ActiveMedAgent: cost-aware sequential test acquisition for multimodal diagnosis

Ma et al. introduce trajectory-based reinforcement learning for clinical decision making, finding that selective diagnostic acquisition outperforms full-modality inputs.

Primary source: arXiv:2610.11140: https://arxiv.org/abs/2610.11140

What’s new: In hospital practice, diagnostic evaluations are sequential and budget-sensitive: physicians begin with inexpensive screenings (such as patient history or basic blood panels) and escalate to costly, invasive procedures (like contrast CT, MRI, or biopsies) only when uncertainty remains high. In ActiveMedAgent: Cost-Aware Trajectory Learning for Multimodal Medical Diagnosis (arXiv:2610.11140, 8 Oct 2026), Ma et al. propose ActiveMedAgent, bringing cost-aware sequential decision theory to multimodal medical models.

Operating on top of a frozen, API-accessible vision-language model, ActiveMedAgent tracks the entropy of diagnostic beliefs across candidate conditions. A lightweight multilayer perceptron (MLP) controller is trained offline on scored trajectories, evaluating each potential diagnostic action by its expected information gain minus clinical and economic cost. Across three benchmark datasets, ActiveMedAgent outperformed both unguided greedy acquisition and full-modality models. Crucially, the authors identify an “information overload” effect: across 175 evaluation cases, the agent produced a correct diagnosis with a restricted subset of tests, whereas presenting the full set of modalities simultaneously led the baseline model to make incorrect predictions.

Why it matters: Medical AI research often assumes that providing more data modalities automatically yields higher accuracy. ActiveMedAgent proves that excessive or redundant inputs can confuse vision-language representations, demonstrating that learning what not to order is as vital for model accuracy as it is for healthcare economics.

Caveats: The policy controller was trained and validated on retrospective benchmark datasets using simulated test costs rather than active hospital workflows. The cost-to-utility weighting coefficient is a static hyperparameter that must be manually tuned to match individual institutional budgets and clinical risk tolerances.