Artificial intelligence-based health remaining life prediction system and method

Abstract

The application discloses a health remaining life prediction system and method based on artificial intelligence. S1, six kinds of multi-dimensional characteristics X of population characteristics, chronic disease spectrum, physical examination, biochemical examination, function evaluation and lifestyle are collected; S2, a mixed survival model of Cox and DeepSurv combined through Stacking is trained by using a desensitization queue, and a risk score h(X) and a baseline cumulative risk H_0(t) are output; S3, a survival function S(t|X) and a conditional remaining life E[T|T>t_0,X] are calculated; S4, the uncertainty is quantified by using a Monte Carlo Dropout and a bootstrap method, and a 95% confidence interval is output; S5, key factors are analyzed by using SHAP, and natural language explanation is generated by using a template; S6, a report is pushed to a tested person, a housekeeper and a family member end; and S7, actual evolution and prediction are compared periodically, and retraining and gray switching are triggered. The method has high prediction accuracy, is explainable, uncertainty is quantifiable, and is self-optimizable.

Description

[0001] Invention Title: Health Life Prediction System and Method Based on Artificial Intelligence Technical Field

[0002] This invention relates to the field of computer-aided decision-making technology at the intersection of smart health and artificial intelligence, specifically to a method based on multi-source health data and employing hybrid... A method and system for accurate prediction of an individual's healthy life expectancy using a combined survival model, including uncertainty quantification, interpretable output, and continuous online optimization. Background Technology

[0003] 2.1 Current Status of Technology As my country enters a deeply aging society, accurate assessment of individual healthy life expectancy is crucial for elderly care services and health. It has significant value in scenarios such as management, commercial insurance pricing, and medical resource allocation. Current related technologies can be divided into three categories: 1. Traditional method based on actuarial tables: using demographic statistics and life tables (such as the "Chinese Life Expectancy Table") to obtain averages by gender and age. The remaining lifespan. This method is coarse-grained and does not take into account individual health status, lifestyle, and chronic disease conditions.

[0004] 2. Prediction methods based on single regression models: such as linear regression, logistic regression, and the traditional Cox proportional hazards model, using a few indicators. (e.g., BMI, blood pressure) modeling mortality risk. This method is interpretable but has limited accuracy and cannot capture nonlinear relationships between features.

[0005] 3. Deep Learning-Based End-to-End Prediction Methods: Utilizing LSTM, Transformer, and other architectures to model time-series health data and output survival probability. The method has high accuracy but poor interpretability, unquantifiable uncertainty, and susceptibility to sample bias.

[0006] 2.2 Technical Issues Based on the analysis of the prior art described above, the present invention aims to solve the following technical problems: Problem 1: The granularity is too coarse. Relying solely on demographic characteristics (age, sex) for prediction lacks the ability to differentiate between highly individualized middle-aged and elderly populations. Elderly people of the same age and gender but with vastly different chronic disease profiles may have a life expectancy that differs by more than 10 years.

[0007] Problem 2: Poor interpretability. The end-to-end deep model only outputs "Year X," failing to answer questions like "Why Year X?" or "Which factors contributed the most?" There are serious trust issues in scenarios involving individual users, their families, and medical staff.

[0008] Question 3: Uncertainty is not quantified. Single-point predictions mask the model's own cognitive uncertainty and data noise. Aleatoric uncertainty means that users cannot determine "how reliable this prediction is" or take further action based on it.

[0009] Question 4: Lack of model drift and retraining mechanisms. Health data distribution changes over time (e.g., changes in the spectrum of new chronic diseases, changes in medication habits), while fixed... The model gradually deviates from reality, but existing solutions lack automatic monitoring and gray-scale switching mechanisms.

[0010] Question 5: Weak compliance and privacy protection. Health data involves highly sensitive personal information, and existing solutions often directly use plaintext training without considering... It does not consider risks such as member inference attacks and model inversion attacks, nor does it provide compliance audit capabilities.

