A safety-oriented hypothetico-deductive framework for AI-assisted differential diagnosis

2026-07-11 · Source: cs.AI updates on arXiv.org · Field: Health & Wellbeing — Clinical Care & Medical Practice, Medical Devices & Health Technology, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

AegisDx is a novel safety-oriented diagnostic reasoning framework designed to enhance AI-assisted differential diagnosis, moving beyond conventional one-shot predictions. It coordinates specialized large language model (LLM) components through structured outputs, evidence retrieval, and verification gates to generate broad differential diagnoses, explicitly screen for dangerous "must-not-miss" conditions, and verify reasoning against medical evidence like PubMed. Evaluated across three layers, AegisDx consistently outperformed standalone LLMs. On literature-derived cases, it achieved Top-3 diagnostic accuracy of 59.9% on JAMA cases (+7.8 percentage points) and 62.7% on NEJM cases (+11.3 percentage points) with GPT-oss-120B. For emergency medicine cases, Top-3 accuracy was 85.7% (+17.1 percentage points), capturing must-not-miss conditions in 78.0% of cases (+26.0 percentage points). A blinded physician evaluation of 43 real-world ED notes from YNHHS showed AegisDx improved the composite safety score from 4.31 to 4.55 (adjusted p=2.1\u00d710⁻⁴) compared to GPT-5, with gains in critical condition identification.

Key takeaway

For Machine Learning Engineers developing clinical AI, prioritizing diagnostic safety over raw predictive accuracy is crucial. Your systems should adopt hypothetico-deductive reasoning frameworks like AegisDx, explicitly screening for "must-not-miss" conditions and grounding reasoning in external medical evidence. This approach, even if it impacts output conciseness, significantly improves physician-rated safety and clinical utility, especially in acute care settings.

Key insights

AI diagnostic systems benefit from hypothetico-deductive reasoning and safety-oriented frameworks over one-shot predictions.

Principles

• Clinical diagnosis is an iterative, hypothetico-deductive process.
• Separate likely-diagnosis generation from must-not-miss screening.
• Diagnostic AI should prioritize safety-critical functions, not just accuracy.

Method

AegisDx coordinates specialized LLM agents via role contracts, structured intermediate outputs, evidence-retrieval interfaces, and verification gates across three stages: hypothesis generation, evidence gathering/verification, and clinical management.

In practice

• Implement explicit "must-not-miss" condition screening in diagnostic AI.
• Ground AI reasoning in external medical evidence like PubMed and clinical guidelines.
• Structure AI outputs for actionable next diagnostic and management steps.

Topics

• AegisDx
• Diagnostic AI
• Patient Safety
• Large Language Models
• Clinical Decision Support
• Multi-agent Systems

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Domain Expert

Related on AIssential

• A safety-oriented hypothetico-deductive framework for AI-assisted differential diagnosis — Engineering diagnostic AI as a safety-oriented reasoning framework provides safer, more transparent clinical decision support.
• Active Evidence-Seeking and Diagnostic Reasoning in Large Language Models for Clinical Decision Support — LLMs exhibit a substantial performance gap in active, multi-turn clinical diagnosis compared to static, full-context evaluations.
• MedDDC-Eval: Diagnosis-Decoupled Evaluation of Multi-Turn Medical Consultation Agents — MedDDC-Eval decouples medical agent history elicitation from diagnosis for more accurate, interpretable evaluation.
• LiteOdyssey: A Lightweight Reasoning AI Agent for Interpretable Rare-Disease Diagnosis — Structured reasoning with public biomedical tools enables highly accurate rare-disease diagnosis using a single lightweight AI agent.
• Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents — Aegle virtualizes MDT reasoning for clinical intake using a multi-agent framework, improving diagnostic accuracy and documentation quality.
• Operation-Mechanism Alignment for Reliable Clinical Reasoning over Electronic Health Records — Clinical reasoning reliability improves by aligning heterogeneous operations with specialized execution mechanisms in a DAG framework.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.