Amoebanator: Calibrated Triage for Severe CNS Infections

Clinical ML • Conformal • OOD

Abstract

Amoebanator is a 9-class differential diagnosis system for severe central nervous system infections, with primary amoebic meningoencephalitis (PAM) caused by Naegleria fowleri as the highest-stakes target class. The system pairs a calibrated probability over the 9 conditions with explicit abstention when the input falls outside the validated population.

Methods

PyTorch implementation built on the RigoBERTa Clinical encoder. A 5-model deep ensemble produces the predictive distribution. Out-of-distribution detection is dual-gated, combining Mahalanobis distance in feature space with a logit-energy criterion. Abstention is governed by Mondrian per-class conformal prediction so coverage is controlled within each diagnostic class instead of pooled across the 9.

Calibration

Precision–Recall

Decision Curve (Net Benefit)

Results

Results will be reported in the medRxiv preprint targeted for May 28, 2026. The current development build evaluates calibration, ensemble disagreement, and OOD detection on held-out splits.

Limitations

PAM is rare. The ensemble is trained on a limited number of confirmed cases and a substantially larger set of differential diagnoses, which means OOD performance under true deployment shift is the primary unknown. External prospective validation is the next milestone.

Ethics

Decision support, not autonomous diagnosis. The abstention gate is the safety control. The model is intended to flag suspected PAM early enough that confirmatory CSF testing can be ordered, not to replace it.

Demo content for review. No clinical use.