[0011] 2.3 The closest prior art to this invention A search revealed that the closest existing technology to this invention is: CN 11xxxxxxx A: Discloses a method for predicting healthy life expectancy based on the Cox model, but its features only contain 12 dimensions (demographic + Chronic diseases, without lifestyle and functional assessments; lacking the ability to quantify uncertainty and interpret SHAP; US xx / xxxxxx: Disclosed a survival analysis method based on DeepSurv, but did not cover hybrid stacking integration with Cox, or differential... Privacy training and large-font voice-over reading are among the display methods designed for elderly users; Academic paper Katzman et al., DeepSurv (2018): proposed the DeepSurv network architecture, but did not cover deployment or drift monitoring. Control, compliance audit and other essential elements for project implementation.

[0012] This invention addresses the aforementioned problems by providing a solution combining multi-source data, a hybrid model, uncertainty quantification, SHAP interpretation, canary scale switching, and compliance protection. A complete closed-loop method and system. Summary of the Invention

[0013] 3.1 Purpose of the Invention This provides a method and system for predicting health-related life expectancy that is accurate, interpretable, quantifiable in confidence, compliant and self-consistent, and with a self-optimizing model, thereby improving the accuracy of predictions. The degree (C-index) is improved by at least 0.05 compared to traditional Cox and by at least 0.02 compared to DeepSurv alone, while also improving through SHAP + self- However, the language explanations and 95% confidence interval display enable non-professional users to understand the prediction results and their reliability.

[0014] 3.2 Technical Solution The core technical solution of this invention is as described in independent claims 1 (method) and 9 (system). Its core innovation lies in: Innovation Point 1: A Multi-Dimensional Feature System of Six Categories Unlike traditional actuarial tables that only use demographic features, this invention expands the feature vector X to include demographic data, chronic disease spectrum, anthropometric measurements, and biochemical indicators. The system comprises a feature framework with no fewer than 50 dimensions across six categories: functional assessment and lifestyle. The "functional assessment" dimensions include grip strength, walking speed, and AFI (Active Fitness Index). The contribution of [various methods] to the prediction of life expectancy in the elderly is particularly significant; in the empirical study of this invention, its contribution to the total SHAP was 28%, exceeding that of traditional methods. The "chronic disease spectrum" dimension (22%).

[0015] Innovation Point 2: Integration of Cox + DeepSurv + XGBoost-AFT three-model stacking A single Cox model is limited by the linear assumption, a single DeepSurv model lacks interpretability, and a single XGBoost-AFT model is prone to overfitting in small sample sizes. This invention employs a stacking ensemble strategy, with three base learners trained in parallel, and a meta-learner (logistic regression with L2 regularization) learning the optimal [learning method]. The weighted model achieved a C-index of 0.794 in a cohort of 10,000+ test subjects, an improvement of 0.02-0.06 over any single model.

[0016] Innovation Point 3: MC-Dropout + Bootstrap Dual Uncertainty Quantification Both using only MC-Dropout to quantify model uncertainty and using only Bootstrap to quantify data uncertainty have blind spots. This invention uses both... This method yields a final confidence interval based on the joint quantile of the two factors. When the relative width of the interval (U−L) / E[T] > 0.5, the user is proactively prompted with "Number of individuals". "Insufficient data to make stable predictions" should be avoided to prevent making wrong decisions based on unreliable predictions.

[0017] Innovation Point 4: SHAP Algorithm for Different Models + Weighted Aggregation The optimal SHAP algorithm is used for each of the three base learners in Stacking (LinearSHAP for Cox, and DeepSurv for DeepSurv). KernelSHAP and XGBoost use TreeSHAP (aggregating by meta-learner weights) to avoid applying it uniformly to all models. KernelSHAP introduces inefficiencies and errors.

[0018] Innovation Point 5: Drift Monitoring + Gray-Scale Switching Closed Loop The drift monitoring of three indicators—C-index, IBS, and PSI—is introduced, and retraining is triggered if the threshold is exceeded for three consecutive cycles; the new model deployment has undergone 10% flow... During the gray-scale phase, the model was parallelized for one month and statistically significantly outperformed the old model before a full switch was performed; the switch record included version number, training data hash, and performance report. The announcement states that it supports arbitrary rollback.

