Polycystic ovarian syndrome screening and typing hierarchical management system and method based on dynamic fusion of multi-modal eye image sclera features and clinical data

Abstract

The invention provides a polycystic ovarian syndrome screening and typing hierarchical management system and method based on dynamic fusion of multi-modal eye image sclera features and clinical data, and is applied to the field of data processing application. According to the method, data preprocessing is performed on data type information, a non-invasive eye image preprocessing algorithm and a space-time alignment algorithm are introduced, and multi-modal dynamic feature information is generated in combination with block chain evidence storage and differential privacy protection; abnormal feature recognition information is generated through a dynamic weight distributor and space-time correlation analysis in combination with feature probability distribution calculation and abnormal feature value filtering; the multi-modal dynamic feature information is processed, and a polycystic ovarian syndrome screening and typing result is generated through a three-level decision system; and fusing an androgen attention module and a metabolism contrast learning result, processing a screening and typing result, a risk influence factor and abnormal feature identification information, and generating a dynamic early warning and intervention result of the polycystic ovarian syndrome.

Description

Technical Field

[0001] This invention relates to the field of data processing applications, and in particular to a screening and classification management system and method for polycystic ovary syndrome based on the dynamic fusion of multimodal eye image scleral features and clinical data. Background Technology

[0002] Polycystic ovary syndrome (PCOS) is a common gynecological endocrine disorder affecting approximately 6%-10% of women of reproductive age worldwide, and its incidence is increasing annually. Currently, the clinical diagnosis of PCOS primarily relies on the Rotterdam criteria, which comprehensively consider indicators such as hyperandrogenism (e.g., hirsutism, acne), ovulation abnormalities (e.g., menstrual cycle disorders), and polycystic ovarian changes (as seen on ultrasound). However, these criteria have significant limitations; their sensitivity and specificity are approximately 70% and 80%, respectively, leading to frequent misdiagnosis and missed diagnosis.

[0003] In terms of risk assessment and early warning, current technologies cannot achieve dynamic risk assessment and accurate early warning for PCOS. Existing risk assessments are mostly based on static clinical indicators, failing to consider the dynamic changes of the disease over time and individual differences. For example, most risk assessment models rely solely on fixed indicators such as patient age, BMI, and sex hormone levels, which cannot monitor disease progression in real time, leading to delayed risk assessment. Furthermore, the lack of an effective early warning mechanism makes it difficult to issue timely warnings in the early stages of the disease or before it worsens.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this disclosure, and therefore includes information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this application is to provide a screening, classification, and grading management system and method for polycystic ovary syndrome (PCOS) based on the dynamic fusion of multimodal ocular scleral features and clinical data. This system at least partially overcomes the problems existing in current technologies. It acquires scleral feature images through non-invasive imaging from eight eyes, combines this with multi-source data including 16 structured clinical features, and generates multimodal dynamic feature information through improved U-Net preprocessing, spatiotemporal alignment algorithms, and blockchain notarization. An abnormal feature is extracted using a dynamic fusion engine, and screening and classification are achieved through a three-level decision-making system combined with a Hidden Markov Model. Risk influencing factors are generated by fusing androgen attention modules and metabolic contrastive learning. Finally, a spatiotemporal correlation early warning engine generates dynamic early warning and intervention results.

[0006] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of this application, a method for screening, subtyping, and grading management of polycystic ovary syndrome (PCOS) based on dynamic fusion of multimodal ocular scleral features and clinical data is provided. The method includes: acquiring data type information, influencing factor data, and abnormal state information; preprocessing the data type information by introducing a non-invasive ocular image preprocessing algorithm and a spatiotemporal alignment algorithm, combined with blockchain notarization and differential privacy protection, to generate multimodal dynamic feature information; processing the influencing factor data and abnormal state information based on a dynamic fusion engine, using a dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and abnormal feature value filtering, to generate abnormal feature identification information; processing the multimodal dynamic feature information to generate PCOS screening and subtyping results through a three-level decision system; processing the screening subtyping results, subtypes, and risk levels in the influencing factor data using a multi-task decision matrix, fusing androgen attention module and metabolic contrast learning results, to generate PCOS risk influencing factors; and processing the screening subtyping results, risk influencing factors, and abnormal feature identification information based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision strategy to generate dynamic early warning and intervention results for PCOS.

[0008] Another aspect of this application discloses a device for screening, classifying, and grading polycystic ovary syndrome (PCOS) based on the dynamic fusion of multimodal ocular scleral features and clinical data. The device comprises: an acquisition module for acquiring data type information, influencing factor data, and abnormal state information; the data type information includes scleral feature images and clinical indicator data; the influencing factor data includes 16 structured clinical features, 22 manual scleral image features, and 1024-dimensional scleral image depth features; and the abnormal state information includes ocular image acquisition interference information and clinical time-series mismatch data. A processing module is used to preprocess the data type information, introducing a non-invasive ocular image preprocessing algorithm and a spatiotemporal alignment algorithm, combined with blockchain notarization and differential privacy protection, to generate multimodal dynamic feature information. The dynamic fusion engine processes influencing factor data and abnormal state information. Through a dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and abnormal feature value filtering, it generates abnormal feature identification information. It processes multimodal dynamic feature information and generates polycystic ovary syndrome (PCOS) screening and subtyping results through a three-level decision-making system. It processes the screening and subtyping results, subtypes, and risk levels in the influencing factor data using a multi-task decision matrix, fusing androgen attention module and metabolic contrast learning results to generate PCOS risk influencing factors. Finally, based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision-making strategy, it processes the screening and subtyping results, risk influencing factors, and abnormal feature identification information to generate dynamic early warning and intervention results for PCOS.

[0009] According to another aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a second processor, implements the above-described method for screening, typing, and grading polycystic ovary syndrome based on dynamic fusion of multimodal ocular scleral features and clinical data.

