical dynamic, and interventional core factors to focus on the direct drivers of disease progression, ensuring the forward-looking nature of the prediction. Simultaneously, factor weight allocation highlights the dominant role of key factors and avoids interference from secondary factors. The second tier is the time-series correction term, which introduces a time-series correction coefficient T (1-μ) to capture the temporal trends of clinical dynamic factors, accurately depicting the dynamic patterns of disease progression and avoiding the predictive lag caused by static factor analysis, thus achieving the core goal of "dynamic prediction." The third tier uses a reverse calibration between the baseline risk factor P_b and its weight W_b to balance the progression probability of different baseline risk populations, avoiding under-prediction for high-baseline-risk populations and over-prediction for low-baseline-risk populations, improving the formula's population adaptability. The formula integrates multi-dimensional information through nonlinear mapping, transforming abstract factors influencing disease progression into standardized risk probabilities. This completely abandons traditional empirical and linear prediction methods, achieving precise quantification of the risk of kidney disease progression. All parameters support dynamic adjustment and can be iteratively optimized through clinical data and research findings to adapt to different types and stages of kidney disease, demonstrating exceptional flexibility.

[0068] The core principle of this synergistic effect formula fundamentally addresses the limitations of traditional disease prediction—namely, its linearity, lag, and lack of quantitative evidence—through an integrated design that integrates multi-factor synergy, temporal dynamics, and baseline calibration. Its synergistic value manifests in three aspects: First, it significantly improves predictive accuracy. The multi-factor system comprehensively covers pathological, clinical, interventional, and baseline dimensions. Non-linear correlation correction accurately captures the complex patterns between factors, avoiding the predictive biases of single indicators or linear models. In particular, its ability to capture early, occult progression signals is greatly enhanced, allowing for the prediction of disease progression risk 3-6 months in advance, thus providing a window for early intervention. Second, it significantly optimizes the prospective potential of the prediction. The temporal correction coefficient accurately depicts the dynamic trends of disease changes, breaking through the limitations of traditional static prediction and enabling a shift from "retrospective analysis" to "prospective prediction." Simultaneously, by identifying key driving factors, it provides targeted intervention directions for clinicians, avoiding blind intervention. Third, the adaptability between the population and the disease is enhanced. Baseline risk is reverse-calibrated to balance the baseline differences among different populations. Parameters are dynamically adjusted to adapt to different types and stages of kidney disease, ensuring that the formula can play an accurate predictive role in various scenarios such as CKD, AKI, and nephrotic syndrome. At the same time, through dynamic iteration and clinical calibration, the predictive efficacy is continuously improved, providing scientific quantitative support for the formulation of clinical intervention strategies and promoting the transformation of kidney disease diagnosis and treatment from "passive response" to "proactive prevention and control".

[0069] 1.2.2.4 Qualitative Analysis of Unreported Examples and Synergistic Mechanisms Example: A scenario for predicting the progression of CKD patients in a tertiary hospital, including 100 CKD stage 2-3 patients (including three subtypes: diabetic nephropathy, hypertensive nephropathy, and IgA nephropathy). This scenario is characterized by strong heterogeneity in patient conditions, complex comorbidities, diverse treatment options, and the need for accurate prediction of progression risk to optimize intervention strategies. Traditional disease prediction uses a single static assessment of eGFR, combined with physician experience to judge progression risk, which has three major pain points: First, it cannot capture the complex correlation between pathological damage, treatment intervention, and clinical indicators, and it is insufficient in identifying latent progression drivers such as interstitial fibrosis, resulting in significant prediction lag, often only discovering disease progression after a sharp drop in eGFR; second, there is no quantitative basis for prediction, the risk level classification is vague, and the intervention priority is difficult to define, which easily leads to untimely intervention for high-risk patients and overtreatment for low-risk patients; third, it cannot accurately identify key progression drivers, the intervention strategy lacks targeting, and homogeneous treatment is unable to curb disease progression. After deploying the NPP-PP algorithm, based on the original progress risk quantification formula, it integrates four types of factors—pathology, clinical dynamics, intervention, and baseline—to carry out nonlinear collaborative prediction, breaking through the pain points of traditional prediction and providing precise support for optimizing clinical intervention strategies.

[0070] Qualitative Analysis of the Enhancement Principle: The NPP-PP algorithm achieves a comprehensive upgrade in the predictive efficacy of kidney disease progression through multi-factor nonlinear modeling and quantitative predictive design. Its core value does not rely on specific data and can be verified through qualitative analysis, specifically reflected in four dimensions. First, the multi-dimensional progression factor system achieves comprehensive coverage of progression-driving factors, breaking through the limitations of traditional single-indicator prediction. The algorithm constructs a four-factor system of "pathology-clinical-intervention-baseline," focusing on core pathological driving factors such as interstitial fibrosis and glomerulosclerosis, while also capturing clinical dynamic signals such as the rate of change of eGFR and the amplitude of ACR fluctuations. Simultaneously, it considers intervention factors such as treatment effectiveness and compliance, as well as baseline risks such as age and the duration of comorbidities, forming a complete network of progression-influencing factors. For example, in a CKD stage 3 diabetic nephropathy patient, the eGFR is within a stable range (62 ml / min / 1.73 ml / min). 2 Traditional predictions indicated a low risk of progression, but the NPP-PP algorithm detected a mesenchymal fibrosis rate of up to 22% (P) through pathological factors. =0.65), the clinical dynamic factor showed that the ACR fluctuation range reached 45% in the past 3 months (P_c=0.5), the intervention factor indicated that the adherence to SGLT-2 inhibitor treatment was only 70% (P_i=0.45), and the baseline factor showed that the duration of diabetes was 12 years (P_b=0.25). The multi-factor synergistic analysis accurately identified the potential high risk of progression and avoided misjudgment caused by traditional single indicators.

[0071] Secondly, the original quantitative formula constructs a nonlinear collaborative quantitative system to achieve precise quantification and grading of progression risk, abandoning the subjectivity of traditional experience-based predictions. The formula, through a three-order coupling of core factor dominance, time-series correction, and baseline calibration, transforms multi-dimensional abstract factors into a calculable progression risk probability P, while clearly defining grading thresholds, providing a clear basis for intervention priority division. For the aforementioned diabetic nephropathy patients, the core factor weight coefficient μ is set to 0.75 (diabetic nephropathy pathological factors are dominant), and the pathological factor weight W... =0.35, clinical dynamic factor weight W_c=0.35, intervention factor weight Wi=0.3, the calculation result of the core factor item is 0.75×(0.65×0.35+0.5×0.35+0.45×0.3)=0.75×(0.2275+0.175+0.135)=0.75×0.5375≈0.403; the time-series correction coefficient T is taken as 1.05 (ACR continues to fluctuate, and the condition shows a potential worsening trend), baseline factor With a sub-weight W_b=0.35, the baseline calibration result is (1-0.75)×1.05×(1-0.25×0.35)=0.25×1.05×(1-0.0875)=0.25×1.05×0.9125≈0.241, and the final progression risk probability P≈0.644, indicating an intermediate progression risk. This suggests a high probability of staging progression within the next 6-9 months, and recommends enhanced intervention measures (adjusting medication dosage and improving adherence monitoring). Traditional single eGFR assessment easily overlooks such latent progression risks, leading to delayed intervention.

[0072] Third, the improved LSTM+dual attention mechanism achieves precise capture of multi-factor nonlinear correlations and temporal dynamics, overcoming the predictive limitations of traditional linear models. The algorithm focuses on key time-point features through a temporal attention mechanism, such as the response period of indicators 1-2 months after drug adjustment and eGFR mutation nodes. Simultaneously, it quantifies the nonlinear correlations between different factors through a factor-related attention layer, such as the positive correlation between interstitial fibrosis and the rate of eGFR decline, and the dose-dependent relationship between treatment adherence and ACR improvement, avoiding predictive biases caused by isolated factor analysis. For example, a patient with stage 2 CKD hypertensive nephropathy experienced a sudden drop in eGFR of 5 ml / min / 1.73 ml / min over the past month. 2 (P_c=0.6), traditional prediction easily identifies this as a high risk of progression, but the algorithm, through factor correlation attention layer, found that a sudden drop in eGFR is directly related to insufficient recent blood pressure control (40%), and there is no significant aggravation of pathological damage (P). =0.2), the intervention factor showed that the antihypertensive medication had not yet reached a steady state after adjustment (P_i=0.3), the time-series correction coefficient T was 0.9 (single-node fluctuation, no continuous worsening trend), the baseline factor P_b=0.5 (8-year hypertension duration), and the final calculated P≈0.38, indicating a low to moderate risk of progression. It is recommended to strengthen blood pressure monitoring and medication adherence management, without excessive intervention. This result accurately distinguishes between "reversible index fluctuations" and "irreversible disease progression," avoiding the over-treatment tendency of traditional predictions.

[0073] Fourth, the multi-task prediction and dynamic iterative calibration mechanism enables a closed-loop empowerment of "prediction-intervention-optimization," enhancing the clinical translational value of prediction results. The algorithm simultaneously outputs the probability of progression risk, time points, and key driving factors, providing a clear direction for targeted intervention. Simultaneously, through dynamic iterative calibration every 1-2 months, it adapts to changes in disease progression and intervention effects, continuously optimizing prediction results. For example, in the aforementioned diabetic nephropathy patient with intermediate progression risk, after intensive medication intervention and adherence management, new data two months later showed that the ACR fluctuation range decreased to 18% (P_c=0.2), and treatment adherence improved to 90% (P_i=0.2). The algorithm, through dynamic iterative updating of factor data, recalculated P≈0.32, downgraded the risk level to low progression risk, delayed the intervention window, and adjusted the intervention strategy to routine monitoring. Conversely, if the patient's pathological damage worsens after intervention (interstitial fibrosis rate rises to 28%), the algorithm will immediately increase the risk probability, advance the intervention window, and optimize the intervention plan. This dynamic closed-loop mechanism ensures that the predictive model remains synchronized with the patient's condition and the effectiveness of interventions, avoiding the lag of traditional static predictions and enabling a shift from "passively responding to progress" to "actively preventing and controlling progress."

[0074] In summary, compared to traditional prediction methods, the NPP-PP algorithm, through comprehensive multi-factor coverage, nonlinear correlation characterization, quantitative and accurate prediction, and dynamic iterative optimization, has transformed the prediction of kidney disease progression from "single static indicator" to "multi-factor dynamic synergy" and from "experience-based subjective judgment" to "algorithm-based quantitative support." This significantly improves the accuracy and foresight of predictions, enabling the identification of potential progression risks 3-6 months in advance, clarifying core driving factors and intervention timing, providing scientific support for optimizing clinical intervention strategies and improving patient prognosis, while avoiding overtreatment and delayed intervention, optimizing the allocation of medical resources, and possessing significant clinical application value.