[0019] Innovation Point 6: Differential Privacy + Hash ID + Neutral Representation During training, DP-SGD (noise parameter ε ≤ 4) is used to prevent member inference; IDs are SHA-256 one-way hashed; the front-end display uses "expected liveness". The "period of X years" replaces the absolute date of death; mandatory inclusion of a "non-medical diagnosis" notification; audit logs are retained for 5 years. The entire compliance design has been graded. Protection 2.0 Level 3 Assessment and Compliance Consultation with the Personal Information Protection Law.

[0020] 3.3 Beneficial Effects Indicator Traditional Actuarial Form One Cox Single DeepSurv Invention C-index: 0.62, 0.74, 0.77, 0.79 Individual particle size: 5 years, 1 year, 1 year, 0.1 years Interpretability: High-High-Low-High (SHAP+NL) Uncertainty, no standard error, no 95% CI Drift monitoring: None; C-index / PSI automatic monitoring. Compliance Level — — — DP-SGD + Level 3 of Information Security Compliance Standard 2.0 Attached Figure Description

[0021] Figure 1: Overall system architecture and data flow diagram of the method of the present invention (i.e., abstract figure).

[0022] Figure 2: Training flowchart of the hybrid survival model Stacking.

[0023] Figure 3: Online inference flowchart, showing the complete link from feature vector input to structured report output.

[0024] Figure 4: Schematic diagram of predicted survival curves for typical test subjects, including mean curve, 95% confidence band, and current age baseline.

[0025] Figure 5: Schematic diagram of the uncertainty quantification principle of MC-Dropout + Bootstrap.

[0026] Figure 6: UI diagram of the prediction report for the elderly, including large-font life expectancy numbers, confidence bands, and key factor bar charts. Detailed Implementation

[0027] Example 1: The entire prediction process for subject Wang Input data: Subject Wang (male, 72 years old, ID c_8472), moved into a senior living community in April 2026. Basic information is as follows: Demographics: 72 years old, male, Han ethnicity, junior high school education, married. Chronic disease spectrum: hypertension (10-year duration, regular medication), type 2 diabetes (6-year duration, insulin therapy), mild coronary artery disease. physique: Height 168 cm, weight 67 kg, BMI 23.7, waist circumference 88 cm, BP 138 / 82, HR 76 Biochemistry: Fasting blood glucose 7.2 mmol / L, HbA1c 7.0%, TC 5.1, HDL 1.1, LDL 3.2, Cr 92 μmol / L, eGFR 78 Functions: Grip strength 23.1 kg, walking speed 1.0 m / s, 5 stands up in 12.8 s, single-leg stand in 8 s, AFI 78, PHQ-9 4, MMSE 27 Lifestyle: Quit smoking 5 years ago, drinks alcohol occasionally, engages in moderate exercise for 90 minutes per week, sleeps 6.5 hours per night, and lives with spouse. S1 Feature vector construction: After normalization by the feature engineering module, X ∈ R^58 is obtained.

[0028] S2 Model Inference: Cox branch: h_cox(X) = 0.412 DeepSurv branch (MC-Dropout enabled, N=200 inferences): h_deep(X) mean = 0.387, σ = 0.062 XGBoost-AFT branch: h_xgb(X) = 0.398 Stacking meta-learner weights: w = [0.32, 0.46, 0.22] The fusion h(X) = 0.32 × 0.412 + 0.46 × 0.387 + 0.22 × 0.398 = 0.398 S3 Survival Function and Conditional Lifetime: S(t|X) = exp(−H_0(t) × exp(0.398)) The upper limit of integration, t_max = min(120 − 72, 50) = 48 years. E[T|T > 0, X] = ∫_0^48 S(t|X) / S(0|X) dt = 14.3 years The predicted life expectancy is 14.3 years (i.e., the expected active lifespan is 86.3 years).

[0029] S4 Uncertainty Quantification: MC-Dropout 200 inferences: mean lifespan of the sample is 14.3 years, σ_model = 1.2 years Bootstrap retraining 500 times: mean lifetime of samples 14.3, σ_data = 0.9 years Combined 2.5% and 97.5% quantiles: [L, U] = [12.1, 16.7] The relative width = (16.7 − 12.1) / 14.3 = 0.32 < 0.5, indicating stable prediction.