[0010] This application presents a screening, subtyping, and grading management system and method for polycystic ovary syndrome (PCOS) based on the dynamic fusion of multimodal ocular scleral features and clinical data. It pioneers a dynamic fusion architecture of ocular and clinical data, overcoming three major technical challenges: non-invasive biomarker discovery, multimodal fusion, and dynamic modeling. By extracting 22 scleral features through the PCOS-SE-ResNet50 network and combining dynamic weight allocation and spatiotemporal correlation analysis, it achieves subtype transformation threshold capture, resulting in subtype identification accuracy higher than traditional methods. Clinical validation shows that this scheme achieves a screening accuracy exceeding 90% and can predict metabolic risks earlier, such as gestational diabetes, compared to traditional models. The system is interpretable, providing decision-making support through SHAP value ranking and spatiotemporal heatmaps. It supports modular expansion and real-time monitoring, forming a closed-loop management system of "screening-subtyping-intervention-tracking," providing non-invasive and precise diagnosis and treatment options for PCOS patients.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0012] Figure 1 This document illustrates a flowchart of a method for screening, classifying, and managing polycystic ovary syndrome based on the dynamic fusion of multimodal ocular scleral features and clinical data, according to an embodiment of this application. Figure 2 This illustration shows a schematic diagram of a scleral feature image preprocessing process based on dynamic fusion of multimodal ocular scleral features and clinical data, provided in an embodiment of this application. Figure 3 This illustration shows a schematic diagram of a three-level decision-making system based on the dynamic fusion of multimodal ocular scleral features and clinical data, provided in an embodiment of this application. Figure 4 This illustration shows a schematic diagram of the working principle of a spatiotemporal correlation early warning engine based on the dynamic fusion of multimodal ocular scleral features and clinical data, provided in an embodiment of this application. Figure 5 This illustration shows a schematic diagram of a polycystic ovary syndrome screening and classification management device based on the dynamic fusion of multimodal ocular scleral features and clinical data, according to an embodiment of this application. Detailed Implementation

[0013] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0014] The following is combined Figure 1 This application describes a method for screening, classifying, and grading polycystic ovary syndrome (PCOS) based on the dynamic fusion of multimodal ocular scleral features and clinical data, according to exemplary embodiments of this application. In one embodiment, this application also proposes a PCOS screening, classifying, and grading management system and method based on the dynamic fusion of multimodal ocular scleral features and clinical data. Figure 1 As shown: S101, obtain data type information, impact factor data, and abnormal status information.

[0015] In one embodiment, scleral images of the patient are acquired using an 8-eye non-invasive imaging device. For example, a polarized light source with a wavelength of 532 nm is used to acquire scleral images containing microvascular distribution, achieving a resolution of 300 dpi and covering eight positions, including the nasal and temporal sides. The patient's age, BMI, serum testosterone level, AMH value, and menstrual cycle records are collected. For example, a patient has a BMI of 26.5 kg / m², serum testosterone of 1.8 nmol / L, and irregular menstrual cycles (45-60 days).

[0016] Sixteen structured clinical features were included, such as age, BMI, testosterone level, fasting blood glucose, insulin level, and lipid profiles. For example, a 30-year-old patient had a fasting blood glucose of 5.8 mmol / L and an insulin resistance index (HOMA-IR) of 2.5. Twenty-two hand-crafted scleral image features, such as microaneurysm density (3 per square millimeter), vascular tortuosity index (1.8), and blood flow pulsatility coefficient of variation (25%), were obtained through a combination of manual annotation and software measurement. 1024-dimensional scleral image depth features were extracted using the PCOS-SE-ResNet50 network. For example, after enhancement by the androgen attention module, the vascular texture feature dimension in the feature vector output from the 5th convolutional layer was 0.72 (range 0-1).

[0017] Interference in eye image acquisition, such as blurred images due to patient blinking or impaired vascular recognition due to scleral pigmentation (e.g., an image with a signal-to-noise ratio below 30dB due to blinking), necessitates triggering a re-image mechanism. Mismatched clinical timelines, where the time interval between clinical indicator acquisition and eye image acquisition exceeds one month (e.g., serum testosterone testing on May 1, 2025, and eye image acquisition on June 15, 2025, a 45-day interval), require timeline alignment.

[0018] S102 performs data preprocessing on data type information, introduces non-invasive eye image preprocessing algorithm and spatiotemporal alignment algorithm, and combines blockchain notarization and differential privacy protection to generate multimodal dynamic feature information.

[0019] In one implementation, scleral feature images, clinical indicator data, blockchain evidence storage requirements, and differential privacy protection parameters are extracted and processed. A non-invasive eye image preprocessing algorithm and a spatiotemporal alignment algorithm are introduced to generate preprocessed scleral feature images, spatiotemporally aligned clinical indicator data, blockchain evidence storage identifiers, and differentially privacy-protected feature data. Figure 2 As shown, improved U-Net segmentation is performed on scleral feature images (such as scleral vascular images of 8 eyes at 300dpi) to remove eyelash occlusion; the acquisition time of clinical indicator data (such as patient BMI=26.5, serum testosterone 1.8nmol / L) is aligned using the Dynamic Time Warping (DTW) algorithm; at the same time, blockchain evidence storage requirements (such as clinical sensitive data fragmentation storage rules) and differential privacy parameters (ε=0.3) are extracted.

[0020] A patient's eye image was blurred due to blinking, and was deemed unacceptable by a quality screening CNN, triggering a real-time re-image capture. In the clinical data, the interval between serum testosterone testing and eye image acquisition was 45 days; this time difference was compressed to within 7 days using the DTW algorithm. The final result includes a pre-processed clear scleral image, spatiotemporally aligned clinical data (e.g., uniformly collected in June 2025), blockchain-based notarized hash values ​​(e.g., unique identifiers generated by SHA-256), and features processed for differential privacy (e.g., adding Laplacian noise to glycated scleral chromaticity values).

[0021] The preprocessed scleral feature images, spatiotemporally aligned clinical indicator data, blockchain-based evidence markers, and differentially protected privacy feature data are integrated to generate multimodal intermediate data containing image preprocessing features, temporally aligned clinical features, data evidence information, and privacy protection features. The preprocessed eye image features (e.g., vascular tortuosity index 1.8), aligned clinical features (e.g., fasting blood glucose 5.8 mmol / L), evidence markers, and privacy protection features (e.g., differential privacy markers with ε=0.3) are structurally integrated. The intermediate data structure is constructed as {eye image features: [vascular density, tortuosity], clinical features: [BMI, testosterone], evidence hash: "abc123", privacy marker: "ε=0.3"}.

[0022] The final generated multimodal intermediate data includes image preprocessing features (such as a denoised vascular clarity score of 90), time-aligned clinical features (such as a table of correspondence between menstrual cycles and hormone testing times), evidence information (blockchain block height 123456), and privacy features (insulin resistance index of 2.5±0.1 after differential privacy processing).

[0023] Based on the integrity, consistency, and security standards of multimodal intermediate data, abnormal data in the data preprocessing process is marked and filtered to generate data preprocessing anomaly records, including the abnormal data type, source module, processing stage, and impact level. Abnormal data is marked according to integrity (e.g., missing two eye positions in the eye image), consistency (e.g., discrepancies between BMI and body fat percentage), and security (e.g., blockchain notarization failure). For example, a patient's eye image showed a vessel recognition rate of <70% due to scleral pigmentation, and was marked as "image quality abnormality"; a clinical AMH value of 3.5 ng / mL was inconsistent with the patient's age of 30 years, and was judged as "indicator contradiction." The final generated anomaly record is as follows: {Type: Image quality abnormality, Source: Eye image acquisition module, Stage: Preprocessing, Impact level: High}.

