Patients switch topics mid-consultation: PRISM estimates clinical state as a hidden variable and lifts decision accuracy 22.40 points over the strongest baseline
Synopsis
In online medical consultation, patients often volunteer symptoms or pivot to new concerns, so the clinical state is only progressively revealed. PRISM splits triage into belief estimation, policy optimization and response realization: it maintains a posterior over plausible consultation states, then selects the next inquiry from the maximum a posteriori state. On 450 real online consultations, its structured decision accuracy exceeds the strongest baseline by 22.40 percentage points, and BERTScore_F1 by 5.51 points.
Interpretation
The consultation trajectory is no longer treated as a predictable fixed path but as a partially observable decision process: the system first estimates which clinical state the consultation is in, then decides what to ask next. Earlier intelligent triage systems were largely developed on simulated or offline encounters, where patients supply information only when the physician asks directly and interaction stays anchored to predefined clinical profiles and diagnostic goals. Across 15 public medical dialogue datasets plus a real online corpus, authentic online corpora average an ECoh of 0.564, below 0.778 for simulated online and 0.694 to 0.865 for offline; authentic online BPB ranges from 1.373 to 1.466, above 1.035 to 1.272 for simulated online and 0.761 for offline.
The system separates triage into belief estimation, policy optimization and response generation, so the clinical decision and the wording of the question are produced independently. The belief module scores each candidate consultation state, uses an availability mask to zero out states already satisfied, and passes the maximum a posteriori state to the policy. The policy first picks a high-level inquiry action and then a specific inquiry target under the state constraint, trained with group relative policy optimization; evaluation uses 450 real online consultation dialogues from a major telemedicine platform, 50 cases in each of nine disease categories.
Supplying the consultation state explicitly improves response quality on its own, not only through PRISM's own architecture. An ablation gives every baseline a unified oracle state policy that supplies only the ground-truth consultation state, while each model still plans actions through its own chain-of-thought reasoning. All ten models improve on ROUGE-L and BERTScore_F1, by an average of 10.75 and 5.58 points respectively; MiMoCode gains 12.97 and 3.76 points, and MeDxAgent gains 10.96 and 3.18 points.
Perspective
Evaluation uses 450 real online consultation dialogues from a major telemedicine platform, 50 cases in each of nine disease categories, alongside 15 public medical dialogue datasets; structured decisions and response generation are both measured within this triage task and annotation schema.
Privacy requirements and partner agreements restrict raw consultation records to internal research, limiting opportunities for independent clinical expert evaluation and calibration of medical annotations; as a compensatory measure, each sample is annotated three times by a privately deployed model and only samples with complete agreement across all three passes are retained.