Protocol for an In Silico Evaluation of an Agentic AI Clinical Decision-Support System for Febrile Illness Triage and Referral in Nigerian Primary Healthcare

This record contains the prospectively deposited protocol for a preclinical in silico evaluation of a human-supervised agentic AI clinical decision-support system for febrile illness triage and referral in Nigerian primary healthcare. The study compares a plain large language model, the same model with retrieval-augmented generation, and a full agentic configuration on a locked synthetic benchmark. The primary safety outcome is dangerous under-triage. Secondary outcomes include overall triage agreement, urgent-referral sensitivity, overtriage, resource appropriateness, missing-information handling, and appropriate abstention or escalation. The record also documents benchmark-locking procedures, reference-standard clinician adjudication, system versioning, and prespecified statistical analyses.