[0024] The process involves validating and fusing multimodal intermediate data and preprocessed abnormal records to generate multimodal dynamic feature information. This information includes standardized eye image features, time-consistent clinical features, blockchain-based evidence, privacy-preserving features, and abnormal data markers. The intermediate data and abnormal records are validated, normal features are fused, and abnormal markers are retained. Eye image features (microaneurysm density 3 / mm²) and clinical features (HOMA-IR=2.5) in the intermediate data are normalized, and a "medium image quality" marker is added to the abnormal records.

[0025] The generated multimodal dynamic feature information includes standardized eye image features (vascular density feature vector [0.72, 0.85]), time-consistent clinical features (correlation matrix between pregnancy cycle and metabolic indicators), blockchain evidence identifier (latest block hash), privacy-preserving features (deep feature vector after differential privacy) and abnormal data markers (such as "clinical data time-series deviation, manual review required").

[0026] S103 processes influencing factor data and abnormal state information based on a dynamic fusion engine. Through a dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and abnormal feature value filtering, it generates abnormal feature identification information.

[0027] In one implementation, a spatiotemporal alignment algorithm and dynamic weight allocation technology are introduced based on a dynamic fusion engine. Combining patient state characteristics, features are extracted from 16 structured clinical features, 22 manual scleral features, and 1024-dimensional depth features in the influencing factor data, as well as from eye image acquisition interference and clinical time-series mismatch data in abnormal state information. This generates a multimodal feature vector set containing dynamic weight coefficients. The Dynamic Time Warping (DTW) algorithm is used to align the timelines of clinical data (such as BMI and testosterone levels) with scleral features (such as vascular tortuosity). The weight coefficients are dynamically adjusted according to the patient's pregnancy status (such as gestational age) (e.g., increasing the weight of insulin during pregnancy by 3.2 times).

[0028] A patient at 20 weeks of gestation had her fasting blood glucose (5.8 mmol / L) and scleral microvascular pulse entropy (0.65) weighted using the formula... The value is dynamically adjusted to 1.8 (1.0 in non-pregnant states). The multimodal feature vector set is as follows: [[BMI:26.5, vascular tortuosity:1.8, depth feature dimension 1:0.72,...], weight coefficients: [1.2,1.8,1.0,...]].

[0029] By comparing normal and abnormal feature vector sets and combining the temporal correlation of scleral abnormality regions located by spatiotemporal thermal mapping, the variation patterns of abnormal features and underlying shared features of polycystic ovary syndrome (PCOS) with disease progression and menstrual cycle were determined, generating a PCOS-related feature mapping relationship with scleral abnormality region localization. Comparing patient feature vectors with the mean values ​​of the normal population (e.g., microaneurysm density <2 / mm²), spatiotemporal thermal mapping revealed an abnormal increase in temporal scleral vessel density on day 14 of the menstrual cycle. One patient's temporal vessel density (3.5 / mm²) was 67% higher than the normal mean (2.1 / mm²), and spatiotemporal thermal mapping showed that this area remained abnormal for three menstrual cycles, mapping to a "hyperandrogen-angiogenesis" association pattern. The generated associated feature mapping relationship is as follows: {Abnormal area: temporal 12 o'clock position, Feature association: testosterone↑→vessel density↑, Temporal pattern: mid-menstrual cycle peak}.

[0030] The mapping relationship of polycystic ovary syndrome (PCOS) associated features is compared with a pre-set normal feature library and a PCOS risk feature library. Combined with clinical validation data, a feature deviation assessment result is generated. The deviation score (e.g., Euclidean distance) is calculated by comparing the patient's testosterone level of 1.8 nmol / L and vascular tortuosity of 1.8 with the pre-set hyperandrogenism feature template. The patient's testosterone level of 1.8 nmol / L and vascular tortuosity of 1.8 deviate from the hyperandrogenism template by 0.32 (threshold 0.5), thus classifying the patient as having a "hyperandrogenism tendency." The generated deviation assessment result is as follows: {Hyperandrogenism deviation: 0.32, Metabolic deviation: 0.18, Risk level: Medium}.

[0031] Combining atypical symptoms, pregnancy status data, and eye image quality from the patient's medical history, the feature deviation assessment results were weighted using the eye position importance weight distribution to generate an abnormal feature vector for polycystic ovary syndrome (PCOS) with fused feature contribution. The deviation was adjusted by weighting eye position importance weights (e.g., nasal side 0.2, temporal side 0.3) and atypical symptoms (e.g., regular but infrequent menstruation). The patient's eye image quality was moderate (low signal-to-noise ratio, weight coefficient 0.8), the temporal abnormal region weight was 0.3, and the atypical symptom correction coefficient was 0.9. The final deviation for hyperandrogenic type was adjusted to 0.32 × 0.3 × 0.8 × 0.9 = 0.069. The final generated abnormal feature vector is shown as [Hyperandrogenic: 0.069, Metabolic: 0.21, Complex: 0.58].

[0032] The results of the feature deviation assessment and the abnormal feature vectors of polycystic ovary syndrome (PCOS) were normalized and comprehensively calculated. Combined with the model's real-time processing efficiency, an abnormal feature identification information set containing feature correlation coefficients, abnormal probability values, and clinical warning lines was generated. After normalization, the abnormal probability was calculated using the Softmax function, and a clinical warning line was set (e.g., an abnormal probability > 0.7 triggers an alert). After comprehensive calculation, the abnormal probability of hyperandrogenism was 0.75, triggering a yellow warning line. The feature correlation coefficient showed a correlation of 0.82 between testosterone and vascular density. The final abnormal feature identification information set is as follows: {Feature correlation coefficient: 0.82, Abnormal probability: 0.75, Warning line: Yellow, Recommendation: Repeat testosterone testing + scleral imaging}.

[0033] S104 processes multimodal dynamic feature information and generates polycystic ovary syndrome screening and classification results through a three-level decision-making system.