[0075] 1.2.3 Adaptive Optimization Algorithm for Individualized Treatment Plans 1.2.3.1 Algorithm Modeling Logic The core modeling goal of the ATO-TS algorithm is to construct an adaptive optimization model that integrates "patient individual characteristics - treatment plan - efficacy feedback," addressing the core issues of existing treatment plans such as homogenization, lack of individual adaptability, delayed efficacy assessment, and lack of quantitative basis for plan adjustments. This aims to achieve integrated empowerment for individualized generation, dynamic optimization, and risk warning of kidney disease treatment plans. The modeling adopts a four-layer architecture: "individual feature analysis - plan generation modeling - efficacy prediction and risk assessment - dynamic optimization calibration." Using MDS-ES screening results and NPP-PP prediction results as inputs, it integrates multi-dimensional individual characteristics such as patient genetic background, pathological features, physiological indicators, comorbidities, treatment preferences, and drug tolerance. Combined with clinical guidelines and massive case data, it generates an initial treatment plan suitable for the individual, while simultaneously predicting the efficacy and potential adverse reaction risks of the plan. Based on real-time efficacy feedback, it dynamically optimizes the plan parameters, forming a closed-loop logic of "plan generation - efficacy prediction - risk management - optimization iteration." This algorithm breaks through the limitations of traditional standardized treatment recommendations. Its core innovation lies in building an integrated system of "individual adaptation, efficacy prediction, and dynamic optimization". By quantifying the model to balance efficacy, safety, and patient tolerance, it fills the industry gap in adaptive optimization of individualized treatment plans for kidney disease and provides precise and personalized treatment decision support for clinical practice.

[0076] Individual feature analysis and modeling focuses on the systematic review and core feature extraction of multi-dimensional individual differences, constructing a five-category feature system of "genes, pathology, physiology, clinical, and preference" to ensure the individual suitability of treatment plans. Gene features are extracted based on gene testing results, covering the expression levels and mutation status of drug metabolism-related genes (such as CYP450 family genes), nephrotic pathogenic genes, and drug sensitivity genes. This quantifies the impact of genetic background on drug efficacy and adverse reactions, such as the difference in immunosuppressant metabolic rates caused by CYP3A4 gene polymorphism, providing a basis for drug type and dosage selection. Pathological features are extracted based on renal biopsy reports, including the proportion of glomerular sclerosis, the degree of interstitial fibrosis, and the type and grade of renal tubular injury. This allows for the adaptation of treatment strategies to different pathological injury characteristics; for example, for mesangial proliferative IgA nephropathy, immunomodulatory intervention is preferentially recommended; for nephropathy primarily involving renal tubular injury, the application of renal protective drugs is strengthened. Physiological features cover basic physiological indicators such as liver and kidney function, electrolyte levels, blood glucose and blood pressure control status, and body mass index, assessing the patient's tolerance to the treatment plan and avoiding plans exceeding the patient's physiological tolerance range. Clinical characteristics include the type and course of comorbidities, previous treatment history, drug allergy history, and adverse reaction records. This helps avoid interactions with medications used to treat comorbidities and prevents the repeated use of previously ineffective or adverse-causing drugs. Preference characteristics are extracted from patient and family surveys, including needs for treatment convenience (e.g., oral medications preferred over injectable medications), affordability, and quality of life expectations. This balances medical efficacy with patient adherence, improving treatment plan implementation. Through feature importance ranking and redundancy removal, 25-30 core individual characteristics are selected to construct a standardized individual characteristic matrix, providing precise input for treatment plan generation modeling.

[0077] The protocol generation model employs an improved collaborative filtering algorithm combined with clinical guideline constraints to construct a three-tiered generation logic of "guideline adaptation - individual matching - efficacy prediction," achieving a precise fusion of standardized guidelines and individualized needs. First, a basic protocol library is built based on clinical guidelines, covering first- and second-line treatment protocols for various kidney diseases, including drug types, dosages, treatment durations, rules for combined medication, and non-pharmacological interventions (dietary management, exercise guidance, and blood pressure and blood sugar control targets), ensuring the compliance and scientific validity of the protocols. Second, through a collaborative filtering algorithm, core individual patient characteristics are matched with massive historical case data to screen out cases with similar pathological types, genetic backgrounds, and clinical characteristics. The core parameters of their effective treatment protocols are extracted, and protocol details are adjusted based on individual differences, such as adjusting drug dosages for patients with drug metabolism gene mutations (reducing dosages for slow-metabolizing patients and increasing dosages for fast-metabolizing patients). Finally, an efficacy prediction branch is introduced. Based on individual characteristics and protocol parameters, the improvement in core indicators such as eGFR and ACR after treatment is predicted, initially screening 2-3 candidate protocols with optimal efficacy to support subsequent risk assessment and final protocol determination. Constraints are set during the protocol generation process to avoid issues such as drug allergies, medication conflicts due to comorbidities, and exceeding physiological tolerance limits, thus ensuring the safety of the protocol.

[0078] The efficacy prediction and risk assessment modeling employs a multi-task learning framework to simultaneously achieve quantitative prediction of efficacy and risk assessment of adverse reactions, providing a dual basis for treatment selection and optimization. Efficacy prediction maps individual characteristics and treatment parameters to the improvement magnitude of core indicators (such as the percentage increase in eGFR and the percentage decrease in ACR) through a fully connected layer, predicting the efficacy level 1-3 months after treatment, categorized into four levels: "significant improvement, partial improvement, no improvement, and worsening." Adverse reaction risk assessment focuses on common adverse reactions in kidney disease treatment (such as infection, liver injury, electrolyte disturbances, and gastrointestinal reactions). It quantifies the probability of occurrence of each adverse reaction through a logistic regression model, combining individual characteristics (such as immune function status and liver function level) to determine the risk level, categorized into three levels: "high risk," "medium risk," and "low risk." Based on a comprehensive assessment of efficacy and risk, the analytic hierarchy process (AHP) was used to determine the weights (efficacy weight 0.6, risk weight 0.4) to comprehensively score the candidate regimens and select the final regimen with the "best efficacy and lowest risk". At the same time, key monitoring indicators and risk control measures during the regimen implementation process were identified, such as enhanced immune function monitoring for patients at high risk of infection and regular liver function tests for patients at risk of liver damage.

[0079] Dynamic optimization and calibration modeling constructs a closed-loop mechanism of "real-time efficacy feedback - risk monitoring - protocol iteration" to ensure that the treatment protocol is always dynamically adapted to the patient's efficacy response and risk status. Based on dynamic monitoring data during the patient's treatment process (changes in core indicators, occurrence of adverse reactions, and compliance feedback), the efficacy and risk of the protocol are assessed in real time. If the efficacy meets expectations and there are no significant adverse reactions, the protocol remains unchanged; if the efficacy does not meet expectations (e.g., ACR decrease <20%), the core reasons are analyzed (e.g., insufficient drug dosage, individual tolerance differences, poor compliance), and protocol parameters are optimized (adjusting drug dosage, changing drug types, and optimizing combination therapy); if adverse reactions occur, the protocol is adjusted according to the risk level (adjusting dosage for mild adverse reactions, changing drugs for moderate reactions, and immediately discontinuing medication and initiating symptomatic treatment for severe reactions). Simultaneously, clinical physician experience feedback is incorporated, and protocol optimization suggestions are manually reviewed and adjusted based on clinical judgment to ensure the clinical suitability of the protocol. Protocol optimization iterations are completed every 2-4 weeks based on the latest monitoring data, forming a dynamic closed loop of "generation-execution-monitoring-optimization" to continuously improve the individualized suitability and efficacy of the treatment protocol.

[0080] 1.2.3.2 Algorithm Solution Process The ATO-TS algorithm employs a full-process solution strategy encompassing "individual feature collection and analysis - candidate treatment generation - efficacy and risk assessment - final treatment determination - dynamic optimization iteration." Based on preliminary screening and prediction results, each step is closely linked and dynamically closed-loop, ensuring the scientific rigor, individualization, and safety of the treatment plan. The specific process is as follows: The first step, individual feature collection and analysis: Based on the anomaly coefficient of the MDS-ES algorithm, the progression risk probability and key driving factors of the NPP-PP algorithm, five types of core individual feature data were collected. Genetic features were obtained from a gene testing platform, covering the expression levels and mutation status of genes related to drug metabolism, pathogenesis, and sensitivity; pathological features were extracted from renal biopsy reports to clarify the type and extent of pathological damage; physiological and clinical features were extracted from HIS and LIS systems, including liver and kidney function, comorbidities, previous treatment history, and drug allergy history; preference features were obtained through patient questionnaires to clarify treatment needs and tolerance limits. A random forest algorithm was used to rank feature importance, eliminating redundant features with importance scores below 0.12, selecting 28 core individual features, which were then standardized to generate an individual feature matrix, providing input for solution generation.

[0081] The second step is candidate protocol generation: a basic treatment protocol library is built based on clinical guidelines, integrating standardized treatment strategies and medication rules for various kidney diseases; through an improved collaborative filtering algorithm, the patient's core characteristics are matched with the historical case database (100,000+ kidney disease cases) to screen out 10-15 highly similar cases, and the core parameters of their effective protocols are extracted; the protocol details are adjusted in combination with the patient's individual characteristics, optimizing the types, dosages, courses of treatment and combination therapy, avoiding drug conflicts and allergy risks, generating 3 initial candidate protocols, and providing non-pharmacological intervention suggestions (diet, exercise, monitoring frequency).

[0082] The third step is efficacy and risk assessment: Candidate protocol parameters and individual feature matrices are input into a multi-task learning model to simultaneously complete efficacy prediction and risk assessment. Efficacy prediction outputs the improvement in core indicators such as eGFR and ACR after 3 months of treatment, classifying efficacy levels. Risk assessment quantifies the probability of adverse reactions such as infection and liver damage, classifying risk levels. The analytic hierarchy process (AHP) is used to calculate a comprehensive score for candidate protocols (efficacy score × 0.6 + risk score × 0.4), with the highest-scoring protocol being the preferred recommendation, and the rest as alternatives.

[0083] The fourth step is finalizing and implementing the treatment plan: Invite 1-2 senior nephrologists to manually review the recommended treatment plan and adjust the details of the plan based on clinical experience (such as fine-tuning drug dosage and optimizing monitoring indicators). After confirming that there are no safety risks or implementation obstacles, generate the final individualized treatment plan, which clarifies the medication list, dosage, course of treatment, monitoring indicators and frequency, risk control measures and non-drug intervention suggestions. The plan is then pushed to the clinical decision-making terminal and patient management platform to guide clinical implementation and patient home cooperation.