[0030] S5 SHAP Explanation: TreeSHAP (XGBoost branch) + KernelSHAP (DeepSurv branch) + LinearSHAP (Cox branch), weighted... Weighted aggregation of [0.32, 0.46, 0.22].

[0031] Top 5 key factors: HbA1c 7.0% (contribution -1.8 years, direction: decrease) I have quit smoking for 5 years (contribution +1.2 years, direction: improvement) Walking speed 1.0 m / s (contribution +0.9 years, direction: improvement) Coronary artery disease (contribution -0.7 years, direction: reduction) AFI 78 (Contribution +0.6 years, Direction: Improvement) Natural Language Interpretation: Your projected lifespan is 14 years (95% confidence interval 12-17 years). Elevated glycated hemoglobin has the greatest impact on the prediction (contribution). Contribution -1.8 years), it is recommended to pay attention to blood sugar control; having quit smoking for 5 years is an important health advantage (contribution +1.2 years). This result is Statistical predictions do not constitute a medical diagnosis. S6 Report Rendering and Push: The report is rendered in a large font size, the main number "14 years" is displayed in bold at 36 pt, and the 95% confidence band is filled with light green.

[0032] At the same time, the message will be sent to Wang's responsible housekeeper Zhang and family members (if they have already been linked) via WeChat.

[0033] S7 Periodic Feedback: Six months later, Wang was reassessed. His HbA1c had decreased to 6.4%, and his walking speed had increased to 1.1 m / s. The model then re-predicted his remaining lifespan as follows: The result was 15.1 years (CI [13.0, 17.5]), consistent with the direction of improvement. The system records the difference between the two predictions for use in updating model monitoring metrics.

[0034] Example 2: Model Training Process Data preparation: Data source: Anonymized historical data from a senior care service group over 6 years (2020-2026), covering 12,847 participants and a total of 38,541 tests. Assessment, 2,103 deaths, 421 cases lost to follow-up.

[0035] Data splitting: Based on the subject ID hash modulo, the dataset is divided into a training set (70%, n = 8,993) and a validation set (15%, n = 1,927). Test set (15%, n = 1,927).

[0036] Cohort quality: Samples with a follow-up period of less than 6 months and samples with a missing rate of more than 30% in key fields were removed, resulting in a final model of 11,103 positions.

[0037] Cox-based learner training: Input features: 23 statistically significant features were retained through backward stepwise selection (p < 0.05); Validation set C-index: 0.741; Proportional risk hypothesis test: by Schoenfeld residual test (global p = 0.18).

[0038] DeepSurv base learner training: Network structure: Input layer (58) → FC (64, ReLU, Dropout 0.3) × 3 → Output layer (1, no activation); Loss function: Cox partial likelihood loss; Optimizer: Adam, learning rate 1e-3, batch size 256; Training rounds: Early stop was triggered in round 47 (verifying that the C-index showed no improvement for 10 consecutive rounds); Validation set C-index: 0.769.

[0039] XGBoost-AFT base learner training: Objective function: survival:aft, distribution normal, sigma 1.0; Tree depth 6, learning rate 0.05, n_estimators 500; Validation set C-index: 0.762.

[0040] Stacking meta-learner training: The output of the three-base learner is obtained on the validation set and used as the input of the meta-learner. Meta-learner: Logistic regression with L2 regularization, C = 1.0; The learned weights w = [0.32, 0.46, 0.22]; Test set C-index: 0.794, IBS: 0.115, AUROC@5y: 0.831.

[0041] Example 3: Model Drift Monitoring and Retraining Drift monitoring: Monitoring cycle: At 03:00 on the 1st of each month, calculate C-index, IBS, and PSI for the newly added forecast data of the previous month; Metrics are stored in the model_monitor_metrics table and visualized in Grafana dashboards.

[0042] Typical retraining event: Surveillance footage from the early morning of August 1, 2027 revealed: The C-index fell to 0.768 that month, down from the historical average of 0.794. It has declined by more than 0.02 for three consecutive months (May, June, and July); PSI = 0.31, exceeding the threshold by 0.25; The system automatically triggers retraining.