[0034] In one implementation, feature extraction processing is performed on multimodal dynamic feature information to generate scleral feature indicators, clinical indicator parameters, blockchain evidence storage features, differential privacy features, data standardization features, and spatiotemporal alignment features. The data standardization features include preprocessed uniform-format eye image data and standardized values ​​of clinical indicators. The spatiotemporal alignment features include temporal matching accuracy and dynamic time warping parameters. Scleral features (such as vascular tortuosity), clinical indicators (such as BMI), blockchain evidence storage hash values, differential privacy tags, standardized eye image / clinical data format, and temporal alignment parameters are extracted from the multimodal dynamic feature information. For example, a patient's standardized eye image data shows an average vascular density of 0.68 in 8 eye positions (standardization range 0-1), a BMI of 26.5 (standardized to 0.72), and a DTW algorithm matching accuracy of 92% (threshold ≥ 80%) in the spatiotemporal alignment features. The following indicators were generated: scleral feature indicators: [vascular tortuosity = 1.8, microaneurysm density = 3 / mm²]; clinical indicator parameters: [BMI = 26.5, serum testosterone = 1.8 nmol / L]; blockchain evidence features: SHA-256 hash value = abc123; differential privacy features: Laplacian noise with ε = 0.3 added as a marker; data standardization features: eye image format unified to PNG, clinical indicator Z-score standardized results; spatiotemporal alignment features: DTW time difference compressed to 7 days, matching accuracy 92%.

[0035] Feature extraction processing is performed on the screening and typing data to generate three-level decision features, Hidden Markov Model features, multi-task learning features, and adaptive weight features. The three-level decision features include the probability output of the screening layer, the subtype mapping of the typing layer, and the threshold parameters of the risk grading layer. The Hidden Markov Model features include the state transition matrix and subtype transformation probabilities. The multi-task learning features include androgen attention weights and metabolic contrast loss functions. The adaptive weight features include dynamic weight update coefficients and feature importance ranking. Figure 3 As shown, the three-level decision probabilities, hidden Markov model parameters, multi-task learning weights, and adaptive weight coefficients are extracted from the screening and typing data.

[0036] The screening layer output a PCOS positive probability of 0.85, and the subtype layer maps it to hyperandrogenism (probability 0.62). The Hidden Markov Model (HMM) state transition matrix shows a transformation probability of 0.18 / year for hyperandrogenism to mixed subtype, and the androgen attention weight in the temporal region is 0.35. The final extracted feature data are as follows: Level 3 decision features: screening layer probability = 0.85, subtype layer subtype = hyperandrogenism, risk grading = medium risk (threshold 0.5); HMM features: state transition matrix T = [[0.7, 0.18, 0.12], ...], subtype transformation probability = 12-18% / year; multi-task learning features: androgen attention weight = 0.35 (temporal side), metabolic contrast loss function value = 0.28; adaptive weight features: dynamic weight update coefficient = 1.2 (pregnancy status), feature importance ranking = vascular density > testosterone.

[0037] Based on scleral feature indicators, clinical indicator parameters, blockchain evidence storage features, differential privacy features, data standardization features, and spatiotemporal alignment features, combined with three-level decision features, Hidden Markov Model features, multi-task learning features, and adaptive weight features, analysis and processing are performed to generate screening and genotyping results. These results characterize the screening probability, subtype, and correlation with multimodal features for polycystic ovary syndrome (PCOS), forming a PCOS status assessment result that includes screening accuracy probability, subtype conversion trend, and risk grading assessment. By integrating scleral / clinical features and decision features, a multi-task model is used to calculate the screening probability, subtype, and risk grading, and a Hidden Markov Model is used to predict the conversion trend.

[0038] After considering all characteristics, a screening probability of 0.85 was determined to be positive, with the subtype being hyperandrogenic (confidence level 0.62) and the risk level being medium. The Hidden Markov Model predicted a 0.15 probability of conversion to mixed subtype within one year, triggering an alert. The final screening subtype result was PCOS positive (probability 0.85), subtype = hyperandrogenic, risk level = medium; the status assessment result was screening accuracy probability = 90%, subtype conversion trend = hyperandrogenic → mixed (probability 0.15 / year), and risk level assessment = medium risk.

[0039] S105 processes the screening and typing results, subtypes, and risk levels in the impact factor data using a multi-task decision matrix, and integrates the androgen attention module and metabolic contrast learning results to generate risk impact factors for polycystic ovary syndrome.

[0040] In one implementation, feature extraction processing is performed on the screening classification results, subtype types, and risk levels in the impact factor data to generate screening probability features, subtype type features, risk level features, multi-task decision matrix features, androgen attention features, and metabolic contrastive learning features. The metabolic contrastive learning features include the feature distance between insulin-resistant and normal samples and the contrastive loss function parameters. Figure 3 As shown, screening probability, subtype type, risk level, multi-task decision matrix parameters, androgen attention weight, and metabolic contrast learning parameters were extracted from the influencing factor data. A patient's screening layer output a PCOS positive probability of 0.85, a subtype of hyperandrogenous (probability 0.62), and a risk level of medium risk. In the multi-task decision matrix, the hyperandrogenous → mixed subtype conversion weight was 0.25, the androgen attention weight in the temporal region was 0.35, and the metabolic contrast loss function value was 0.28 (the feature distance between insulin-resistant samples and normal samples was 1.2).

[0041] The final extracted features are as follows: screening probability feature: 0.85; subtype feature: hyperandrogenism (confidence 0.62); risk level feature: medium risk; multi-task decision matrix feature: [[0.7,0.25,0.05],...] (state transition weights); androgen attention feature: temporal region 0.35; metabolic contrastive learning feature: insulin resistance feature distance 1.2, loss function 0.28.

[0042] Feature extraction was performed on the fused data to generate risk impact factor features, dynamic weight features, association analysis features, and comprehensive assessment features. The association analysis features included the association matrix between subtype type and risk level, and the interaction relationship between androgens and metabolic characteristics. Risk impact factor parameters, dynamic weights, subtype-risk association matrices, and comprehensive assessment indicators were extracted from the fused data. Among the risk impact factors, the weight of hyperandrogenism on metabolic syndrome was 1.8, and the dynamic weight was adjusted to 1.2 due to pregnancy status. The subtype-risk association matrix showed a probability of 0.6 for hyperandrogenism → intermediate risk, and an interaction coefficient of 0.75 between androgens and metabolic characteristics (for every 0.5 nmol / L increase in testosterone, vascular density increases by 0.1 vessels / mm²).

[0043] The final extracted features are as follows: Risk influencing factors: Hyperandrogenism influence weight 1.8, metabolic syndrome risk coefficient 0.9; Dynamic weight features: Pregnancy status adjustment coefficient 1.2; Association analysis features: Subtype-risk matrix [[0.6,0.3,0.1],...], androgen-metabolism interaction coefficient 0.75; Comprehensive assessment features: Probability of risk escalation in the next 6 months 0.3.

