How to Build a Verifiable Reasoning System for Medical Triage

A verifiable reasoning system for medical triage combines a neural network for pattern recognition with a symbolic logic engine for rule-based decision-making. The neural component analyzes unstructured patient data—such as symptom descriptions from a telemedicine transcript—to extract structured features like chest_pain or high_fever. This structured output is then passed to the symbolic component, which applies deterministic triage protocols like the Emergency Severity Index (ESI). This hybrid architecture ensures the system's decisions are not just accurate but also explainable, a critical requirement for high-stakes medical applications.

The practical value lies in generating an auditable reasoning trace. For each patient, the system outputs a final triage level (e.g., ESI Level 2) alongside a logical proof: 'Assigned Level 2 due to rule R4: chest pain AND age > 50 triggers high-risk cardiac pathway.' This trace allows clinicians to verify the logic, builds institutional trust, and meets regulatory demands for transparency under frameworks like the EU AI Act. Building this system requires careful integration of tools like PyTorch for the neural model and a rule engine like CLIPS or Prolog for the symbolic layer.

The Neural Symptom Analyzer is the perception layer of your neuro-symbolic system. Its primary function is to map raw, often messy patient data—like a chief complaint of "chest pain and shortness of breath"—into a structured, machine-readable format for logical evaluation. You typically implement this as a fine-tuned Small Language Model (SLM), such as Microsoft's Phi-3 or a distilled Llama variant, optimized for medical named entity recognition and symptom classification. This model must output a normalized set of clinical entities (e.g., symptom: dyspnea, severity: 7/10, duration: 2 hours) that serve as facts for the downstream symbolic reasoner.

To build it, start with a clinical corpus for fine-tuning, focusing on symptom lexicons like SNOMED CT. Use a multi-label classification head to tag inputs with relevant symptoms, signs, and patient demographics. Crucially, the model should also output a confidence score for each extracted entity. This score is a key signal for the symbolic layer and potential Human-in-the-Loop (HITL) Governance Systems, triggering human review for low-confidence interpretations. The output is not a diagnosis, but a cleaned, structured fact set ready for logical triage protocol application.