Experimental AI system improves identification of drugs in emergency poisoning cases

Researchers report that DeToxR, an AI system designed to identify multiple substances involved in emergency poisonings, outperformed an unadapted language model and supervised baselines. It combines accounts from paramedics and patients with medical measurements, then uses a training reward that penalizes missed substances and false identifications. In a preliminary clinical validation, the researchers reported micro-F1 scores of 0.644 versus 0.473 and higher recall. The system is intended to support decisions when ingestion histories are incomplete and symptoms do not clearly identify the substances involved.