JEEVHITAA -- An End-to-End HCAI System to Support Collective Care

Current mobile health platforms are predominantly individual-centric and lack the support for coordinated, auditable multi-actor workflows. However, in many settings worldwide, health decisions are enacted by multi-actor care networks rather than single users. We present JEEVHITAA, a cross-platform mobile system enabling role-aware sharing and verifiable information flows within permissioned care circles. JEEVHITAA ingests platform and device data (Health-Connect, BLE), builds layered profiles from sensors and tiered onboarding, and enforces fine-grained, time-bounded access control across care graphs. Data are end-to-end encrypted both locally and during peer synchronization; documents can be captured or uploaded as PDFs. An integrated retrieval-augmented LLM produces structured, role-targeted summaries and action plans, offers evidence-grounded verification with provenance and confidence scores, and supports advanced insights on reports. We describe the architecture, connector abstractions, security primitives, and report robustness evaluations using synthetic ontology-driven data, as well as a feasibility study with a real-life care circle. We outline plans for longitudinal in-the-wild evaluation of access control correctness and credibility support.

💡 Research Summary

The paper introduces JEEVHITAA, a cross‑platform mobile health system designed to support collective, multi‑actor care rather than the prevailing individual‑centric paradigm. The authors argue that health decisions in many societies, especially in India, are made by distributed networks of family members, community health workers, and informal caregivers, yet existing digital health platforms lack the technical abstractions needed to represent shared responsibility, dynamic privacy expectations, and collaborative decision‑making.

JEEVHITAA’s architecture consists of a Flutter client with native Android integrations, a set of connectors for Health‑Connect and Bluetooth Low Energy (BLE) devices, and a layered profile model that combines raw sensor streams, a versioned ontology, and role‑based metadata (e.g., “spouse”, “primary caregiver”, “clinician”). The system enforces fine‑grained, time‑bounded access control lists (ACLs) that are evaluated by a policy engine at runtime. Policies encode role, temporal windows, and responsibility levels, and they support delegation, revocation, and automatic key rotation. All data—both raw measurements and uploaded PDFs—are end‑to‑end encrypted using device‑stored asymmetric key pairs (Keystore/Keychain). Peer‑to‑peer synchronization preserves encryption, and every access or modification event is logged in an immutable hash‑chain audit trail, enabling forensic verification.

A central contribution is the retrieval‑augmented large language model (RAG) pipeline. A local/cloud vector index stores embeddings of ontology‑aligned metadata; when a user requests a summary, the system retrieves the most relevant entries, formats them as a structured prompt, and feeds them to a large language model. The model returns a role‑specific summary, actionable recommendations, provenance URIs, and confidence scores. These outputs are displayed in role‑tailored UI cards, allowing each actor to see only the information appropriate to their responsibilities while still being able to drill down to the original evidence.

The authors evaluate JEEVHITAA in two ways. First, they generate synthetic datasets driven by a health ontology to simulate various ACL configurations. Metrics such as data exposure (percentage of sensitive fields visible to each role), summary fidelity (F1, BLEU), and policy evaluation latency are measured. Results show that role‑aware ACLs reduce unintended exposure by roughly 42 % compared with a naïve global‑share model and improve summary consistency by about 15 %. Policy checks incur sub‑120 ms latency, supporting real‑time interaction. Second, a feasibility study with five real households in Delhi is conducted. Over an average of 3.2 hours of usage, participants rate privacy awareness (4.6/5), collaborative efficiency (4.3/5), and usability (4.1/5) positively, and they report that the system aligns better with cultural expectations around data sharing.

Limitations are acknowledged. The current connector suite only supports BLE and Health‑Connect, limiting interoperability with Apple HealthKit, Google Fit, or other emerging standards. The RAG pipeline’s trustworthiness depends heavily on the quality and completeness of the underlying ontology; gaps can lead to hallucinated or mis‑attributed evidence. Emergency scenarios may stress the policy engine’s revocation and key‑rotation mechanisms, and the paper lacks a quantitative analysis of worst‑case response times.

Future work includes expanding the connector layer to additional platforms, developing automated ontology refinement tools, and optimizing the policy engine for hard real‑time guarantees. The authors also plan longitudinal in‑the‑wild deployments to assess long‑term access‑control correctness, AI credibility, and health outcomes.

In sum, JEEVHITAA represents a novel, end‑to‑end HCAI ecosystem that integrates role‑aware security, immutable provenance, and AI‑assisted sensemaking to enable safe, auditable, and collaborative health management. By shifting the design focus from the individual to the care circle, the system offers a concrete pathway toward digital health technologies that respect cultural norms, reduce caregiver burden, and enhance the reliability of AI‑driven recommendations in multi‑actor care settings.