Excellence in research: trustworthy AI for healthcare: evidence-grounded, compliant, human-in-the-loop

The National Science Foundation Office of Integrative Activities awarded Tuskegee University $699,999 on October 1, 2026, under the Integrative Activities program (CFDA 47.083) to develop trustworthy artificial intelligence systems for healthcare that balance clinical utility with accuracy, privacy, and accountability.

The project addresses hallucination, privacy breaches, and overconfident claims in large language models applied to healthcare by designing AI systems that tie clinical recommendations to validated medical evidence, minimize exposure of patient information to individual AI components, and engage clinicians during high-uncertainty or high-risk scenarios. The work comprises three integrated research thrusts. The first develops a clinical knowledge-grounded retrieval-augmented generation framework that structures validated medical sources as an evidence graph, optimizes query-specific subgraphs, constrains generation to high-confidence evidence paths, and provides provenance linking claims to sources. The second builds a compliance-focused large language model agent that audits and remediates data through redaction, generalization, or synthesis, paired with a compliance-aware orchestrator that constructs least-privilege task graphs for specialized agents. The third implements risk-aware uncertainty monitoring that routes selected outputs to clinicians for human review.

The award runs through September 30, 2029, with performance at Tuskegee, Alabama. Beyond the healthcare application, the research methods are intended to transfer to regulated fields such as law and finance. The project strengthens AI research and education at Tuskegee University, expands student opportunities, broadens participation in the national AI workforce, and shares resources with partner institutions and other historically Black colleges and universities.

Abstract

Artificial intelligence (AI) could help healthcare professionals summarize records, coordinate care, and make faster decisions, but today?s large language models can invent facts, expose sensitive patient information, and provide answers confidently without evidence for the projected confidence. This project asks how healthcare AI can be made accurate, private, and accountable by design. The central idea is to tie important claims to trusted medical evidence, limit each AI component to the minimum patient information needed for its task, and involve clinicians whenever uncertainty or risk is high. Rather than replacing clinical judgment, the system is intended to help experts work more efficiently while preserving their authority over final decisions. The research could enable safer clinical summarization and decision support while producing methods that transfer to other regulated fields such as law and finance. IT will also strengthen AI research and education at tuskegee university, expand student opportunities, broaden participation in the national AI workforce, and share resources with partner institutions and other historically black colleges and universities. The project integrates three research thrusts. First, a clinical knowledge-grounded retrieval-augmented generation framework will organize validated medical sources as an evidence graph of atomic claims and typed relationships, optimize query-specific support subgraphs, constrain generation to high-confidence evidence paths, and return provenance linking claims to sources. Second, a compliance-focused large language model agent will semantically audit and remediate data through redaction, generalization, or synthesis, while a compliance-aware orchestrator builds least-privilege task graphs for specialized agents. Third, risk-aware uncertainty monitoring will route selected outputs to clinicians; structured diagnostic traces will localize failures in evidence, agents, or data flows; and a remediation planner will propose reversible fixes with targeted tests for expert approval. Evaluation will use medical question-answering and DE-identification benchmarks, a mock electronic health record, simulated clinical workflows, penetration testing, and clinician studies with clinical partners to measure accuracy, privacy, efficiency, diagnostic value, and human-AI team performance. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are planned for this award.