From Black-Box to Glass-Box: Knowledge-Constrained Neuro-Symbolic Acute Pharyngitis Triage under Data Scarcity

Abstract Acute pharyngitis is a major challenge in primary care and a key contributor to antimicrobial resistance (AMR) due to inappropriate prescribing for viral etiologies. While Deep Learning shows potential, "Black-Box" opacity and data scarcity hinder clinical adoption. We propose a Knowledge-Constrained Neuro-Symbolic architecture decoupling visual perception from diagnostic reasoning. Trained on minimal data (N=89) and validated externally (N=742), our system significantly outperformed a ResNet-50 baseline (Specificity: 90.3% vs. 69.7%). Crucially, it resolved the "Pediatric Trap"—where baseline specificity collapsed to 55.0% in children due to physiological hypertrophy—by maintaining 84.7% specificity via age-stratified causal logic. Furthermore, it achieved double the sensitivity of adapted clinical scores (14.8% vs. <7%) while safely rejecting 46% of low-quality inputs. This "Glass-Box" approach prevents dangerous hallucinations, offering a transparent safety net for autonomous triage.