[0044] Based on screening probability features, subtype features, risk level features, multi-task decision matrix features, androgen attention features, and metabolic contrastive learning features, combined with risk impact factor features, dynamic weight features, association analysis features, and comprehensive assessment features, a risk impact factor for polycystic ovary syndrome (PCOS) is generated. The risk impact factor is calculated using a multi-task matrix by integrating screening, subtype, risk, and association features. With a comprehensive screening probability of 0.85, a high androgen weight of 1.8, a baseline value of 0.6 for medium risk, a dynamic weight of 1.2, and an interaction coefficient of 0.75, the calculated risk impact factor is 0.85 × 1.8 × 0.6 × 1.2 × (1 + 0.75) = 1.33 (threshold 1.0 triggers an early warning). The generated result is a PCOS risk impact factor of 1.33 (high-risk warning); risk composition: screening probability 30%, subtype 40%, risk level 20%, and interaction features 10%.

[0045] S106, based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision-making strategy, processes screening and typing results, risk influencing factors and abnormal feature identification information to generate dynamic early warning and intervention results for polycystic ovary syndrome.

[0046] In one implementation, a spatiotemporal correlation early warning engine analyzes and processes screening classification results, risk impact factors, and abnormal feature identification information to generate a PCOS risk assessment value. The abnormal feature identification information includes the abnormal subtype, the time period of the abnormality, and the degree of correlation with multimodal features. Simultaneously, the system accesses screening classification results (e.g., subtype type, screening probability) output by the three-level decision-making system, the quantitative risk value generated by the risk impact factor calculation module (e.g., 1.33), and data such as scleral abnormality areas and temporal change trends marked by the abnormal feature identification module. Figure 4 As shown, a three-dimensional matrix of "time-space-feature" is constructed. For example, using a month as the time granularity, features such as vascular density and clinical hormone levels in eight ocular positions of the sclera are mapped to a spatiotemporal coordinate system to analyze dynamic evolution patterns. Dynamic thresholds are set based on clinical validation data (e.g., a cohort of 400 PCOS patients), pre-setting thresholds for key feature changes (e.g., monthly increase in vascular density ≥ 0.05 vessels / mm², testosterone fluctuation ≥ 0.5 nmol / L). Exceeding these thresholds triggers risk escalation. A logistic regression model maps the integrated feature parameters to risk levels (low / medium / high). For example, a risk impact factor ≥ 1.0 and abnormal features persisting for more than 3 months are classified as high risk.

[0047] Specifically, the three-level decision-making system outputs a probability of 0.85 for the Kaohsiung type, with a risk classification of medium risk (initial threshold 0.5). This is achieved through the formula... (Where 1.8 is the weighting for hyperandrogenism, 0.6 is the baseline value for medium risk, 1.2 is the pregnancy status adjustment coefficient, and 0.75 is the androgen-metabolism interaction coefficient), exceeding the risk threshold of 1.0.

[0048] Spatiotemporal heatmaps showed that the vascular density at the 12 o'clock position on the temporal side increased from 3.0 vessels / mm² to 3.18 vessels / mm², with a monthly growth rate of 0.06 vessels / mm² (threshold 0.05). Testosterone levels also increased from 1.8 nmol / L to 2.3 nmol / L within 3 months (an increase of 0.5 nmol / L), consistent with the correlation pattern of "increased testosterone → increased vascular density". The monthly increase in vascular density exceeding the threshold (+20%) triggered a warning of "enhanced hyperandrogen activity". The risk factor 1.33 > 1.0, and the abnormal characteristics persisted for 3 months, upgrading the risk level from medium to high. Multimodal association analysis showed a dynamic correlation between testosterone and vascular density (Pearson coefficient 0.82), validating the consistency of pathological progression.

[0049] The final output results are as follows: PCOS risk assessment value: High risk (based on both risk impact factors and the persistence of abnormal features). Abnormal subtype: Hyperandrogenism (screening classification confidence level 0.85, and consistent with dynamic changes in androgen-related characteristics). Abnormal period: Recent 3 months (spatiotemporal matrix shows that the abnormal features started 3 months ago and show a continuous worsening trend). Multimodal association: For every 0.5 nmol / L increase in testosterone, the corresponding temporal vessel density increases by 0.1 vessels / mm² (the time sequence matching accuracy is 92% by aligning hormone detection and eye image acquisition time using a dynamic time warping algorithm), consistent with the "androgen-scleral angiogenesis" association map.

[0050] Dynamic threshold adaptation, unlike traditional static thresholds (such as the Rotterdam standard), dynamically adjusts risk assessment thresholds based on individual patient pathological characteristics (such as pregnancy status and BMI). For example, after increasing the insulin weight by 3.2 times during pregnancy, the vascular density threshold is correspondingly lowered to 0.04 vessels / mm². By sorting SHAP values ​​(e.g., vascular density contribution of 35%, testosterone contribution of 28%) and using spatiotemporal heatmaps (marking abnormal areas in the temporal region), clinically interpretable risk assessment reports are generated, addressing the limitations of black-box models. The Dynamic Time Warping (DTW) algorithm aligns the timelines of clinical data (such as hormone testing) and ocular image data (such as vascular density), eliminating temporal mismatch errors (e.g., compressing a 45-day interval to within 7 days) and improving the timeliness of risk assessment.

[0051] The multi-objective dynamic decision-making strategy parameter set within the spatiotemporal correlation early warning engine is processed based on the PCOS risk assessment value to generate a decision parameter correction set. This set includes subtype conversion threshold parameters, risk grading weight parameters, and intervention matching parameters. Based on the PCOS risk assessment value (e.g., high / medium / low risk), a predefined parameter adjustment rule library is invoked. For example, at high risk, the subtype conversion threshold is lowered (accelerating early warning); at medium risk, the corresponding weight is increased (strengthening monitoring); and at low risk, the default parameters are maintained. A decision tree model trained on clinical data from 400 PCOS patients shows that, for example, when the risk assessment value is ≥1.0, 92% of cases show subtype conversion within 6 months, requiring threshold adjustment for early warning.

[0052] The multi-parameter collaborative correction mechanism is as follows: Subtype conversion threshold: This affects the state transition probability of the Hidden Markov Model. For example, the annual conversion threshold from hyperandrogenic to mixed subtype is lowered from 0.18 to 0.15, meaning an earlier conversion warning is triggered. Risk grading weights: The weight ratios of low / medium / high risk are redistributed through the Softmax function. For example, increasing the weight of medium risk by 20% enhances the model's sensitivity to this level of feature. Intervention matching parameters: This is linked to the intervention protocol library of clinical guidelines, dynamically matching measures according to subtype and risk level (e.g., metformin treatment for hyperandrogenic). Patient risk assessment value = high risk (1.33), abnormal subtype = hyperandrogenic, temporal vessel density increased by 0.06 vessels / mm² per month in the past 3 months (exceeding the threshold of 0.05), indicating a high probability of conversion to mixed subtype.