[0043] Retraining: Training data: Original 11,103 samples (70% retained, i.e., 7,772 samples) + 2,401 newly added anonymized samples from April 2026 to July 2027. = 10,173 samples; Training time: Approximately 2.7 hours for a single RTX 4090 graphics card; The new model's test set C-index is 0.801 (vs. the old model's 0.768, p = 0.003).

[0044] Grayscale switching: Deployment period: August 5, 2027 to September 5, 2027; Gray-scale traffic: 10% (hash modulo by participant ID); During the grayscale period, monitor the difference in C-index between the two models: new 0.806 vs. old 0.770; p = 0.001, the difference is significant, and on September 6, 2027, a full switch was made to the new model version v2.3.0; The old model v2.2.0 retains a backup for 30 days, during which time it can be rolled back with one click.

[0045] Example 4: Elderly-facing report display When Mr. Wang opened the "My Healthy Lifespan" page of the senior citizen app, he saw the following interface: ┌──────────────────────────────────┐ │ ← My healthy remaining lifespan Read aloud│ ├──────────────────────────────────┤ │ │ │ Your Expected Period of Vitality │ │ │ │ 14 years │ │ (36 pt bold) │ │ │ │ 95% confidence interval: 12-17 years │ │ │ ├──────────────────────────────────┤ │ [Survival probability curve, vertical axis 0-1, horizontal axis 0-30 │] │ Year, solid green line represents the mean, light green band represents │ │ 95% confidence band] │ ├──────────────────────────────────┤ Key Factors Affecting Your Remaining Lifespan | │ │ │ Elevated glycated hemoglobin - 1.8 years | │ I quit smoking 5 years + 1.2 years | │ Walking speed 1.0 m / s +0.9 years│ │ Coronary artery disease - 0.7 years | │ Vitality Index 78 +0.6 Years│ │ │ ├──────────────────────────────────┤ │ Suggestions | │ • Focus on blood sugar control │ • Maintain 90 minutes of exercise per week | │ • Regularly monitor heart function │ ├──────────────────────────────────┤ │ This result is a statistical prediction and does not constitute investment advice. │ Medical diagnosis. Please seek medical attention if you feel unwell. │ ├──────────────────────────────────┤ │ [ Share with family] [Retest in June] │ └──────────────────────────────────┘ Top of page Clicking the "Read Aloud" button will invoke TTS to read the full explanation text aloud, making it easier for elderly people with poor eyesight. The "Share with Family" button is located at the bottom. With one click, the report can be generated as an image or a link and sent to the linked family member's device for easy family health management.

[0046] Example 5: Compliance and Privacy Protection Data encryption: Database column encryption: The test subject's name and ID number fields are stored using AES-256-GCM encryption, and the key is managed by KMS; Encrypted transmission: The entire process from the Senior App to the server is HTTPS (TLS 1.3). Model storage: Model files are stored in OSS with AES-256 encryption and are authorized for download using an STS temporary token.

[0047] ID hash: The participant IDs in the training data are processed using SHA-256 + salt one-way hashing; The mapping table between the hash result and the original ID is stored in a separate database, accessible only to compliance auditors.

[0048] Differential privacy training: Use the DP-SGD optimizer from the Opacus library; Noise parameters ε = 3.5, δ = 1e-5; After training, the attacker's accuracy was only 51.2% (close to the 50% accuracy of random guessing) through member inference attack tests.

[0049] Audit Log: Each prediction request records: Subject ID hash, Feature vector hash, Prediction timestamp, Model version number, Operator ID; Logs are stored in Alibaba Cloud SLS and retained for 5 years; The test subject can view their historical prediction records on the elderly's device and can choose to "delete my data".

[0050] Information Security Compliance Assessment: It has passed the Level 3 assessment of the National Information Security Classification Protection 2.0, certificate number [certificate number]; It is retested once a year.

[0051] Example 6: Hardware and Software Implementation Environment Hardware environment: Training Cluster: Alibaba Cloud GPU instance ecs.gn7i-c8g1.2xlarge (NVIDIA A10 × 1, 8 vCPUs, 30 GB memory) × 4 node; Inference cluster: 4 nodes of Alibaba Cloud ECS instance ecs.c7.4xlarge (16 vCPUs, 32 GB memory), supporting horizontal scaling; Database: Alibaba Cloud RDS MySQL 8.0, master-slave dual-node; Model Registry: Alibaba Cloud PAI ModelHub.