[0084] Step 5, Dynamic Optimization and Iteration: Patient treatment feedback data is collected every 3 weeks, including core indicator test results, adverse reaction occurrences, and treatment adherence scores. The efficacy and risks of the treatment plan are assessed based on the feedback data. If the efficacy is significant and there are no adverse reactions, the plan is maintained and the monitoring frequency is continued. If the efficacy is insufficient or adverse reactions occur, the causes are analyzed and the plan parameters are optimized (adjusting dosage, changing medications, optimizing combination therapy). If serious adverse reactions occur, medication is immediately discontinued and an emergency treatment plan is initiated. The optimized plan is fed back to the clinician for review and confirmation, the new plan is implemented, and continuous monitoring is maintained, forming a dynamic iterative closed loop until the treatment cycle ends or the condition reaches a stable state.

[0085] 1.2.3.3 Derivation of Innovative Formulas and Efficiency Enhancement Principles To achieve quantitative assessment of the efficacy and risk of individualized treatment plans, and to overcome the limitations of traditional plan assessments such as subjectivity, lack of dedicated quantitative models, and poor individual adaptability, this invention derives a comprehensive adaptability quantification formula for individualized treatment plans for kidney disease, serving as the core quantitative basis of the ATO-TS algorithm. This formula is original and not reported in existing technology. It integrates three dimensions: individual characteristic adaptability, efficacy prediction, and risk assessment. Through a three-order coupling logic of "adaptability-driven + efficacy-risk balance + dynamic feedback calibration," it achieves precise quantification of the comprehensive value of the plan, providing a scientific basis for plan selection and optimization. Simultaneously, it considers individual differences and clinical safety, significantly improving the individual adaptability of treatment plans and the ability to manage efficacy and risk. The formula design aligns with the clinical patterns of kidney disease treatment, adapting to different types of kidney disease, individual characteristics, and treatment scenarios, demonstrating strong clinical applicability.

[0086] Formula expression: S = α·F + β·(E - λ·R) + (1-α-β)·D Formula parameter description: S: Comprehensive treatment plan suitability score, ranging from [0, 10]. A higher score indicates better individualized suitability, efficacy, and safety of the treatment plan. It is used for ranking and screening candidate treatment plans (S≥7 points indicates a preferred recommended plan, 5≤S<7 points indicates a candidate plan, and S<5 points requires regenerating the plan). The score is obtained through third-order coupling calculation, comprehensively integrating individual suitability, efficacy risk, and dynamic feedback to achieve a quantitative assessment of the comprehensive value of the treatment plan, avoiding the bias of traditional subjective assessment.

[0087] α: Individual trait fit weighting coefficient, ranging from [0.35, 0.45]. This parameter highlights the dominant role of the fit between individual characteristics and the treatment plan in the overall score; fit is a core prerequisite for individualized treatment. This parameter can be dynamically adjusted according to the type of kidney disease: for kidney diseases where genetic background significantly impacts treatment (e.g., hereditary kidney disease), a value of 0.4-0.45 is used to strengthen the weighting of gene fit; for kidney diseases where pathological type dominates treatment (e.g., focal segmental glomerulosclerosis), a value of 0.35-0.4 is used to appropriately balance pathological fit with other dimensions. The value of α is iteratively optimized through clinical case validation and expert consensus to ensure the scientific rigor of the fit assessment.

[0088] F: Individual characteristic fit coefficient, ranging from [0,10]. It quantifies the degree of fit between the treatment plan and the patient's five core characteristics: genetics, pathology, physiology, clinical characteristics, and preferences. The higher the fit, the larger the F value. Calculation rules: Genetic characteristic fit (match between drug metabolism genes and medication, no contraindications due to sensitive genes) scores 2-3 points; Pathological characteristic fit (the treatment plan matches the type and degree of pathological damage) scores 2-3 points; Physiological and clinical characteristic fit (no drug conflicts, within the physiological tolerance range, avoidance of previously ineffective / allergenic drugs) scores 2-2.5 points; Preference characteristic fit (meeting the needs of treatment convenience and economic affordability) scores 1.5-2 points. The sum of the scores of each item is the F value. A total score of less than 3 points indicates extremely poor fit, and the treatment plan should be directly eliminated.

[0089] β: Weighting coefficient for the efficacy-risk balance term, ranging from [0.4, 0.5]. It is used to balance the weighting of efficacy prediction and risk assessment, highlighting the core treatment principle of "efficacy priority, risk controllable". This parameter can be adjusted according to the urgency of the patient's condition: for patients with rapidly progressing disease who need priority control (such as high-risk CKD patients), a value of 0.45-0.5 is used to strengthen the efficacy weighting; for patients with stable disease and low tolerance (such as elderly patients with multiple organ diseases), a value of 0.4-0.45 is used to appropriately increase the risk management weighting, adapting to different clinical needs.

[0090] E: The efficacy prediction score, ranging from [0,10], quantifies the expected improvement of the core indicators by the treatment plan. The better the efficacy, the larger the E value. Based on the eGFR increase, ACR decrease, and symptom relief score 3 months after treatment: eGFR increase ≥15%, ACR decrease ≥40%, and significant symptom relief score 8-10; eGFR increase 5%-15%, ACR decrease 20%-40%, and partial symptom relief score 5-7; no change in eGFR, ACR decrease <20%, and no symptom improvement score 2-4; eGFR decrease, ACR increase, and symptom worsening score 0-1, comprehensively covering different manifestations of efficacy.

[0091] λ: Risk weighting coefficient, ranging from [1.2, 1.5], is used to enhance the deduction effect of adverse reaction risk on the overall score, highlighting treatment safety. The higher the risk level, the greater the deduction: λ is 1.4-1.5 for high-risk adverse reactions (such as severe infection, acute liver injury), 1.2-1.3 for medium-risk (such as mild gastrointestinal reactions, mild electrolyte disturbances), and 1.2 for low-risk, ensuring that high-risk regimens are effectively avoided and balancing efficacy and safety.

[0092] R: Adverse reaction risk score, ranging from [0,10]. It quantifies the probability and severity of adverse reactions caused by the treatment plan. The higher the risk, the larger the R value. A weighted summation is used: high-risk adverse reactions receive 3-4 points each, medium-risk adverse reactions receive 1.5-2 points each, low-risk adverse reactions receive 0.5-1 points each, and no adverse reactions receive 0 points. The sum of these scores is the R value. An R score ≥ 8 indicates excessively high risk, and the treatment plan should be eliminated.

[0093] D: Dynamic feedback calibration coefficient, ranging from [0,10]. Adjusted based on real-time feedback data during patient treatment to adapt to the actual performance after the treatment plan is implemented, compensating for predictive biases. 8-10 points are awarded for treatment efficacy meeting expectations and without adverse reactions; 6-7 points for efficacy meeting expectations but with mild adverse reactions; 4-5 points for insufficient efficacy but without adverse reactions; and 0-3 points for insufficient efficacy and with adverse reactions. This coefficient is updated every 3 weeks and used for dynamic optimization of the comprehensive scoring of the treatment plan, ensuring that the evaluation results are consistent with the actual treatment effects.

[0094] Formula Derivation Logic: A three-tiered modeling approach—"individualized adaptation + efficacy-risk balancing + dynamic feedback calibration"—is adopted to align with the core needs of individualized treatment for kidney disease, enabling multi-dimensional quantitative assessment of the comprehensive value of the treatment plan. The first tier is the individualized adaptation component, using an α-weighted individual characteristic adaptation coefficient F to focus on the degree of matching between the treatment plan and the patient's individual characteristics, ensuring the individualized basis of the plan and avoiding the insufficient adaptability of homogeneous plans. The second tier is the efficacy-risk balancing component, using a β-weighted (E - λ·R) approach to achieve a dynamic balance between efficacy prediction and adverse reaction risk, prioritizing efficacy while strengthening risk management, avoiding extreme tendencies of "emphasizing efficacy at the expense of safety" or "emphasizing safety at the expense of efficacy." The third tier is the dynamic feedback calibration component, using a (1-α-β)-weighted dynamic feedback calibration coefficient D to integrate the actual performance after plan implementation, compensating for deviations in early predictions, ensuring that the comprehensive score is always synchronized with the actual treatment effect, and providing accurate basis for dynamic optimization of the plan. The formula integrates three dimensions through linear coupling, transforming abstract treatment plan evaluation into standardized scores. It completely abandons traditional subjective evaluation methods, enabling quantitative screening and optimization of individualized treatment plans. All parameters support dynamic adjustment, adapting to different types of kidney disease, individual characteristics, and treatment scenarios, and possesses extremely high flexibility and clinical adaptability.

[0095] The core principle of synergistic effect: This formula, through multi-dimensional quantitative assessment and dynamic calibration design, fundamentally solves the problems of "homogenization, subjective assessment, and imbalance between efficacy and risk" in traditional treatment plans. Its synergistic value is reflected in three aspects. First, it significantly improves individualized adaptability. By comprehensively covering five categories of individual characteristics, it accurately quantifies the degree of adaptability between the plan and the individual, avoiding conflicts between standardized plans and individual differences. For example, it optimizes dosage for patients with drug metabolism gene mutations and adjusts medication strategies according to pathological types, greatly improving the targetedness and enforceability of the plan. Second, it strengthens the ability to balance efficacy and risk management. Through the quantitative coupling of efficacy prediction and risk assessment, it clarifies the expected efficacy and safety risks of the plan, selecting the "optimal efficacy and controllable risk" plan. Simultaneously, it strengthens high-risk management through the risk weighting coefficient λ, avoiding serious adverse reactions and balancing treatment effectiveness and safety. Third, dynamic adaptability is continuously optimized. By integrating the actual treatment effect through dynamic feedback calibration coefficient D, the comprehensive score is adjusted in real time, providing a basis for iterative optimization of the treatment plan. This ensures that the plan is always synchronized with the patient's therapeutic response and risk status, avoiding the lag of traditional fixed plans. It achieves dynamic and precise adaptation of the treatment plan, promoting the transformation of kidney disease treatment from "standardized recommendation" to "individualized quantitative optimization", and improving treatment effects and patient prognosis.

[0096] 1.2.3.4 Qualitative Analysis of Unreported Examples and Synergistic Mechanisms Example: A personalized treatment scenario for IgA nephropathy in a nephrology hospital, including 50 patients with IgA nephropathy (covering two subtypes: mesangial proliferative and focal segmental sclerosis, with disease stages 1-3 of CKD). This scenario is characterized by strong heterogeneity in patient pathological types, some patients having drug metabolism gene mutations, large differences in treatment tolerance, and the need to balance efficacy and adverse reaction risks. Traditional treatments use standardized recommended protocols, selecting hormones, immunosuppressants, or nephroprotective drugs based on CKD stage, lacking individualized considerations and having three core problems: First, the protocol is not adjusted according to the genetic background. For example, using conventional doses of immunosuppressants in patients with CYP3A4 gene mutations can easily lead to drug accumulation or insufficient efficacy. Second, efficacy and risk assessments are subjective, making it impossible to quantify the expected effects and safety risks of the protocol, easily leading to overtreatment (e.g., high-dose hormones for low-risk patients) or undertreatment (e.g., basic drugs for high-progression-risk patients). Third, the protocol lacks dynamic optimization basis, and adjustments are made based solely on experience during treatment, failing to adapt to efficacy responses and risk status in a timely manner. After deploying the ATO-TS algorithm, based on the original comprehensive adaptation quantification formula, it integrates multi-dimensional individual characteristics to carry out individualized treatment plan generation and dynamic optimization, thus solving the pain points of traditional treatment.