[0053] The specific corrections are as follows: The subtype conversion threshold is lowered. The original threshold for the annual conversion probability from hyperandrogenism to mixed subtype was 0.18 (i.e., an alert is triggered when the predicted conversion probability is ≥0.18). The current vascular density growth rate of the patient corresponds to an annual conversion probability prediction of 0.21 (exceeding the original threshold of 0.18). To capture the conversion threshold earlier, the threshold is lowered to 0.15, advancing the alert trigger time by approximately 4 months. Risk grading weights are adjusted: Original weights: [low 0.3, medium 0.4, high 0.3]. After correction: the medium-risk weight is increased by 20% to 0.5, the high-risk weight remains at 0.3, and the low-risk weight is reduced to 0.2. The model's priority in identifying medium-risk characteristics is enhanced; for example, when patient indicators fluctuate within the medium-risk range, the system response speed is improved by 50%.

[0054] The original plan for hyperandrogenemia was routine intervention consisting of "lifestyle adjustments + follow-up examination after 3 months". The revised plan includes "testosterone monitoring + metformin 500mg / day", with the addition of monthly scleral imaging (originally every 3 months). Metformin was verified to reduce the rate of vascular density increase in patients with hyperandrogenemia by up to 30%, and combined with monthly eye imaging, pathological progression can be tracked in real time. The subtype conversion threshold is 0.15 / year. This parameter directly affects the state transition matrix T of the Hidden Markov Model. For example, in the revised matrix, T[Hyperandrogen → Mixed] is adjusted from 0.18 to 0.15, meaning the model considers a 15% probability of subtype conversion per year; values ​​below this value do not trigger an alert. Risk grading weights are set to [Low 0.2, Medium 0.5, High 0.3]. These weight vectors are used for weighted calculations of the multi-task decision matrix. For example, the output value of a medium-risk feature is multiplied by a coefficient of 0.5, making the final risk score more biased towards the medium-risk level. Intervention matching = Kaohsiung type → Metformin + monthly scleral imaging: This mapping relationship is stored in a blockchain smart contract. When the system detects a high risk of Kaohsiung type, it automatically calls the contract to generate intervention instructions, and at the same time, it is stored on the blockchain to ensure that it cannot be tampered with (hash value SHA-256: abc123).

[0055] The decision parameter correction set is parsed and transformed to generate dynamic early warning and intervention results for PCOS. These results characterize the risk warning levels for different subtypes, the intervention plans corresponding to the gestational period, and the optimization effect of individualized intervention pathways. The system receives the decision parameter correction set (such as subtype conversion thresholds, risk grading weights, and intervention matching parameters) and converts these technical parameters into executable early warning rules and intervention instructions through a clinical semantic mapper. For example, "subtype conversion threshold = 0.15 / year" is mapped to "trigger an early warning when the predicted conversion probability is ≥15%". Combining real-time patient data (such as gestational period and eye image quality), a predefined intervention protocol template library (based on standardized pathways validated in 400 clinical cases) is invoked.

[0056] A three-level interpretable system based on spatiotemporal heatmaps and feature importance ranking is employed: Level 1 warning: risk level color coding (e.g., yellow represents high risk); Level 2 interpretation: contribution of key features (e.g., vascular density contributes 35%, testosterone contributes 28%); Level 3 localization: spatiotemporal heatmap of scleral abnormality areas (e.g., abnormal vascular density at the 12 o'clock position on the temporal side). Individualized intervention pathways are generated, and a dynamic weight allocator adjusts the intensity and frequency of interventions based on patient-specific characteristics (e.g., pregnancy status, BMI). For example, the frequency of testosterone monitoring for pregnant patients is increased from once a month to once every two weeks.

[0057] The target patient was 16 weeks pregnant and screened as having hyperandrogenic cystic disease (probability 0.85), with a risk assessment value of 1.33 (high risk). The decision parameter adjustment set included: lowering the subtype conversion threshold to 0.15 / year; increasing the weight of intermediate risk by 20%; and the intervention matching being "metformin + monthly scleral imaging". When the risk assessment value 1.33 > the threshold 1.0, and the pregnancy status belonged to the high-risk group (the risk of metabolic complications is increased 2.3 times in pregnant PCOS patients), a yellow alert (high risk) was triggered. The heatmap showed a monthly increase of 0.06 vessels / mm² in temporal vessel density (exceeding the threshold of 0.05), and SHAP value ranking showed that this feature contributed the most to the risk (35%); it was associated with an increase of 0.5 nmol / L in testosterone levels, forming a pathological correlation map of "↑ testosterone → ↑ vessel density".

[0058] Metformin 500mg / day can reduce the homeostasis model index (HOMA-IR) by 22% in pregnant women with hyperandrogenism and slow the increase in vascular density. Monitoring frequency is adjusted to monthly scleral imaging: a shorter interval than the usual 3 months, allowing for real-time capture of subtype transformation signs. If Δvascular density > 0.05 / month → referral to a reproductive specialist: when the rate of increase in abnormal characteristics exceeds a threshold, a cross-departmental collaboration process is triggered, automatically generating a referral form through the hospital's HIS system (interface conforming to the HL7FHIR standard).

[0059] Dynamic early warning system indicates a high risk of conversion from hyperandrogenic to mixed types (yellow). The warning information includes the risk type: conversion from hyperandrogenic to mixed types; supporting evidence: monthly increase in vascular density of 0.06 vessels / mm², and an increase in testosterone of 0.5 nmol / L; time window: predicted conversion probability of 21% within the next 6 months (exceeding the revised threshold of 15%). Metformin 500mg / day + monthly scleral imaging before 20 weeks of gestation. Drug dosage: 500mg / day (adjusted based on BMI=26.5kg / m², maximum dose not exceeding 1500mg / day); monitoring method: 8-eye scleral imaging (wavelength 532nm polarized light, resolution 300dpi), data stored on blockchain (hash value SHA-256:xyz789). Referral requests are automatically pushed to the reproductive specialist's workstation via the hospital API, along with a multimodal patient characteristic report (including ocular images, clinical indicators, and risk evolution map).

[0060] Using "monthly increase in vessel density > 0.05 vessels / mm²" as an independent early warning trigger, the system detects risks 4-6 months earlier than traditional static threshold models. Through reinforcement learning, the intervention intensity is dynamically adjusted. For example, when a patient's response to metformin is insufficient (HOMA-IR decrease < 10%), the system automatically suggests combining it with spironolactone. The decision logic for optimizing the pathway is stored in a smart contract (immutable). The early warning report includes three levels of interpretability (feature importance, eye position localization, and pathological correlation), meeting the FDA's transparency requirements for medical AI. Intervention results are fed back in real time through the patient's app (e.g., medication adherence, monitoring data uploads), forming a closed-loop management process of "screening-classification-intervention-tracking".

