Decentralised Federated Learning Framework for Privacy-Preserving Medical Insurance and Sustainable Rural Economic Development in Tamil Nadu

Technical Summary: Decentralised Federated Learning Framework for Privacy-Preserving Medical Insurance

Problem Statement
Rural medical-insurance entrepreneurship in India, particularly in Tamil Nadu, faces a critical bottleneck: fragmented and siloed health data. Insurance agents, primary health centers, and cooperative micro-insurers possess client data but cannot legally or practically pool it due to privacy concerns and regulatory constraints. This data isolation prevents the training of accurate, equitable risk-assessment models, hindering the development of rural Insurtech. While Federated Learning (FL) offers a theoretical solution by enabling collaborative model training without centralizing raw data, existing implementations often fail to address the specific trade-offs between privacy guarantees, model accuracy, and the heterogeneity of rural data distributions. Furthermore, previous claims of "cost-free" equivalence between federated and centralized models have been challenged regarding the rigor of privacy accounting and statistical validation.

Methodology
The paper proposes and evaluates TW-DP-FedAvg, a federated architecture that integrates gradient-clipped Gaussian Differential Privacy (DP) with a trust-weighted client aggregation rule. The study addresses three specific methodological gaps identified in prior work:

1. Rigorous Privacy Accounting: Instead of reporting uncomposed per-round privacy parameters, the authors fix a total privacy budget ($\epsilon_{total} = 8$) across 40 communication rounds. Using the Advanced Composition Theorem, they derive a stricter per-round budget ($\epsilon_0 \approx 0.207$) to ensure the cumulative privacy guarantee is defensible.
2. Trust-Weighted Aggregation: The framework weights client updates based on a trust metric ($trust_c = 1/(1 + BCE_c)$), where $BCE_c$ is the local binary cross-entropy loss, multiplied by the local sample size.
3. Experimental Design:
   • Datasets: Due to the lack of real rural Tamil Nadu insurance records, the study utilizes two public benchmarks: the Medical Cost Personal Dataset ($n=1,338$) and the Pima Indians Diabetes Dataset ($n=768$).
   • Partitioning: Data is partitioned into five simulated district shards using a Dirichlet distribution ($\alpha$) to model non-IID (heterogeneous) data, replacing previous ad-hoc splits.
   • Statistical Validation: The study employs Two One-Sided Tests (TOST) for equivalence testing with bounds of $\pm 3$ accuracy percentage points, alongside paired t-tests, Wilcoxon signed-rank tests, and Cohen's $d$ effect sizes.
   • Ablation: A $2 \times 2$ ablation study isolates the independent effects of the DP mechanism and the trust-weighting rule.

Key Results
The revised analysis yields more conservative and statistically robust conclusions compared to the original submission:

• Accuracy vs. Privacy Trade-off: Under the corrected, composed privacy budget, the federated model incurs a measurable accuracy cost compared to a centralized baseline.
  • On the Medical Cost dataset, TW-DP-FedAvg achieved 81.7% accuracy versus 88.8% for the centralized baseline (a 7.1% gap). The difference is statistically significant ($p=0.046$), and TOST equivalence testing failed to establish that the federated model is within the $\pm 3\%$ threshold of the centralized model ($p=0.893$).
  • On the Pima Diabetes dataset, the accuracy gap was negligible (71.1% vs. 71.4%), but the F1-score for the minority class dropped significantly (0.366 vs. 0.612) under the tighter privacy budget, indicating a negative impact on recall.
• Ablation Findings: The accuracy penalty is primarily driven by the Differential Privacy mechanism (noise injection), not the trust-weighting rule. The trust-weighted aggregation performed comparably to standard sample-size weighting in terms of raw accuracy.
• Heterogeneity Impact: The proposed Ecosystem Inclusion Score (EIS) demonstrated a monotonic decrease as data heterogeneity increased (Dirichlet $\alpha$ decreased). EIS dropped from 0.970 (low heterogeneity) to 0.868 (high heterogeneity) on the Medical Cost benchmark, confirming the metric's sensitivity to non-IID data structures.

Significance and Claims
The paper positions TW-DP-FedAvg not as a perfect, cost-free replacement for centralized analytics, but as a viable ecosystem-level infrastructure that requires a conscious trade-off.

• Revised Contribution: The study explicitly retracts claims of "equivalence" between federated and centralized models under a defensible privacy budget. Instead, it establishes that federated training comes at a measurable accuracy cost.
• Policy and Entrepreneurial Context: The findings are grounded in India's Digital Personal Data Protection Act (DPDPA) 2023 and the IRDAI regulatory sandbox. The authors argue that while the technology offers a pathway for rural Insurtech to collaborate without centralizing sensitive data, stakeholders (regulators and entrepreneurs) must acknowledge the accuracy penalty.
• Limitations: The authors emphasize that the study is a technical proof-of-concept using proxy datasets. It does not validate unit economics, venture viability, or field deployment with actual rural intermediaries. The discussion on entrepreneurial ecosystems is a conceptual extrapolation, noting that infrastructure alone is insufficient to catalyze ecosystem emergence without addressing financing, mentoring, and distribution channels.

In summary, the paper provides a rigorous, corrected evaluation of a privacy-preserving federated learning framework, demonstrating that while it enables data collaboration in fragmented rural markets, it does so at the expense of predictive performance, a trade-off that must be managed by future regulatory and market participants.