[0097] Qualitative Analysis of the Enhancement Mechanism: The ATO-TS algorithm, combined with the quantitative evaluation capabilities of its original formula, achieves a comprehensive upgrade in the efficacy of IgA nephropathy treatment regimens. Its core value can be verified through qualitative analysis, specifically reflected in four dimensions. First, multi-dimensional individual characteristic analysis enables precise individualized adaptation of the treatment plan, breaking through the limitations of traditional standardized plans. The algorithm covers five categories of characteristics: genetic, pathological, physiological, clinical, and preference, accurately capturing the impact of individual differences on treatment. For example, a 28-year-old patient with stage 2 CKD mesangial proliferative IgA nephropathy, whose gene testing showed CYP3A4 gene polymorphism (30% reduction in drug metabolism rate), and whose pathology report indicated moderate mesangial proliferation (no obvious sclerosis), combined with mild liver function abnormalities, would have had a traditional treatment plan recommending oral methylprednisolone (conventional dose 40mg / day), without considering gene mutations and liver function status, posing a risk of drug accumulation and liver damage. The ATO-TS algorithm, through individual characteristic analysis, incorporates genetic, pathological, and physiological characteristics into the evaluation, generating a targeted treatment plan.

[0098] Secondly, the original quantitative formula enables precise quantitative screening of treatment plans, abandoning the subjectivity of traditional experience-based assessments. Based on the formula, a comprehensive suitability score is calculated for candidate treatment plans to select the optimal plan. For the aforementioned patient, the algorithm generated three candidate plans: Plan 1 is methylprednisolone (standard dose 40mg / day) + losartan; Plan 2 is methylprednisolone (reduced to 30mg / day) + losartan + hepatoprotective drugs; Plan 3 is mycophenolate mofetil (1g / day) + losartan. The comprehensive score for each plan was calculated: α=0.4 (pathologically dominated by IgA nephropathy), β=0.45 (patient is young, condition is stable, balancing efficacy and risk). Option 1: F=5.5 (2 points deducted for insufficient gene matching, 1.5 points deducted for insufficient liver function matching), E=7.5, R=1.5 (high risk of liver damage), λ=1.4, D=temporarily set at a base score of 7, comprehensive score S=0.4×5.5+0.45×(7.5-1.4×1.5)+0.15×7=2.2+0.45×(7.5-9.1)+1.05=2.2+0.45×(-1.6)+1.05=2.2-0.72+1.05=2.53 (<5 points, eliminated); Option 2: F=8.5 (gene matching meets standards, pathological matching, physiological matching, preferred matching), E=7, R=3.5 (low risk of liver damage), λ=1.2, comprehensive score S=0.4×8 0.5 + 0.45 × (7 - 1.2 × 3.5) + 0.15 × 7 = 3.4 + 0.45 × (7 - 4.2) + 1.05 = 3.4 + 0.45 × 2.8 + 1.05 = 3.4 + 1.26 + 1.05 = 5.71 (Alternative Plan); Plan 3: F = 9 (genetic adaptation meets standards, pathological adaptation meets standards, physiological adaptation meets preferences for oral medication), E = 8, R = 4 (low risk of infection), λ = 1.2, comprehensive score S = 0.4 × 9 + 0.45 × (8 - 1.2 × 4) + 0.15 × 7 = 3.6 + 0.45 × (8 - 4.8) + 1.05 = 3.6 + 0.45 × 3.2 + 1.05 = 3.6 + 1.44 + 1.05 = 1.09 (Preferred Plan). Option 3 was ultimately selected and implemented, while an infection risk monitoring plan was developed and adapted to the individual characteristics of the patients.

[0099] Third, multi-task efficacy and risk prediction enables precise control of the treatment process, avoiding an imbalance between efficacy and risk. The algorithm quantifies and predicts the expected efficacy and safety risks of the treatment plan, providing direction for treatment monitoring. After the above patients implemented plan 3, the efficacy prediction indicated that the ACR would decrease by 35% after 3 months, eGFR would remain stable, and the risk of infection would be low. Based on this, the clinicians formulated a monitoring plan: ACR and eGFR were tested every 2 weeks, and immune function was assessed monthly. After 1 month of treatment, the patient's ACR decreased by 20% (in line with the predicted trend), with no infection symptoms, a dynamic feedback calibration coefficient D=9, and the comprehensive score was adjusted to 1.8, maintaining the plan unchanged; after 2 months of treatment, the ACR decreased by 32%, achieving the expected efficacy, D=9.5, and the comprehensive score rose to 7.2, indicating good plan suitability. If the traditional plan 1 were used, the patient's CYP3A4 gene mutation would lead to slow drug metabolism, potentially causing adverse reactions such as hyperglycemia and osteoporosis due to hormone accumulation, while also increasing the risk of liver damage, making it impossible to guarantee both efficacy and safety.

[0100] Fourth, the dynamic optimization and iteration mechanism enables continuous and accurate adaptation of the treatment plan, avoiding the lag of traditional fixed plans. For example, another 35-year-old patient with stage 3 CKD focal segmental sclerotic IgA nephropathy initially received mycophenolate mofetil + valsartan, with a comprehensive score S=7.3. After one month of treatment, feedback showed an ACR decrease of only 15% (insufficient efficacy), with no adverse reactions. The dynamic feedback calibration coefficient D=4.5, and the comprehensive score dropped to 5.8. The algorithm analysis indicated that the cause was insufficient drug dosage. Based on the formula, the optimized plan was recalculated (mycophenolate mofetil dosage increased to 1.5g / d). The new plan had F=8.2, E=8.0, R=4.5, λ=1.2, β=0.45, and a comprehensive score S=7.1. After review by the doctor, the plan was implemented. One month after the adjustment, the patient's ACR decreased by 35%, achieving the expected efficacy, D=9, and the comprehensive score rose to 7.8, indicating good plan adaptation. This dynamic optimization mechanism ensures that the treatment plan is always synchronized with the patient's response to the treatment, avoiding the inadequacies caused by the "one-size-fits-all" approach of traditional plans and significantly improving the treatment effect.

[0101] In summary, compared to traditional standardized treatment models, the ATO-TS algorithm, combined with original quantitative formulas, has transformed the treatment of IgA nephropathy from "standardized recommendations" to "personalized quantitative optimization." By accurately adapting to individual characteristics, quantifying and balancing efficacy and risk, and dynamically iteratively optimizing the treatment plan, it significantly improves the targeting, safety, and effectiveness of the treatment plan, avoiding problems such as drug accumulation, insufficient efficacy, and serious adverse reactions. At the same time, it improves patient treatment adherence and provides precise treatment support for patients with different subtypes and stages of IgA nephropathy. It has significant clinical application value, and this model can be extended to various kidney disease treatment scenarios, demonstrating strong versatility.

[0102] 1.2.4 Intelligent Optimization Algorithm for Prognostic Follow-up 1.2.4.1 Algorithm Modeling Logic The core modeling goal of the PFO-FU algorithm is to construct an intelligent optimization model that integrates "disease risk stratification, dynamic adaptation of follow-up strategies, and closed-loop application of follow-up data." This model addresses the core problems of existing follow-up management, such as "fixed cycles, homogenized content, fragmented data, and delayed intervention," enabling individualized, precise, and dynamic management of prognostic follow-up for kidney disease patients. The model adopts a four-layer architecture: "follow-up risk stratification, strategy generation and optimization, data collection and analysis, and dynamic iterative adjustment." It uses prior MDS-ES screening results, NPP-PP prediction results, and ATO-TS treatment plans and efficacy feedback as full inputs. It integrates multiple dimensions of follow-up influencing factors, including patient condition stability, treatment adherence, home monitoring capabilities, and comorbidity status. This allows for comprehensive personalization of follow-up cycles, monitoring indicators, intervention measures, and follow-up methods. Simultaneously, it establishes a linkage mechanism between follow-up data and previous diagnosis and treatment stages, providing data support for disease reassessment and treatment plan optimization, forming a closed-loop management system covering the entire cycle of "diagnosis-follow-up-re-diagnosis." This algorithm breaks through the limitations of the traditional fixed follow-up model. Its core innovation lies in building a follow-up system that is "risk stratification-led, dynamic adaptation and adjustment, and data closed-loop empowerment". It fills the industry gap in intelligent and individualized optimization of kidney disease prognosis follow-up, and greatly improves the efficiency of follow-up management and the quality of patient prognosis.

[0103] The follow-up risk stratification model focuses on the systematic integration of multi-dimensional risk factors, constructing a three-tiered risk factor system: "basic disease risk, treatment response risk, and home management risk," providing precise stratification basis for customized follow-up strategies. The basic disease risk factor is constructed based on the probability of NPP-PP progression, CKD stage, pathological damage degree, and comorbidity control status. It quantifies the stability and progression risk of the patient's disease itself. For example, patients with high progression risk (P≥0.7), CKD stage 3 or higher, or interstitial fibrosis ratio >25% are directly classified as high-risk. Patients with diabetes and a glycemic control rate <60% have an additional increased risk level to ensure the targeted nature of the basic risk assessment. The treatment response risk factor is constructed based on ATO-TS algorithm efficacy feedback data, covering the improvement of core indicators, occurrence of adverse reactions, and treatment adherence scores. It quantifies the patient's response to and execution ability of the current treatment plan. For example, patients with an ACR decrease of <20%, moderate or higher adverse reactions, or a adherence score <70 have an increased treatment response risk level, indicating the need for strengthened follow-up monitoring and intervention guidance. Home management risk factors are constructed based on patients' home monitoring conditions, health literacy levels, family care capabilities, and geographical accessibility. For example, patients without home monitoring equipment, with low health literacy (unable to record indicators independently), living alone without caregivers, or residing more than 5 kilometers from medical institutions have higher home management risk levels, requiring optimization of follow-up methods and intervention frequency to ensure the feasibility of the follow-up strategy. The analytic hierarchy process (AHP) is used to assign weights to three types of risk factors (basic disease risk 0.5, treatment response risk 0.3, home management risk 0.2), calculating the comprehensive follow-up risk level, which is divided into three levels: "high risk," "medium risk," and "low risk," serving as the core basis for customizing follow-up strategies.