[0061] This application acquires scleral feature images through non-invasive imaging from eight eyes, combines them with multi-source data including 16 structured clinical features, and generates multimodal dynamic feature information through improved U-Net preprocessing, spatiotemporal alignment algorithms, and blockchain notarization. Anomaly features are extracted using a dynamic fusion engine, and screening and classification are achieved through a three-level decision-making system combined with a Hidden Markov Model. Risk influencing factors are generated by fusing an androgen attention module and metabolic contrastive learning. Finally, a spatiotemporal correlation early warning engine generates dynamic early warning and intervention results.

[0062] This pioneering dynamic fusion architecture of ocular and clinical data overcomes three major technical challenges: non-invasive biomarker discovery, multimodal fusion, and dynamic modeling. By extracting 22 scleral features through the PCOS-SE-ResNet50 network and combining dynamic weight allocation and spatiotemporal correlation analysis, it achieves subtype transformation threshold capture, resulting in subtype identification accuracy higher than traditional methods. Clinical validation shows that this approach achieves a screening accuracy exceeding 90% and can predict metabolic risks earlier, such as gestational diabetes, compared to traditional models. The system is interpretable, providing decision-making support through SHAP value ranking and spatiotemporal heatmaps. It supports modular expansion and real-time monitoring, forming a closed-loop management system of "screening-subtyping-intervention-tracking," providing non-invasive and precise diagnosis and treatment solutions for PCOS patients.

[0063] In one implementation, such as Figure 5 As shown, this application also provides a device for screening, typing, and grading polycystic ovary syndrome based on the dynamic fusion of multimodal ocular scleral features and clinical data, comprising: The acquisition module 501 is used to acquire data type information, impact factor data and abnormal status information. The data type information includes scleral feature images and clinical indicator data. The impact factor data includes 16 structured clinical features, 22 manual features of scleral images and 1024-dimensional scleral image depth features. The abnormal status information includes eye image acquisition interference information and clinical time sequence mismatch data. Processing module 502 is used to preprocess data of data types, introduce non-invasive eye image preprocessing algorithm and spatiotemporal alignment algorithm, and combine blockchain notarization and differential privacy protection to generate multimodal dynamic feature information; based on the dynamic fusion engine, it processes the influencing factor data and abnormal state information, and generates abnormal feature identification information by combining feature probability distribution calculation and abnormal feature value filtering through dynamic weight allocator and spatiotemporal correlation analysis; it processes the multimodal dynamic feature information and generates polycystic ovary syndrome screening and subtyping results through a three-level decision system; it processes the screening subtyping results, subtypes, and risk levels in the influencing factor data with a multi-task decision matrix, and integrates the androgen attention module and metabolic contrast learning results to generate polycystic ovary syndrome risk influencing factors; based on the spatiotemporal correlation early warning engine and combined with multi-objective dynamic decision strategy, it processes the screening subtyping results, risk influencing factors and abnormal feature identification information to generate polycystic ovary syndrome dynamic early warning and intervention results.

[0064] The various embodiments in this application are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for evaluating the method, electronic device, electronic device, and readable storage medium for screening and classifying polycystic ovary syndrome (PCOS) based on the dynamic fusion of multimodal ocular scleral features and clinical data are basically similar to the PCOS screening and classifying management method embodiments described above. Therefore, the description is relatively simple, and relevant parts can be referred to in the description of the PCOS screening and classifying management method embodiments described above.

Claims

1. A method for screening, classifying, and managing polycystic ovary syndrome (PCOS) based on dynamic fusion of multimodal ocular scleral features and clinical data, characterized in that... include: Acquire data type information, impact factor data, and abnormal status information. Data type information includes scleral feature images and clinical indicator data. Impact factor data includes 16 structured clinical features, 22 manual scleral image features, and 1024-dimensional scleral image depth features. Abnormal status information includes eye image acquisition interference information and clinical time series mismatch data. Data preprocessing is performed on data type information, introducing non-invasive eye image preprocessing algorithm and spatiotemporal alignment algorithm, combined with blockchain notarization and differential privacy protection to generate multimodal dynamic feature information; Based on the dynamic fusion engine, the impact factor data and abnormal state information are processed. Through dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and abnormal feature value filtering, abnormal feature identification information is generated. Multimodal dynamic feature information is processed, and polycystic ovary syndrome screening and classification results are generated through a three-level decision-making system; The screening and typing results, subtypes, and risk levels in the impact factor data are processed in conjunction with a multi-task decision matrix. The results of the androgen attention module and metabolic contrast learning are integrated to generate risk impact factors for polycystic ovary syndrome. Based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision-making strategy, the screening and typing results, risk influencing factors and abnormal feature identification information are processed to generate dynamic early warning and intervention results for polycystic ovary syndrome.

2. The method as described in claim 1, characterized in that, Data preprocessing is performed on the data type information, introducing non-invasive eye image preprocessing algorithms and spatiotemporal alignment algorithms. Combined with blockchain notarization and differential privacy protection, multimodal dynamic feature information is generated, including: The system extracts and processes scleral feature images, clinical indicator data, blockchain evidence storage requirements, and differential privacy protection parameters. It introduces non-invasive eye image preprocessing algorithms and spatiotemporal alignment algorithms to generate preprocessed scleral feature images, spatiotemporally aligned clinical indicator data, blockchain evidence storage identifiers, and differential privacy-protected feature data. The preprocessed scleral feature images, spatiotemporally aligned clinical indicator data, blockchain evidence identification, and differentially privacy-protected feature data are integrated to generate multimodal intermediate data that includes image preprocessing features, temporally aligned clinical features, data evidence information, and privacy protection features. Based on the integrity, consistency and security standards of multimodal intermediate data, abnormal data in the data preprocessing process is marked and filtered to generate data preprocessing abnormal records, including abnormal data type, source module, processing stage and impact level; The multimodal intermediate data and preprocessed abnormal records are verified and fused to generate multimodal dynamic feature information, which includes standardized eye image features, time-consistent clinical features, blockchain evidence identification, privacy protection features and abnormal data markers.

