AI Diagnostic Agent Learns from Experience Like Clinicians, Boosting Accuracy 92%

A new approach to AI-assisted diagnosis combines two key elements of human clinical expertise: sound reasoning and the ability to learn from past cases. Researchers have developed SEA (a self-learning diagnostic agent), which mimics how physicians improve over time by building a library of diagnostic patterns alongside their analytical skills.

The Problem with Current AI Diagnostics#

Most AI systems for medical decision support treat each case in isolation. They analyse a patient's symptoms and data from scratch without drawing on lessons from previous cases. This limits their capacity to develop the kind of pattern recognition that distinguishes experienced clinicians from novices. SEA addresses this gap by enabling the system to accumulate and reuse diagnostic insights across multiple encounters.

How SEA Works#

The system features a dual-memory module inspired by cognitive science. Rather than storing raw case data, this memory consolidates patterns and diagnostic rules that emerge from accumulated experience. The researchers paired this architecture with a reinforcement training framework that jointly optimises both the agent's reasoning ability and its memory management—essentially teaching the system when to form new rules and when to apply existing ones.

Strong Performance Across Tests#

On standard benchmarks using the MedCaseReasoning dataset, SEA achieved 92.46% accuracy, surpassing the strongest competing method by nearly 20 percentage points. In a more demanding long-horizon test with the ER-Reason dataset, it reached the highest final accuracy (0.7214) and showed the most consistent improvement, while other approaches plateaued or became unstable.

Clinical Validation#

Beyond raw numbers, expert clinicians reviewed the diagnostic rules that SEA had extracted. They found these rules to be both clinically sound and practically useful—a critical validation that the system wasn't simply gaming the metrics but genuinely capturing reliable medical knowledge. This suggests the dual-memory approach successfully transforms raw case experience into reusable diagnostic wisdom.