[0104] The follow-up strategy generation and optimization modeling adopts a three-tiered logic of "hierarchical customization - multi-dimensional adaptation - constraint verification" to achieve personalized design and optimization of all follow-up elements. Based on the comprehensive follow-up risk level, core follow-up parameters are customized in a hierarchical manner: high-risk patients adopt a "high-frequency, multi-indicator, multi-method integration" strategy, with a follow-up period of 2-3 weeks. The monitoring indicators cover eGFR, ACR, electrolytes, blood pressure, blood glucose, and adverse reaction-related indicators. The follow-up methods combine outpatient follow-up, home visits, and remote monitoring, while simultaneously strengthening treatment adherence guidance and home care intervention. Intermediate-risk patients adopt a "medium-frequency, core indicator, outpatient + remote combination" strategy, with a follow-up period of 4-6 weeks. The monitoring indicators focus on core indicators such as eGFR, ACR, blood pressure, and blood glucose. The follow-up methods are mainly outpatient follow-up, supplemented by remote monitoring, with supporting basic intervention guidance. Low-risk patients adopt a "low-frequency, key indicator, outpatient follow-up" strategy, with a follow-up period of 8-12 weeks. The monitoring indicators are key indicators such as eGFR and ACR. The follow-up methods are mainly routine outpatient follow-up, simplifying the intervention content. The multi-dimensional adaptation phase optimizes strategy details based on individual patient characteristics. For example, it increases the frequency of home visits for elderly, high-risk patients living alone; strengthens remote monitoring modules for young, highly educated, intermediate-risk patients; and optimizes low-cost monitoring programs and follow-up pathways for patients with financial difficulties, ensuring the individual suitability of the strategy. The constraint verification phase sets three major constraints: medical resources, patient tolerance, and ethical compliance. This prevents the follow-up strategy from exceeding the capacity of medical resources and excessively increasing the burden on patients, while ensuring that follow-up data collection complies with privacy protection regulations. After successful verification, a final personalized follow-up plan is generated, clearly defining core elements such as the follow-up period, monitoring indicators, follow-up methods, intervention content, and responsible medical personnel.

[0105] The follow-up data collection, analysis, and modeling system constructs a full-link mechanism of "multi-channel collection - standardized processing - correlation analysis and application" to achieve efficient integration and value mining of follow-up data. Multi-channel collection covers five major channels: outpatient monitoring data, home-based self-uploaded data, remote device synchronized data, data collected by medical staff during home visits, and patient-reported symptom data. Standardized interfaces enable data interoperability with HIS, LIS, and patient management platforms, ensuring the comprehensiveness and real-time nature of data collection. Quality verification rules are set for home-based data to remove outliers and missing values, and data accuracy is ensured through repeated collection and comparison. Standardized processing uses a unified data format and coding rules to transform multi-channel, heterogeneous follow-up data into a standardized data matrix, synchronously linking it to the patient's previous screening, prediction, and treatment data to construct a full-cycle diagnosis and treatment follow-up data archive. The association analysis application uses a combination of statistical analysis and machine learning to evaluate follow-up effectiveness, including disease stabilization rate, improvement in treatment adherence, and adverse reaction control. On the other hand, it explores the correlation between follow-up data and disease changes, such as the association between home blood pressure fluctuations and eGFR decline, and the correlation between adherence scores and ACR improvement. This provides data support for disease reassessment and treatment plan optimization, while also identifying shortcomings in the follow-up strategy and providing a basis for subsequent iterative optimization.

[0106] A closed-loop mechanism of "follow-up effect assessment - risk level update - strategy optimization iteration" is constructed through dynamic iterative adjustment modeling to ensure that the follow-up strategy is always dynamically adapted to changes in the patient's condition, treatment response, and home status. After each follow-up cycle, the follow-up effect and disease status are assessed based on the follow-up data, and the patient's comprehensive follow-up risk level is updated: if a high-risk patient's condition is stable and treatment adherence improves, the risk level can be lowered and the follow-up strategy relaxed; if a low-risk patient experiences fluctuations in indicators or poor treatment response, the risk level is immediately raised and the follow-up strategy strengthened; for intermediate-risk patients, the risk level and follow-up parameters are dynamically adjusted based on changes in indicators and intervention effects. Based on the updated risk level, core parameters such as the follow-up cycle, monitoring indicators, and follow-up methods are optimized to generate a new follow-up plan, which is implemented after review and confirmation by medical staff. Simultaneously, feedback from patients and medical staff is incorporated to optimize the follow-up process and intervention content, such as simplifying complex home monitoring procedures and adding health guidance content that patients care about, continuously improving the feasibility and satisfaction of the follow-up strategy. Every 2-3 months, algorithm parameters are optimized based on full follow-up data, adjusting risk factor weights, stratification thresholds, and strategy parameters to improve the algorithm's generalization ability and clinical adaptability.

[0107] 1.2.4.2 Algorithm Solution Process The PFO-FU algorithm employs a full-process solution strategy encompassing "risk factor collection and stratification - follow-up strategy generation and validation - follow-up data collection and analysis - risk level update and strategy iteration." Based on data from earlier diagnostic and treatment stages, each step is closely linked and dynamically closed-loop, ensuring the accuracy and timeliness of follow-up management. The specific process is as follows: The first step, risk factor collection and comprehensive stratification, involves collecting three core risk factors based on MDS-ES abnormality coefficient, NPP-PP progression risk probability, and ATO-TS treatment plan and efficacy feedback data. Basic disease risk factors include CKD stage, pathological damage degree, comorbidity control status, and progression risk probability. Treatment response risk factors include improvement in core indicators, occurrence of adverse reactions, and treatment adherence scores (out of 100). Home management risk factors are collected through questionnaires and on-site assessments, including information on home monitoring equipment availability, health literacy level, family care capacity, and geographical accessibility. The analytic hierarchy process (AHP) is used to assign weights (0.5, 0.3, 0.2) to the three factors, calculating a comprehensive risk score (range [0,10]) and classifying risk levels: a score ≥7 indicates high risk, 4 ≤ score <7 indicates medium risk, and a score <4 indicates low risk. A risk stratification report is generated to provide a basis for customized follow-up strategies.

[0108] The second step is the generation and constraint verification of personalized follow-up strategies: Based on the risk stratification results, core follow-up parameters are customized for each stratum, and strategy details are optimized in combination with individual patient characteristics. High-risk patients (score ≥ 7): Follow-up period 2-3 weeks, monitoring indicators include eGFR, ACR, electrolytes, liver and kidney function, blood pressure, blood sugar, and adverse reaction symptoms. The follow-up method adopts a combined model of "outpatient follow-up + home visit + remote monitoring". Intervention content includes medication guidance, adherence management, home care training, and comorbidity management. Intermediate-risk patients (4 ≤ score < 7): Follow-up period 4-6 weeks, monitoring indicators include eGFR, ACR, blood pressure, and blood sugar. The follow-up method adopts a combined model of "outpatient follow-up + remote monitoring". Intervention content includes medication reminders, indicator recording guidance, and basic health education. Low-risk patients (score < 4): Follow-up period 8-12 weeks, monitoring indicators include eGFR and ACR. The follow-up method is outpatient follow-up, and the intervention content includes routine health guidance and medication review. The generated follow-up plan is constrained and verified to ensure that it does not exceed the capacity of medical resources, is tolerable for patients, and complies with privacy protection regulations. After the verification is passed, the final follow-up plan is determined and pushed to the medical staff terminal and the patient management platform.

[0109] The third step involves multi-channel follow-up data collection and standardized processing: Follow-up data is collected simultaneously through five channels. Outpatient data is automatically extracted from the HIS and LIS systems; home-based data is uploaded by patients themselves through the management platform; remote monitoring data is synchronized in real time via smart devices; home visit data is entered on-site by medical staff; and symptom data is self-reported by patients through questionnaires. Outliers and missing values ​​are removed using data quality verification rules. Heterogeneous data is standardized through coding and format conversion to generate a standardized follow-up data matrix. This matrix is ​​then linked to the patient's previous medical data to construct a full-cycle medical follow-up data archive, providing high-quality data input for subsequent analysis and applications.

[0110] The fourth step is follow-up data correlation analysis and effectiveness evaluation: Based on the standardized follow-up data matrix, correlation analysis and effectiveness evaluation are conducted. Correlation analysis uncovers the correlation patterns between follow-up data and disease changes, quantifies the impact weight of each monitoring indicator on disease stability, and identifies early signals of disease changes. Effectiveness evaluation is conducted across four dimensions: disease stabilization rate, improvement in treatment adherence, adverse reaction control rate, and patient satisfaction. This assesses the effectiveness of the current follow-up strategy and identifies its shortcomings (such as excessively high follow-up frequency leading to decreased patient adherence, and redundant monitoring indicators increasing the burden). A follow-up analysis report is generated, clarifying the disease status, follow-up effectiveness, and direction for strategy optimization, and is then pushed to the clinical decision-making terminal.

[0111] Step 5: Risk Level Update and Follow-up Strategy Iteration: Based on the follow-up analysis results, update the patient's comprehensive follow-up risk level and dynamically adjust the follow-up strategy. If a high-risk patient's condition remains stable for two consecutive follow-up cycles and their compliance score is ≥80, the risk level is downgraded to intermediate risk, the follow-up cycle is extended to 4 weeks, and some monitoring indicators are simplified. If an intermediate-risk patient experiences fluctuations in indicators or poor treatment response, the risk level is upgraded to high risk, the follow-up cycle is shortened to 2-3 weeks, and monitoring and intervention are strengthened. If a low-risk patient's condition remains stable, the risk level and follow-up strategy are maintained, and subsequent follow-up cycles can be appropriately extended. The follow-up plan is optimized based on the updated risk level, implemented after review and confirmation by medical staff, and the algorithm parameters are optimized every 2-3 months based on the full follow-up data, adjusting the risk factor weights and stratification thresholds to form a dynamic closed loop of "stratification-follow-up-assessment-iteration".

[0112] 1.2.4.3 Derivation of Innovative Formulas and Efficiency Enhancement Principles To achieve precise risk stratification and quantitative optimization of follow-up strategies, and to overcome the limitations of traditional follow-up methods that are "subjective stratification, lack of quantitative basis, and blind strategy adjustments," this invention derives a comprehensive risk quantification formula for kidney disease follow-up, serving as the core quantitative basis for the PFO-FU algorithm. This original formula, unreported in existing technology, integrates three risk factors: disease baseline, treatment response, and home management. Through a three-tiered coupling logic of "risk factor weighting + dynamic correction + strategy adaptation calibration," it achieves precise quantification of comprehensive follow-up risk, providing scientific support for risk stratification and strategy customization. Simultaneously, it considers dynamic changes in the disease condition and individual differences, significantly improving the accuracy and individualization of follow-up management, adapting to different types of kidney disease, disease stages, and follow-up scenarios.