3. The method as described in claim 1, characterized in that, Based on a dynamic fusion engine, influencing factor data and anomalous state information are processed. Through a dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and anomalous feature value filtering, anomalous feature identification information is generated, including: Based on the dynamic fusion engine, a spatiotemporal alignment algorithm and dynamic weight allocation technology are introduced. Combined with patient status characteristics, features are extracted from 16 structured clinical features, 22 scleral manual features and 1024-dimensional depth features in the impact factor data, as well as eye image acquisition interference and clinical time sequence mismatch data in abnormal status information, to generate a multimodal feature vector set containing dynamic weight coefficients. By comparing the normal and abnormal feature vector sets and combining the temporal correlation of the scleral abnormality region located by the spatiotemporal heat map, the change pattern of the abnormal features and underlying shared features of polycystic ovary syndrome with the course of the disease and the menstrual cycle is determined, and a mapping relationship of polycystic ovary syndrome associated features with the location of scleral abnormality region is generated. The mapping relationship of polycystic ovary syndrome (PCOS) associated features is compared with the pre-set normal feature library and PCOS risk feature library, and combined with clinical validation data, feature deviation assessment results are generated. Combining atypical symptoms, pregnancy status data, and eye image quality from the patient's medical history, the feature deviation assessment results are weighted using the eye position importance weight distribution to generate an abnormal feature vector of polycystic ovary syndrome with fused feature contribution. The feature deviation assessment results and abnormal feature vectors of polycystic ovary syndrome are normalized and integrated. Combined with the real-time processing efficiency of the model, an abnormal feature identification information set containing feature correlation coefficients, abnormal probability values ​​and clinical warning lines is generated.

4. The method as described in claim 2, characterized in that, Multimodal dynamic feature information is processed, and polycystic ovary syndrome (PCOS) screening and classification results are generated through a three-level decision-making system, including: Feature extraction processing is performed on multimodal dynamic feature information to generate scleral feature indicators, clinical indicator parameters, blockchain evidence storage features, differential privacy features, data standardization features, and spatiotemporal alignment features. Among them, data standardization features include preprocessed uniform format eye image data and clinical indicator standard values, and spatiotemporal alignment features include temporal matching accuracy and dynamic time warping parameters. Feature extraction processing is performed on the screening and typing data to generate three-level decision features, hidden Markov model features, multi-task learning features, and adaptive weight features. The three-level decision features include the probability output of the screening layer, the subtype mapping of the typing layer, and the threshold parameters of the risk grading layer; the hidden Markov model features include the state transition matrix and the subtype transformation probability; the multi-task learning features include androgen attention weights and the metabolic contrast loss function; and the adaptive weight features include dynamic weight update coefficients and feature importance ranking. Based on scleral feature indicators, clinical indicator parameters, blockchain evidence storage features, differential privacy features, data standardization features, and spatiotemporal alignment features, combined with three-level decision features, hidden Markov model features, multi-task learning features, and adaptive weight features, analysis and processing are performed to generate screening and typing results. Among them, the screening and typing results are used to characterize the screening probability of polycystic ovary syndrome, subtype type, and the degree of correlation with multimodal features, forming a polycystic ovary syndrome status assessment result that includes screening accuracy probability, subtype conversion trend, and risk grading assessment.

5. The method as described in claim 1, characterized in that, The screening and subtype results, risk levels, and risk levels in the impact factor data are processed using a multi-task decision matrix. The results of the androgen attention module and metabolic contrastive learning are then integrated to generate risk factors for polycystic ovary syndrome, including: Feature extraction processing is performed on the screening classification results, subtypes, and risk levels in the impact factor data to generate screening probability features, subtype features, risk level features, multi-task decision matrix features, androgen attention features, and metabolic contrast learning features. The metabolic contrast learning features include the feature distance between insulin resistance and normal samples and the contrast loss function parameters. Feature extraction processing is performed on the fused data to generate risk impact factor features, dynamic weight features, correlation analysis features, and comprehensive evaluation features. Among them, the correlation analysis features include the correlation matrix between subtype type and risk level, and the interaction relationship between androgen and metabolic features. Based on screening probability characteristics, subtype characteristics, risk level characteristics, multi-task decision matrix characteristics, androgen attention characteristics, and metabolic contrast learning characteristics, combined with risk influencing factor characteristics, dynamic weight characteristics, correlation analysis characteristics, and comprehensive evaluation characteristics, risk influencing factors for polycystic ovary syndrome are generated.

6. The method as described in claim 5, characterized in that, Based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision-making strategy, the screening and typing results, risk influencing factors, and abnormal feature identification information are processed to generate dynamic early warning and intervention results for polycystic ovary syndrome, including: Based on the spatiotemporal correlation early warning engine, the screening and classification results, risk impact factors and abnormal feature identification information are analyzed and processed to generate PCOS risk assessment value. Among them, the abnormal feature identification information includes abnormal subtype type, abnormal occurrence time period and degree of correlation with multimodal features. Based on the PCOS risk assessment value, the multi-objective dynamic decision-making strategy parameter set inside the spatiotemporal correlation early warning engine is processed to generate a decision parameter correction set. The multi-objective dynamic decision-making strategy parameter set includes subtype conversion threshold parameters, risk classification weight parameters, and intervention measure matching parameters. The decision parameter correction set is parsed and transformed to generate dynamic early warning and intervention results for PCOS. The dynamic early warning and intervention results are used to characterize the risk warning level of different subtypes, the intervention plan corresponding to the pregnancy cycle, and the optimization effect of individualized intervention path.

7. A device for screening, classifying, and grading polycystic ovary syndrome (PCOS) based on dynamic fusion of multimodal ocular scleral features and clinical data, characterized in that... The device includes: The acquisition module is used to acquire data type information, impact factor data, and abnormal status information. Data type information includes scleral feature images and clinical indicator data. Impact factor data includes 16 structured clinical features, 22 manual scleral image features, and 1024-dimensional scleral image depth features. Abnormal status information includes eye image acquisition interference information and clinical time series mismatch data. The processing module is used to preprocess data of various data types, introducing non-invasive eye image preprocessing algorithms and spatiotemporal alignment algorithms, combined with blockchain notarization and differential privacy protection, to generate multimodal dynamic feature information; based on a dynamic fusion engine, it processes influencing factor data and abnormal state information, using a dynamic weight allocator and spatiotemporal correlation analysis, combined with feature probability distribution calculation and abnormal feature value filtering, to generate abnormal feature identification information; it processes the multimodal dynamic feature information, generating polycystic ovary syndrome (PCOS) screening and classification results through a three-level decision system; it processes the screening and classification results, subtypes, and risk levels in the influencing factor data in combination with a multi-task decision matrix, and integrates the androgen attention module and metabolic contrast learning results to generate PCOS risk influencing factors; based on a spatiotemporal correlation early warning engine combined with a multi-objective dynamic decision strategy, it processes the screening and classification results, risk influencing factors, and abnormal feature identification information to generate dynamic early warning and intervention results for PCOS.

8. An electronic device, characterized in that, include: First processor; and memory for storing executable instructions of the first processor; The first processor is configured to execute the polycystic ovary syndrome screening and classification management method based on dynamic fusion of multimodal ocular scleral features and clinical data as described in any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the second processor, it implements the polycystic ovary syndrome screening and classification management method based on the dynamic fusion of multimodal ocular scleral features and clinical data as described in any one of claims 1 to 6.

Images