Human Interventions Impact Multi-Agent Medical AI Accuracy.

Multi-agent medical AI systems, designed for clinical reasoning, are susceptible to human interventions, particularly at critical "fault points" where an agent's reasoning is most vulnerable. This research explored how such interventions impact diagnostic accuracy. Using the MedQA dataset, the study simulated doctor-patient conversations to analyze the effects of human input on the AI's reasoning and diagnostic outcomes.The findings revealed a significant impact: correctly timed and accurate human interventions could boost the baseline diagnostic accuracy of the AI systems by as much as 40%. Conversely, interventions that were incorrect or introduced biases led to a performance degradation of up to 6%. These detrimental interventions also increased diagnostic drift and uncertainty within the AI's reasoning process.Beyond mere performance shifts, the analysis uncovered behavioral similarities between cognitive biases observed in the simulated AI environments and those commonly found in real-world clinical practice, such as premature closure or susceptibility to misleading cues. This suggests that identifying and strategically guiding these fault points with human input could be a powerful mechanism for enhancing the robustness and reliability of diagnostic outcomes in multi-agent medical AI systems.