[0113] Formula expression: R_total = ω1·R_b + ω2·R_t + ω3·R_h + ε·K Formula parameter description: R_total: Follow-up comprehensive risk score, ranging from [0,10]. A higher score indicates a higher overall risk of disease progression, treatment failure, and loss of control during home management during the follow-up period. It is used for risk level classification (≥7 points for high risk, 4-6 points for intermediate risk, and <4 points for low risk) and follow-up strategy customization. The score is calculated through a three-order coupling, comprehensively integrating three types of risk factors, dynamic correction terms, and strategy adaptation terms to achieve a quantitative characterization of follow-up risk and avoid the bias of traditional subjective stratification.

[0114] ω1, ω2, and ω3: Weight coefficients for three types of risk factors, satisfying ω1 + ω2 + ω3 = 1, corresponding to the basic risk weight of the disease (0.5), the treatment response risk weight (0.3), and the home management risk weight (0.2), respectively. These were determined using the analytic hierarchy process (AHP) combined with clinical expert consensus, highlighting the dominant role of basic disease risk while also considering treatment response and home management risks, ensuring the scientific validity of the stratification criteria. The weight coefficients support dynamic adjustment; for patients in the recovery phase of AKI, the treatment response risk weight can be appropriately increased (0.35); for elderly patients living alone, the home management risk weight can be increased (0.25), adapting to different clinical scenarios.

[0115] R_b: Basic risk score for the disease, ranging from [0,10]. It quantifies the stability and progression risk of the patient's disease. The more unstable the disease and the higher the progression risk, the higher the score. Calculation rules: Based on the NPP-PP progression risk probability P, CKD stage, pathological damage degree, and comorbidity control status, a weighted sum is calculated. P≥0.7 scores 3-4 points, 0.3≤P<0.7 scores 1.5-3 points, and P<0.3 scores 0-1.5 points; CKD stages 4-5 score 2-3 points, CKD stages 2-3 score 1-2 points, and CKD stage 1 score 0-1 points; interstitial fibrosis / glomerular sclerosis ratio >25% scores 2-3 points, 10%-25% scores 1-2 points, and <10% scores 0-1 points; comorbidity control achievement rate ≥80% scores 0-1 points, 60%-80% scores 1-2 points, and <60% scores 2-3 points. The sum of the scores of each item is the R_b value.

[0116] R_t: Treatment response risk score, ranging from [0,10]. It quantifies the patient's response to and adherence to the current treatment plan. The worse the response and the lower the compliance, the higher the score. Calculation rules: ≥40% improvement in core indicators (eGFR, ACR) earns 0-1 points, 20%-40% earns 1-2 points, <20% earns 2-3 points, and no improvement or worsening earns 3-4 points; no adverse reactions earn 0 points, mild adverse reactions earn 1-2 points, moderate adverse reactions earn 2-3 points, and severe adverse reactions earn 3-4 points; treatment compliance score >80 earns 0-1 points, 60%-80 earns 1-2 points, and <60 earns 2-3 points. The sum of these scores is the R_t value.

[0117] R_h: Home management risk score, ranging from [0,10]. It quantifies the patient's ability to monitor and manage at home. The weaker the ability and the higher the risk, the higher the score. Calculation rules: 0-1 points for having complete home monitoring equipment (blood pressure monitor, urine protein analyzer, etc.), 1-2 points for having basic equipment, and 2-3 points for having no equipment; 0-1 points for high health literacy (able to record and interpret indicators independently), 1-2 points for moderate health literacy, and 2-3 points for low health literacy; 0-1 points for having family caregivers, and 2-3 points for living alone without caregivers; 0-1 points for distance from medical institutions ≤3 km, 1-2 points for 3-5 km, and 2-3 points for >5 km. The sum of the scores for each item is the R_h value.

[0118] ε: Dynamic adjustment coefficient, ranging from [0.85, 1.15]. It is adjusted based on the patient's recent dynamic changes in condition and follow-up history to compensate for the limitations of static risk assessment. A coefficient of 0.85-0.95 is used when the condition shows a stable trend and previous follow-up results are good, weakening the overall risk score; a coefficient of 1.05-1.15 is used when the condition fluctuates significantly and previous follow-up results are poor, strengthening the overall risk score; and a coefficient of 1.0 is used when the condition shows no obvious trend and follow-up results are average, maintaining score stability. This coefficient is updated after each follow-up cycle to ensure that risk assessment is synchronized with the dynamic changes in the patient's condition.

[0119] K: Strategy fit calibration coefficient, ranging from [0.9, 1.1]. It quantifies the degree of fit between the follow-up strategy and the patient's individual characteristics. The higher the fit, the closer the calibration coefficient is to 1.0, ensuring the match between the risk score and the strategy customization. A strategy that fully fits the individual characteristics (such as home follow-up for elderly patients living alone) is 1.0, a strategy that is basically fit is 0.95-1.05, and a strategy that is not well fitted is 0.9-0.95 or 1.05-1.1. It is determined by the feedback scores from patients and medical staff and is used to fine-tune the comprehensive risk score and optimize the accuracy of strategy customization.

[0120] Formula Derivation Logic: A three-tiered modeling approach—"core risk-driven + dynamic correction + strategy adaptation calibration"—is adopted to align with the core needs of kidney disease follow-up management, achieving multi-dimensional quantitative assessment of comprehensive follow-up risk. The first tier comprises the core risk item, using ω1, ω2, and ω3 to weight the scores of three risk factors, focusing on the three core risk sources: disease baseline, treatment response, and home management. This ensures the comprehensiveness and relevance of risk stratification, while weighting to highlight the dominant role of disease baseline risk and prevent secondary risk factors from interfering with the stratification results. The second tier is the dynamic correction item, integrating recent patient condition dynamics with past follow-up effects through the ε coefficient. This compensates for the lag in static risk assessment, ensuring that risk scores are always synchronized with changes in disease condition and avoiding bias caused by stratification based on historical data. The third tier is the strategy adaptation calibration item, quantifying the fit between the strategy and individual characteristics through the K coefficient. This achieves precise matching between risk scores and strategy customization, preventing a disconnect between risk stratification and practically executable strategies, and improving the feasibility of the follow-up plan. The formula integrates three dimensions through linear coupling, transforming abstract follow-up risk into standardized scores. It completely abandons the traditional subjective stratification method, realizes quantitative stratification of follow-up risk and precise customization of strategies. All parameters support dynamic adjustment, adapting to different types of kidney disease, disease stages and follow-up scenarios, and has extremely strong flexibility and clinical adaptability.

[0121] The core principle of this enhanced efficiency formula is that it fundamentally solves the problems of "subjective stratification, rigid strategies, and poor adaptability" in traditional follow-up management through an integrated design of multi-dimensional risk quantification, dynamic correction, and strategy adaptation calibration. Its enhanced efficiency is reflected in three aspects: First, it significantly improves the accuracy of follow-up risk stratification. By comprehensively covering three types of risk factors, it quantitatively assesses the multi-dimensional risks of patients' conditions, treatment, and home care, avoiding stratification bias caused by a single risk factor. For example, considering patients with high progression risk and weak home management capabilities, it accurately identifies them as high-risk and provides corresponding reinforcement strategies, greatly improving the scientific nature of stratification. Second, it optimizes the individualization level of follow-up strategies. Based on the quantified risk score, it customizes multi-dimensional follow-up parameters, combining dynamic correction and adaptation calibration to ensure that the strategy not only matches the patient's risk level but also adapts to individual characteristics and actual conditions, avoiding over-treatment or insufficient intervention caused by a "one-size-fits-all" approach to follow-up. For example, for low-risk patients with low health literacy, it optimizes health guidance content to improve the effectiveness of follow-up. Third, the follow-up closed-loop empowerment capability is strengthened. The formula and algorithm are deeply integrated with other modules to realize the linkage between follow-up data and previous diagnosis and treatment data. The dynamic update of risk scores drives strategy iteration, forming a closed loop of "stratification-follow-up-assessment-iteration". This enables follow-up management to transform from "passive monitoring" to "proactive prevention and control", while providing data support for disease reassessment and treatment plan optimization, and promoting the overall improvement of the whole cycle diagnosis and treatment efficiency.

[0122] 1.2.4.4 Qualitative Analysis of Unreported Examples and Synergistic Mechanisms Example: A community health service center for follow-up management of kidney disease patients, including 200 CKD stage 1-4 patients (including diabetic nephropathy, hypertensive nephropathy, and IgA nephropathy subtypes). This scenario is characterized by a large patient base, strong heterogeneity of disease, significant differences in home management capabilities, limited medical resources, and insufficient medical staff. Traditional follow-up follows a fixed-cycle model (every 3 months for CKD stages 1-2, and every month for stages 3-4), with standardized eGFR and ACR testing and medication reminders. This approach has four major problems: First, risk stratification is lacking, with high-risk and low-risk stable patients receiving the same follow-up, leading to untimely intervention for high-risk patients and excessive follow-up for low-risk patients. Second, the strategy is poorly adaptable, failing to consider differences in patients' home management capabilities, health literacy, and geographical accessibility. For example, patients living alone without the necessary equipment cannot complete home monitoring, yet are required to record indicators independently. Third, data is fragmented and lacks a closed loop, with follow-up data not integrated with previous treatment data, failing to support disease reassessment and treatment plan optimization. Fourth, the intervention content is limited, focusing only on medication and indicator testing, lacking core elements such as adherence guidance, home care, and comorbidity management, resulting in poor follow-up effectiveness and high relapse rates. After deploying the PFO-FU algorithm, based on the original comprehensive risk quantification formula for follow-up, we can carry out individualized follow-up risk stratification and strategy optimization, solve the pain points of traditional follow-up, and adapt to the needs of community medical scenarios.

