Making Confidence Scores Comparable in Multi-Agent Disease Screening: A Gateway Architecture with Per-Agent Calibration

This research paper presents a multi-agent disease screening architecture designed to make confidence scores comparable across independently trained machine-learning models. The proposed gateway architecture integrates twelve screening agents covering structured clinical data, medical imaging, and symptom-based text analysis. A per-agent confidence normalisation method using Platt scaling and isotonic regression is evaluated on three public tabular datasets: PIMA diabetes, Cleveland heart disease, and Wisconsin breast cancer. Across 900 gateway checks, the predicted labels remained unchanged. The reference experiment showed improved confidence calibration for support vector machine agents, while also identifying limitations involving small calibration samples, untested cross-condition rules, and the need for further clinical validation. Keywords: Multi-agent systems, Disease Screening, Machine Learning, Confidence Calibration, Platt Scaling, Clinical Decision Support, Medical AI.