FirstCare AI: Probabilistic Triage System for Early Medical Decision Support

Abstract

Emergency departments in Algeria face a persistent crisis of overcrowding, prolonged waiting times, and a triage process that remains entirely manual and unsupported by intelligent decision-aid technology. A critical aggravating factor is linguistic: the majority of Algerian patients express their symptoms in Algerian Darija, a dialect structurally distinct from Modern Standard Arabic and absent from all existing medical NLP resources — creating a systematic barrier to accurate and timely triage assessment. This thesis presents a multilingual, voice-interactive medical chatbot designed to as sist clinicians during initial emergency triage. Fine-tuned on a purpose-built corpus of 5536 clinician-validated patient utterances — the first annotated medical triage dataset for Algerian Darija. the XLM-RoBERTa-based system processes input in Algerian Darija, Modern Standard Arabic, French, and English, and produces a medical specialty recommendation alongside a three-tier urgency classification. Un der a reproducible multi-seed evaluation protocol, the system achieves 90% accuracy on risk-level classification and 95% on medical specialty prediction, demonstrating that robust cross-lingual transfer to an entirely unseen dialect is attainable through disciplined corpus construction and careful fine-tuning. The system is designed to support not replace the human clinician, by structuring symptom elicitation and reducing cognitive load during high-volume shifts, while preserving full clinical decision-making authority.