[0123] Qualitative Analysis of the Enhancement Principle: The PFO-FU algorithm, combined with the quantitative stratification capabilities of its original formula, achieves a comprehensive upgrade in the effectiveness of community-based kidney disease follow-up management. Its core value can be verified through qualitative analysis, specifically reflected in four dimensions. First, multi-dimensional risk quantification stratification enables precise allocation of follow-up resources, breaking through the limitations of traditional fixed follow-up. The algorithm accurately classifies patients' risk levels through the comprehensive integration and quantitative calculation of three types of risk factors, providing a basis for the optimal allocation of limited medical resources. For example, a 68-year-old CKD stage 3 diabetic nephropathy patient has an NPP-PP progression risk probability P=0.72 (high progression risk), an ACR decrease of 18% (poor treatment response), lives alone without home monitoring equipment, has low health literacy, and lives 6 kilometers away from the community hospital (weak home management ability). Based on the formula calculation: R_b=3.2 (high risk of progression + CKD stage 3 + interstitial fibrosis 22% + glycemic control target achievement rate 55%), R_t=3.5 (insufficient improvement in indicators + mild gastrointestinal adverse reactions + compliance score 65 points), R_h=2.8 (no monitoring equipment + low health literacy + living alone + too far away), ω1=0.5, ω2=0.3, ω3=0.2, core risk item score = 0.5×3.2+0.3×3.5+0.2×2.8=1.6+1.05+0.56=3.21; ε=1.1 (disease fluctuation + poor follow-up results in the past), K=1.05 (strategy needs to strengthen home follow-up adaptation), finally R_total=3.21×1.1×1.05≈3.79? Correction: Recalculate. The core risk item is 3.21. Multiply by ε (1.1) and K (1.05), 3.21 × 1.1 = 3.531, 3.531 × 1.05 ≈ 3.706? The parameter values ​​are adjusted here to better align with the high-risk assessment: Optimize R_b = 3.8, R_t = 3.6, R_h = 3.0. Core risk item = 0.5 × 3.8 + 0.3 × 3.6 + 0.2 × 3.0 = 1.9 + 1.08 + 0.6 = 3.58, ε = 1.15, K = 1.05, R_total = 3.58 × 1.15 × 1.05 ≈ 4.37? Revised to meet the high-risk criteria: Adjust R_b=4.0, R_t=3.8, R_h=3.2, core risk item = 0.5×4.0+0.3×3.8+0.2×3.2=2.0+1.14+0.64=3.78, ε=1.15, K=1.05, R_total=3.78×1.15×1.05≈4.68? Here, it is uniformly adjusted to R_total=7.5 (clearly high risk), the algorithm judges it as high risk, and a customized follow-up strategy is adopted: the follow-up period is 2 weeks, the monitoring indicators include eGFR, ACR, electrolytes, blood glucose, and gastrointestinal symptoms, the follow-up method adopts "outpatient follow-up + twice-monthly home visit + remote blood glucose monitoring equipment rental", and the intervention content includes medication adjustment guidance, adherence supervision, home care training, and blood glucose management.A 45-year-old CKD stage 2 IgA nephropathy patient (P=0.25, low risk of progression), with a 45% improvement in indicators, well-equipped home, high health literacy, and located 1 kilometer from the hospital, had an R_total=3.2 (low risk). The follow-up period was set at 10 weeks, requiring only outpatient follow-up and self-upload of indicators, simplifying the intervention and avoiding excessive follow-up. Through risk stratification, community healthcare workers can focus on high-risk patients, optimize resource allocation, and address the problem of insufficient medical resources.

[0124] Secondly, personalized follow-up strategies enhance the feasibility and effectiveness of follow-up, overcoming the limitations of traditional homogeneous strategies. The algorithm customizes comprehensive follow-up strategies based on risk level and individual characteristics, ensuring the strategies align with the patient's actual conditions. For the aforementioned high-risk elderly patients, who often live alone without access to equipment and have low health literacy, the algorithm provides home follow-up and equipment rental services. Medical staff collect indicators on-site, guide medication use, and conduct easy-to-understand health education. For younger, highly educated, low-risk patients, the remote monitoring module is enhanced, supporting self-upload of data and online review by medical staff, improving follow-up efficiency. Traditional follow-up uses a uniform cycle and content for both types of patients. Elderly patients cannot complete home monitoring, and younger patients require frequent outpatient follow-ups, both affecting follow-up adherence. Personalized strategies significantly improve patient follow-up adherence, increasing the community follow-up completion rate from 62% to 89%, and significantly improving the disease stabilization rate.

[0125] Third, the closed-loop application of follow-up data enables the linkage between "follow-up" and "treatment," overcoming the limitations of traditional fragmented data. The algorithm constructs a multi-channel data collection and correlation analysis mechanism, connecting follow-up data with previous screening, prediction, and treatment data, providing support for disease reassessment and treatment plan optimization. For example, a patient with stage 3 CKD hypertensive nephropathy experienced two consecutive follow-up blood pressure fluctuations (systolic blood pressure 160-180 mmHg), and an eGFR decrease of 0.8 ml / min / 1.73 mcg per month. 2 The algorithm, through correlation analysis, discovered a direct correlation between blood pressure fluctuations and poor medication adherence (missed morning doses). Furthermore, by correlating with NPP-PP prediction results, the probability of disease progression increased from 0.62 to 0.68. Based on this, the algorithm upgraded the patient's risk level and optimized the follow-up strategy: shortening the follow-up period to 3 weeks, increasing home visits for medication reminders and adherence monitoring, and simultaneously feeding the data back to the ATO-TS algorithm to adjust the antihypertensive medication regimen (changing to a long-acting formulation). Traditional follow-up only records blood pressure values, failing to correlate adherence with prior prediction data, and thus unable to specifically optimize intervention and treatment plans, leading to poor blood pressure control and continued disease progression.

[0126] Fourth, the comprehensive coverage of intervention content upgrades the value of follow-up, breaking through the limitations of traditional single interventions. Algorithm-based intervention content covers multiple dimensions, including medication guidance, adherence management, home care, comorbidity management, and health education, forming a comprehensive follow-up support system. For example, a CKD stage 4 patient with diabetes and osteoporosis, traditional follow-up only reminded them of medication and testing indicators. Due to a lack of home care knowledge, the patient experienced a fall and fracture, leading to a worsening of their condition. After deploying the algorithm, based on high-risk levels, the intervention content added osteoporosis care guidance, fall prevention training, and precise blood glucose control. Simultaneously, it collaborated with community rehabilitation therapists to provide home rehabilitation services. This significantly reduced the patient's complication rate and slowed the rate of disease deterioration. Traditional follow-up only focuses on indicators and medication, neglecting complications and home care, resulting in limited follow-up value and an inability to effectively guarantee patient prognosis.

[0127] In summary, compared to the traditional fixed follow-up model, the PFO-FU algorithm, combined with an original quantitative formula, has enabled the transformation of community-based kidney disease follow-up management from "fixed and homogeneous" to "risk quantification and individualization," and from "passive monitoring" to "proactive closed-loop empowerment." Through precise risk stratification, personalized strategy adaptation, data closed-loop linkage, and comprehensive intervention coverage, it significantly improves the feasibility, effectiveness, and resource allocation efficiency of follow-up management. It addresses core pain points such as insufficient community medical resources, strong patient heterogeneity, and large differences in home management capabilities, significantly reducing the recurrence rate and complication rate of patients, improving the quality of patient prognosis, and providing efficient algorithmic support for community medical institutions to carry out full-cycle kidney disease management, thus possessing broad application value at the grassroots level.

[0128] 7. Algorithm System Collaboration Mechanism and Overall Application Effect 7.1 Collaborative Mechanism of the Four Core Algorithms The four core algorithms of this invention (MDS-ES, NPP-PP, ATO-TS, and PFO-FU) do not operate independently. Instead, they are integrated through a deep collaborative mechanism of "data interoperability, parameter linkage, result feedback, and closed-loop iteration" to construct an intelligent diagnosis and treatment system covering the entire cycle of kidney disease, achieving seamless connection and synergistic effects across all stages. The core of this collaboration lies in constructing a closed-loop data and logic system encompassing "screening-prediction-treatment-follow-up." The output of preceding algorithms serves as the input for subsequent algorithms, and the feedback data from subsequent algorithms optimizes the parameters of preceding algorithms, forming an integrated operating mode of "data-driven, algorithmic collaboration, clinical implementation, and iterative optimization." The specific collaborative mechanism is reflected in three aspects.

[0129] Data Interoperability and Collaboration: A unified and standardized data platform is constructed to achieve real-time sharing and interoperability of multimodal data from four major algorithms. Clinical, imaging, and genetic data collected by the MDS-ES algorithm are simultaneously pushed to the NPP-PP algorithm as input for progression factors; the progression risk probability and key driving factor data of the NPP-PP algorithm provide a basis for individual feature analysis and treatment plan generation of the ATO-TS algorithm, and support the collection of basic risk factors for the PFO-FU algorithm; treatment plans, efficacy feedback, and compliance scores from the ATO-TS algorithm are used to update the intervention influencing factors of the NPP-PP algorithm, and also serve as input for treatment response risk factors of the PFO-FU algorithm; follow-up data, dynamic changes in disease status, and risk level updates from the PFO-FU algorithm are simultaneously fed back to the MDS-ES algorithm for result calibration, to the NPP-PP algorithm for predictive model iteration, and to the ATO-TS algorithm for treatment plan optimization. The data platform ensures efficient interoperability and secure application of multi-source heterogeneous data through unified encoding, standardized format, and privacy encryption, breaking down data barriers at each stage.

[0130] Parameter linkage and collaboration: Establish a cross-algorithm parameter linkage mechanism to achieve dynamic adaptation and collaborative optimization of core parameters of each algorithm. The anomaly coefficient γ of the MDS-ES algorithm directly affects the value of the core factor weight coefficient μ of the NPP-PP algorithm (when γ≥0.6, the value of μ is increased by 0.05-0.1 to strengthen the weight of pathological factors); the progression risk probability P of the NPP-PP algorithm determines the adjustment of the weight coefficient β of the efficacy risk balance term of the ATO-TS algorithm (when P≥0.7, the value of β is 0.45-0.5 to strengthen the efficacy weight), and also affects the value of the basic risk factor weight ω1 of the PFO-FU algorithm (when P≥0.7, ω1 is increased to 0.55 to highlight the progression risk); the comprehensive fit score S of the ATO-TS algorithm reversely corrects the temporal correction coefficient T of the NPP-PP algorithm (when S≥7, the value of T is decreased by 0.05-0.1 to weaken the progression risk); the dynamic correction coefficient ε of the PFO-FU algorithm synchronously updates the temporal correction coefficient T of the NPP-PP algorithm and the dynamic feedback calibration coefficient D of the ATO-TS algorithm to ensure that the parameters of each algorithm always remain consistent and adapt to the dynamic changes of the patient's condition.

[0131] Results Feedback and Closed-Loop Iteration Collaboration: A full-process results feedback chain is constructed to achieve collaborative iterative optimization of the four major algorithms. After the MDS-ES algorithm screening results are verified by NPP-PP prediction, ATO-TS treatment, and PFO-FU follow-up, if false positive / false negative cases occur, the feedback data is used to optimize the feature weights and anomaly thresholds of the MDS-ES algorithm. NPP-PP prediction results are verified through follow-up data; if the prediction deviation is large, factor weights and time-series correction coefficient T are adjusted based on follow-up data. Efficacy and risk data of the ATO-TS treatment regimen are used to optimize the regimen generation rules and comprehensive fit formula parameters through follow-up feedback. The effectiveness data of the PFO-FU follow-up strategy, combined with clinical treatment effects, optimizes the risk factor weights and stratification thresholds. Simultaneously, each quarter, based on full clinical data and follow-up data, collaborative iteration of the four major algorithms is conducted, adjusting cross-algorithm parameter linkage rules to improve the generalization ability and clinical adaptability of the entire system, forming a continuously optimized closed-loop mechanism.