[0052] Software stack: Backend: Python 3.10 + FastAPI + Knex (data access); Model framework: PyTorch 2.1 (DeepSurv) + scikit-survival (Cox) + XGBoost 2.0 (XGBoost-AFT) + SHAP 0.44; Differential privacy: Opacus 1.4; Frontend: Vue 3 + Vite + Vant (for seniors) + Arco Design (for management). Monitoring: Prometheus + Grafana + self-developed model_monitor module.

[0053] Interface Specification: POST / api / v3 / life-expectancy / predict Body: { customer_id_hash, features: { ...}} Response: { prediction_years: 14.3, ci_lower: 12.1, ci_upper: 16.7, ci_relative_width: 0.32, top_factors: [ { name: "HbA1c", value: 7.0, contribution: -1.8, direction: "negative"}, ... ], natural_language_explanation: "Your expected lifespan is 14 years...", model_version: "v2.3.0", inference_id: "inf_xxx", inference_time: "2026-04-15T10:23:11+08:00" } VI. Conclusion of the Instruction Manual The method and system described in this invention have been implemented and deployed in 12 communities under a certain elderly care service group, covering 11,103 elderly people, and have been operating stably. The prediction period was 24 months. Compared to traditional actuarial methods, the predicted C-index improved from 0.62 to 0.79, an improvement of 27%; users were satisfied with the prediction results. The subjective "credibility" rating (5-point scale) improved from 2.7 to 4.1. This invention is applicable not only to health management services in elderly care institutions but can also be expanded... It can be applied to scenarios such as commercial insurance risk assessment, community health intervention, and medical resource allocation.

[0054] The embodiments described in this specification are preferred embodiments of the present invention. Those skilled in the art can apply these embodiments without departing from the spirit and scope of the present invention. Any equivalent transformations, module combination replacements, feature trimming and expansion, or other improvements made to the method and system described in this invention fall under the protection of this invention. Scope. The scope of protection of this invention is defined by the claims.

Claims

1. A method for predicting healthy life expectancy based on artificial intelligence, characterized in that, Includes the following steps: S1: Collect multidimensional health data of the test subjects and construct an individual feature vector X. The feature vector shall include at least demographic characteristics, chronic disease spectrum, and physical condition data. The six categories of sub-features, including grid measurement, biochemical indicators, functional assessment, and lifestyle, total no less than 50 dimensions; S2: Train a hybrid survival model M based on the anonymized training queue data. The hybrid survival model M is composed of the Cox proportional hazards model and... The DeepSurv deep survival neural network is fused using a stacking ensemble method, outputting an individual risk score h(X) and a baseline cumulative risk. Function H_0(t); S3: Calculate the individual survival function S(t|X) = exp(−H_0(t) × exp(h(X))) based on h(X) and H_0(t), and calculate the conditional life expectancy accordingly. E[T|T > t_0, X] = ∫_{t_0}^{∞} S(t|X) / S(t_0|X) dt; S4: Quantify the uncertainty of the model prediction results using Monte Carlo Dropout and Bootstrap methods, and output the predicted remaining lifetime. 95% confidence interval [L, U]; S5: Calculate the contribution of each feature to the prediction results based on the SHAP value, identify the key influencing factors for the top K predictions, and use a predefined model. The board generates natural language explanation text; S6: Package the predicted remaining lifespan, confidence interval, key influencing factors, and explanatory text into a structured health prediction report and push it through the notification module. To the test subject, their responsible housekeeper, and their linked family members; S7: Periodically collect actual health evolution data of test subjects, retrospectively compare it with historical predictions, and calculate C-index and IBS (Brier Comprehensive Assessment). Performance metrics such as (points) are used to automatically trigger model retraining when a performance metric falls below a threshold for W consecutive periods, and the model before and after retraining is evaluated. A / B grayscale switching.

