Research map / Probabilistic and scientific modeling / Molecules and proteins
IBCA: An Intelligent Platform for Social Insurance Benefit Qualification Status Assessment
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Reports the deployed IBCA platform for verifying retiree survival status in Shandong, China, built around an MVI-TSP survival predictor (improved bidirectional GRU with missing-value interpolation) plus VAE/VRNN and MC Dropout uncertainty scoring.
Problem
Verifying the survival and benefit qualification status of retirees to prevent pension over-payment and social security fraud, where manual data-comparison certification is inaccurate, slow, and suffers from missing and inconsistent government data.
Approach
Combines a GRU-based survival predictor (GRU-B with mask matrix and time decay for missing-value interpolation, plus 1D dilated convolution and bidirectional RNN for local/global features) with uncertainty estimation via a VRNN/VAE for imputation and MC Dropout for prediction, deployed as an end-to-end platform.
Key results
- MVI-TSP achieved AUROC 0.9841 and AUPRC 0.9158 on the offline test, outperforming GRU-D (0.9772 / 0.8748).
- In deployment, the one-month hospitalized-cohort analysis reported recall of 68.8% and accuracy of 89.6%; the non-hospitalized cohort reported recall 75.0% and accuracy 99.3%.
- Deployed since November 2019 across 12 Shandong cities, handling over 50 TB of data and generating cumulative revenue of RMB 52.637 million.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVI-TSP instead of GRU-D | Survival prediction · Government dataset test set | AUROC: 0.69% | Helped |
| Use MVI-TSP instead of GRU-D | Survival prediction · Government dataset test set | AUPRC: 4.1% | Helped |
Surprising findings
- The platform was designed for efficiency at the expense of data privacy protection, according to the authors' deployment lessons.
Limitations
- Dataset is available only upon request under terms of use, limiting reproducibility.
- Privacy protection was sacrificed for efficiency; federated learning is proposed as future work.
- Deployment accuracy is reported only from a preliminary analysis with small death counts (1,137 and 679) and no independent ground-truth audit.
Experimental setup
Datasets: Shandong government dataset (one-year span, personal, medical insurance, civil affairs, health records)
Benchmarks: Shandong government survival dataset (one-year span)
Metrics: AUROC AUPRC recall accuracy
Baselines: Logistic Regression RetainEX AdaCare Diople GRU-D
Builds on: GRU-D Variational Recurrent Neural Network MC Dropout AdaCare Diople
Closest papers by meaning
- The Value of Prediction in Identifying the Worst-Off (ICML 2025 · similarity 0.75)
- A Submodular Optimization Approach to Accountable Loan Approval (AAAI 2024 · similarity 0.74)
- AI-Driven Personalized Fall Prevention for Older Adults (AAAI 2025 · similarity 0.74)
- A Novel Characterization of the Population Area Under the Risk Coverage Curve (AURC) and Rates of Finite Sample Estimators (ICML 2025 · similarity 0.74)
- Trustworthy and Practical AI for Healthcare: A Guided Deferral System with Large Language Models (AAAI 2025 · similarity 0.74)
- Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.73)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.