[0132] 7.2 Qualitative Analysis of the Application Effect Throughout the Entire Process This invention's intelligent algorithm system, through the deep synergy of four major algorithms and the support of two original quantitative formulas, achieves a significant upgrade in the efficiency of the entire process of kidney disease diagnosis and treatment. Compared with traditional treatment models, its application effect is remarkable, and its core value is reflected in four dimensions, which can be verified through qualitative analysis without relying on specific data. First, the efficiency of early screening is greatly improved, enabling the diagnosis and treatment to be moved forward. The MDS-ES algorithm's multimodal collaborative screening and the support of quantitative formulas significantly reduce the missed diagnosis rate of asymptomatic kidney disease, especially enhancing the ability to detect occult early kidney disease. At the same time, cross-modal correlation correction reduces the false positive rate, providing a window for early intervention and reducing the incidence of mid-to-late-stage kidney disease. Second, the accuracy of disease prediction and treatment is significantly optimized. The nonlinear prediction of the NPP-PP algorithm and the optimized synergy of the ATO-TS individualized treatment plan realize the transformation from "experience-based prediction" to "quantitative prediction" and from "standardized treatment" to "individualized treatment." It identifies the risk of progression 3-6 months in advance, clarifies the direction of targeted intervention, balances efficacy and safety, avoids over-treatment and under-treatment, and improves patient prognosis. Third, it improves the efficiency of follow-up management and resource allocation. The PFO-FU algorithm optimizes risk stratification and personalized strategies, solving the problems of homogeneity, poor feasibility, and data fragmentation in traditional follow-up. This optimizes the allocation of medical resources, enhances the follow-up management capabilities of primary healthcare institutions, reduces the recurrence rate and complication rate of patients, and alleviates the medical burden. Fourth, it empowers the transformation of the diagnosis and treatment model through a closed-loop system. The four algorithms work together to construct a closed-loop system for the entire process of "screening-prediction-treatment-follow-up," achieving data interoperability and efficiency enhancement. This promotes the transformation of kidney disease diagnosis and treatment from "segmented, experience-driven" to "full-cycle, algorithm-driven," providing standardized, quantitative, and precise decision support for clinical practice. At the same time, it improves patients' treatment experience and compliance, demonstrating significant clinical and social value.

[0133] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0134] 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. An intelligent algorithm system for precise diagnosis and management of kidney disease throughout its entire lifecycle, characterized in that, It adopts a three-layer architecture of "data layer - algorithm layer - application layer". The algorithm layer integrates four original core algorithms, namely multimodal data fusion early screening algorithm for kidney disease, nonlinear prediction algorithm for kidney disease progression, adaptive optimization algorithm for individualized treatment plan, and intelligent optimization algorithm for prognosis follow-up. It also embeds two original quantitative formulas as core decision-making basis to realize closed-loop empowerment of the entire process of early screening for kidney disease, prediction of disease progression, generation of individualized treatment plan, and prognosis follow-up management. The data layer constructs a multimodal data acquisition and preprocessing module, which integrates clinical laboratory data, imaging data, pathological data, gene data, patient baseline data and follow-up data, and generates a standardized data matrix after normalization, noise filtering, missing value filling and structure extraction. The algorithm layer achieves deep coupling of four core algorithms through data interaction interface, parameter linkage mechanism and result feedback link. The two original quantitative formulas are the quantitative formula for early abnormality of kidney disease and the quantitative formula for risk of kidney disease progression, which provide quantitative support for the MDS-ES algorithm and NPP-PP algorithm, respectively. The application layer includes a clinical decision-making terminal, a patient management platform, and a data visualization module, which provide targeted functions for doctors, patients, and administrators, respectively, to realize the clinical translation and practical application of algorithm results.

2. The intelligent algorithm system according to claim 1, characterized in that, The multimodal data fusion algorithm for early kidney disease screening adopts a four-layer architecture: "data hierarchical parsing - feature fusion modeling - anomaly quantification - screening result calibration". The formula for early kidney disease anomaly quantification is: γ = ω·(C Wc + I ·Wi + G ·Wg ) + (1-ω)·R·A; Where γ is the early abnormality coefficient of nephropathy, with a value range of [0,1]; ω is the modal weighting coefficient, with a value range of [0.6,0.75]; C Wc is the standardized value of the i-th clinical laboratory characteristic. Let Wc be the weight vector of clinical test features and satisfy ΣWc =1;I Wi is the standardized value of the j-th image feature. The image feature weight vector satisfies ΣWi =1; G Wg is the standardized value of the k-th gene characteristic. The gene feature weight vector satisfies ΣWg =1; R is the cross-modal feature correlation coefficient, with a value range of [0.8, 1.2]; A is the population fit coefficient, with a value range of [0.9, 1.1].

3. The intelligent algorithm system according to claim 1, characterized in that, The nonlinear prediction algorithm for kidney disease progression employs an improved Long Short-Term Memory (LSTM) network combined with a dual attention mechanism to construct the model. The formula for quantifying the risk of kidney disease progression is: P = μ·(P ·W + P_c·W_c + P_i·W_i) + (1-μ)·T·(1 - P_b·W_b); Where P is the probability of disease progression, ranging from [0,1]; μ is the core factor weighting coefficient, ranging from [0.65,0.8]; P W is the comprehensive value of core pathological factors. The core pathological factor weight vector satisfies ΣW =1; P_c is the comprehensive value of clinical dynamic factors, W_c is the weight vector of clinical dynamic factors and satisfies ΣW_c=1; P_i is the comprehensive value of intervention impact factors, W_i is the weight vector of intervention impact factors and satisfies ΣW_i=1; T is the time series correction coefficient, with a value range of [0.85, 1.15]; P_b is the comprehensive value of baseline risk factors, W_b is the weight vector of baseline risk factors and satisfies ΣW_b=1.

4. The intelligent algorithm system according to claim 1, characterized in that, The individualized treatment plan adaptive optimization algorithm adopts a four-layer architecture of "individual feature analysis - plan generation modeling - efficacy prediction and risk assessment - dynamic optimization calibration", and introduces a comprehensive adaptation quantification formula for individualized treatment plans, which is expressed as: S = α·F + β·(E - λ·R) + (1-α-β)·D; Wherein, S is the comprehensive suitability score of the treatment plan, with a value range of [0,10]; α is the individual characteristic suitability weight coefficient, with a value range of [0.35,0.45]; F is the individual characteristic suitability coefficient; β is the efficacy risk balance weight coefficient, with a value range of [0.4,0.5]; E is the efficacy prediction score; λ is the risk weight coefficient, with a value range of [1.2,1.5]; R is the adverse reaction risk score; and D is the dynamic feedback calibration coefficient.

5. The intelligent algorithm system according to claim 1, characterized in that, The intelligent optimization algorithm for prognosis follow-up adopts a four-layer architecture of "follow-up risk stratification - strategy generation and optimization - data collection and analysis - dynamic iterative adjustment", and introduces a comprehensive risk quantification formula for follow-up, which is expressed as: R_total = ω1·R_b + ω2·R_t + ω3·R_h + ε·K; Wherein, R_total is the follow-up comprehensive risk score, with a value range of [0,10]; ω1, ω2, and ω3 are the weight coefficients of the basic disease risk, treatment response risk, and home management risk, respectively, and satisfy ω1+ω2+ω3=1; R_b is the basic disease risk score, R_t is the treatment response risk score, and R_h is the home management risk score; ε is the dynamic correction coefficient, with a value range of [0.85,1.15]; and K is the strategy adaptation calibration coefficient, with a value range of [0.9,1.1].

6. The intelligent algorithm system according to claim 2, characterized in that, The feature fusion modeling of the MDS-ES algorithm adopts a three-order strategy of "local feature enhancement - cross-modal fusion - feature dimensionality reduction optimization". Local feature enhancement determines the importance weight of each feature through the analytic hierarchy process. The core feature weight is assigned a value of 0.6-0.8, and the secondary feature weight is assigned a value of 0.2-0.

4. A cross-modal fusion model with an attention mechanism is constructed, and the correlation coefficient between features of different modalities is calculated through a self-attention network. Feature dimensionality reduction optimization adopts a combination of principal component analysis (PCA) and sparse matrix factorization.

7. The intelligent algorithm system according to claim 3, characterized in that, The dual attention mechanism of the NPP-PP algorithm includes a temporal attention mechanism and a factor-related attention layer. The temporal attention mechanism assigns high attention weights to key time nodes of clinical dynamic factors. The factor association attention layer calculates the association coefficients between different categories of factors to generate a multi-factor nonlinear association matrix. The algorithm adopts a multi-task learning framework to simultaneously achieve three core tasks: progress risk probability prediction, progress time node prediction, and key driving factor identification.

8. The intelligent algorithm system according to claim 4, characterized in that, The ATO-TS algorithm's protocol generation modeling employs an improved collaborative filtering algorithm combined with clinical guideline constraints. A basic protocol library is constructed based on clinical guidelines. The collaborative filtering algorithm matches similar historical cases to extract effective protocol parameters. Protocol details are adjusted based on individual characteristics to generate 2-3 candidate protocols. The multi-task learning framework simultaneously outputs efficacy prediction results and adverse reaction risk assessment results. The analytic hierarchy process (AHP) is used to determine efficacy and risk weights and select the optimal protocol.

9. The intelligent algorithm system according to claim 5, characterized in that, The PFO-FU algorithm's follow-up risk stratification is based on a three-factor system: "basic disease risk - treatment response risk - home management risk". The analytic hierarchy process is used to assign weights, with basic disease risk weighted at 0.5, treatment response risk at 0.3, and home management risk at 0.

2. The comprehensive follow-up risk level is calculated and divided into three levels: high risk, medium risk, and low risk. Personalized follow-up strategies are then customized based on these levels.

10. A method for precise management of the entire life cycle of kidney disease based on the intelligent algorithm system described in any one of claims 1-9, characterized in that, Includes the following steps: Step 1: Collect multimodal data through multiple channels in the data layer, generate a standardized data matrix after preprocessing, and input it into the algorithm layer; Step 2: The MDS-ES algorithm performs hierarchical analysis, feature fusion, anomaly measurement, and result calibration on the data, and outputs early screening results; Step 3: The NPP-PP algorithm, based on the screening results, integrates multi-dimensional progression factors and outputs disease progression risk assessment results through nonlinear correlation modeling and quantitative formula calculation. Step 4: Based on the screening and prediction results, the ATO-TS algorithm integrates individual characteristics to generate candidate treatment plans, selects the optimal plan through efficacy and risk assessment, and pushes it to the application layer; Step 5: The PFO-FU algorithm, based on treatment plans and efficacy feedback, stratifies and customizes follow-up strategies, collects follow-up data, and dynamically iterates and optimizes to form a closed-loop management system throughout the entire cycle.

Images