Tiered Desensitization and Lightweight Fusion of Multi-Source Heterogeneous Data for Privacy-Preserving Personalized Health Management in Community Elderly Care: A Case Study of the "Hui-YangLao" Platform in Chengdu, China

Background: Community-based smart elderly-care platforms continuously aggregate wearable, primary-care, and community-service data about older adults. These data are highly sensitive, yet their granularity is exactly what personalized chronic-disease management and home-safety warning require. Resource-constrained county and community platforms, in particular, struggle to reconcile privacy compliance, computational cost, and service accuracy. Methods: We report a 10-month prospective deployment study on the "Hui-YangLao" platform in Pidu District, Chengdu, China, covering 1,267 older adults across six urban communities and three townships. We designed and operationalized a tiered desensitization scheme (high / medium / low privacy risk) coupled with a three-stage lightweight fusion pipeline (data-level cleaning, feature-level weighted fusion, decision-level cross fusion) and a four-tier role-based access control. We evaluated privacy protection, data utility, and computational overhead before and after the intervention, using platform-monitored re-identification risk, risk-identification accuracy, and server CPU/memory consumption. Results: After desensitization, the composite identity re-identification risk score dropped from 9.2/10 to 1.1/10 (88.0% reduction); geolocation re-identification probability fell from 100% to 0.7%; the coefficient of determination for chronic-disease trend fitting decreased only marginally (R² = 0.961 to 0.948). The chronic-risk identification accuracy dropped from 92.36% to 90.82% (1.54 percentage points) while the share of sensitive fields in cloud-stored records fell from 41.7% to 0%. Per-batch computational overhead decreased by 61.7%, allowing the pipeline to run on a low-cost community cloud server. Across the 10-month pilot, 17 unauthorized-access alerts were logged and closed within 24 hours, and no confirmed exfiltration of identifiable elderly health records occurred. Conclusions: For resource-constrained community elderly-care platforms, coupling tiered desensitization with a lightweight three-stage fusion pipeline can substantially reduce privacy exposure without materially degrading the accuracy of personalized health warnings. The scheme offers a low-cost, reproducible reference implementation for other mid-sized Chinese community platforms.