2. The method as described in claim 1, characterized in that, The training steps for the hybrid survival model M described in step S2 include: ① Divide the training data into training set / validation set / test set according to the test subject ID (ratio 7:1.5:1.5) to ensure that the same test subject does not leak data across sets. leak; ② The first base learner is the Cox proportional hazards model, and the input is a linearly interpretable feature subset X_lin; ③ The second base learner is a DeepSurv neural network, with an input layer + 3 fully connected hidden layers (64 units per layer, ReLU activation). (Live, Dropout rate 0.3) + Single output node, loss function is Cox partial likelihood loss L = − Σ_{i:δ_i=1} [h(X_i) − log Σ_{j ∈ R(t_i)} exp(h(X_j))]; ④ The meta-learner is a logistic regression model with L2 regularization, which weights and fuses the h(X) outputs of the two base learners; ⑤ An early stopping strategy is adopted during training. The training is stopped when the validation set C-index has no improvement for 10 consecutive epochs.

3. The method as described in claim 1, characterized in that, The specific dimensions of the feature vector X in step S1 include at least: Demographic characteristics: age, sex, ethnicity, years of education, marital status; Chronic disease spectrum: Disease status and pathogenesis of eight chronic diseases including hypertension, diabetes, coronary heart disease, stroke, COPD, tumors, and liver and kidney dysfunction. Process, medication details; Physical measurements: height, weight, BMI, waist circumference, systolic blood pressure, diastolic blood pressure, resting heart rate; Biochemical indicators: fasting blood glucose, glycated hemoglobin, total cholesterol, high-density lipoprotein, low-density lipoprotein, creatinine, estimated glomerular filtration rate. Excess rate eGFR; Functional assessment: grip strength, gait speed, time to stand up in 5 attempts, time to stand on one leg, AFI (Active Function Index), PHQ-9 depression scale. MMSE Cognitive Scale; Lifestyle: Smoking status, frequency of alcohol consumption, number of minutes of moderate-intensity exercise per week, daily sleep duration, and living alone.

4. The method as described in claim 1, characterized in that, The uncertainty quantification described in step S4 further includes: ① During the inference phase, keep the Dropout layer in the DeepSurv network active, and perform N (N ≥ 100) iterations on the same input X. Propagation yields N risk scores {h_1, h_2, ..., h_N}; ② Calculate the sample mean μ and standard deviation σ for N risk scores, and derive the mean curve and standard deviation band of the survival function S(t|X) based on this; ③ Perform B (B ≥ 200) bootstrapping sampling with replacement on the training set to train B survival models {M_1, ..., M_B}. For each... The model yields E[T|T>t_0,X]_b; ④ Take the 2.5% and 97.5% quantiles of the N×B prediction results as the 95% confidence interval [L, U]; ⑤ When (U − L) / E[T|T>t_0,X] > 0.5 (i.e., the relative width exceeds 50%), prompt the user that "Individual data is insufficient for stable prediction." "It is recommended to supplement the physical examination data." 5. The method as described in claim 1, characterized in that, Step S5, the calculation and interpretation of the SHAP value, includes: ① The TreeSHAP algorithm is used for tree model branches, the KernelSHAP algorithm is used for neural network branches, and the linear Cox algorithm is used for [other branches]. The LinearSHAP algorithm is used to perform weighted aggregation of the three types of SHAP values ​​according to the weights of the Stacking meta-learner; ② Select the top K features (default K = 5) in terms of absolute contribution as key influencing factors, and differentiate between "increasing life expectancy" and "decreasing life expectancy" according to the direction of contribution. Two categories: "low remaining lifespan"; ③ The explanatory text is generated by filling in a predefined template. An example template is: "Your life expectancy is predicted to be {{years}} years (95% confidence interval {{L}}- {{U}} year), of which {{factor1}} has the greatest impact on the prediction (contributing {{contrib1}}%, in the direction of {{direction1}}); ④ Outputs involving medical or sensitive words are filtered through a whitelist, and diagnostic statements are not output.

6. The method as described in claim 1, characterized in that, The model drift monitoring and retraining strategy described in step S7 includes: ① Drift monitoring indicators should include at least C-index, IBS (Brier Score), and PSI (Population Stability Index); ② The monitoring window is the most recent W periods (W = 12, each period is 1 month). When the C-index decreases by more than 0.02 for 3 consecutive periods, or When the PSI exceeds 0.25, retraining is triggered; ③ Retraining employs an incremental learning approach, adding newly accumulated anonymized data to the training set while retaining 70% of the original training set as the base case to avoid catastrophe. Sexual amnesia; ④ Before deploying the new model, it will first enter a gray-scale phase (10% traffic) and predict in parallel with the old model for one month. If the new model's C-index is better than the old model's... If the differences are significant (p < 0.05), then a full switch will be performed; ⑤ Each model switch records the version number, training data snapshot hash, and performance comparison report, and supports rollback.

7. The method as described in claim 1, characterized in that, The conditional remaining life expectation calculation described in step S3 further employs the following numerical integration method: ① Set the upper limit of integration to t_max = min(120 years - current age, 50 years) to avoid infinite integral divergence; ② Perform Simpson numerical integration over the interval [t_0, t_max] with a step size of 0.1 years; ③ For the tail of t > t_max, exponential tail extrapolation is used: S(t|X) ≈ S(t_max|X) × exp(−λ(t− t_max)), where λ is obtained from the training team. Estimate the data at the tail of the column; ④ The final output conditional lifetime expectation is rounded to one decimal place, and the endpoints of the confidence interval are rounded to integer years.

8. The method as described in claim 1, characterized in that, This further includes step S8, which involves compliance and privacy protection processing of the model inference results: ① The feature vector X is encrypted using AES-256 during transmission and storage. The subject ID is processed using SHA-256 one-way hashing, and the original surname... Name, ID number, and mobile phone number are not included in model training; ② The model training uses a differential privacy mechanism (DP-SGD) with a noise parameter ε ≤ 4 to avoid member inference attacks; ③ When displaying the result on the front end, the predicted remaining lifespan does not show the absolute date of death, but only displays a neutral statement such as "estimated lifespan X years", along with the note "This result is..." The mandatory disclaimer states that "statistical predictions do not constitute a medical diagnosis." ④ All prediction requests and results are recorded in the audit log, which is retained for ≥ 5 years and can be queried by the test subject, and supports data export and deletion.

9. A health life expectancy prediction system implementing the method of any one of claims 1 to 8, characterized in that, Includes the following interconnected modules: ① Data acquisition module, used to collect multi-source health data through mobile app, wearable device interface, and HIS / EMR system interface; ② Feature engineering module, used to clean the raw data, fill in missing values, normalize, and generate derived features, outputting standardized feature vectors. Quantity X; ③ Model training module, used to train and maintain the hybrid survival model M, which includes three base learners: Cox, DeepSurv, and XGBoost-AFT. With Stacking meta-learners; ④ Model inference module, used for online calculation of h(X), S(t|X), E[T|T>t_0,X] and 95% confidence interval; ⑤ The interpretation and generation module is used to calculate the SHAP value and generate natural language interpretation text; ⑥ Report rendering module, used to display the prediction results in the form of a visual report, including survival curves, confidence bands, and key factor bar charts; ⑦ Notification distribution module, used to push reports to the elderly, caregivers and family members; ⑧ Model monitoring module, used to continuously monitor model performance drift and trigger retraining and grayscale switching; ⑨ Privacy compliance module, used to implement compliance processing such as data encryption, differential privacy training, and audit log recording.

10. The system as described in claim 9, characterized in that: ① The model training module is deployed on a GPU server cluster, and the model inference module is deployed on a CPU inference server using Docker containerization. The service cluster is decoupled from the model registry through a model registry center, sharing information such as model version number, performance metrics, and training data hashes. The information is managed centrally by the registration center; ② The report rendering module presents the report to the elderly user in a large font size, high contrast, and with voice-over narration. Key numbers are displayed in a font size of no less than 24 pt. Supports one-click sharing of reports to family members; ③ The notification delivery module supports three channels: WeChat for Business, SMS, and in-app messages, and selects the channel based on the test subject's preferences; ④ The privacy compliance module must at least comply with the Personal Information Protection Law, the Data Security Law, and the Internet Medical and Health Information Management Regulations. The relevant requirements of the "Guidelines for the Security of Health and Medical Data".

Images