Diagnosis of Acute Poisoning using explainable artificial intelligence

Medical toxicology is the clinical specialty that treats toxic effects of substances, for example, an overdose, a medication error, or scorpion sting. The volume toxicological knowledge and research has, as with other medical specialties, outstripped ability individual clinician to entirely master stay current it. application machine learning/artificial intelligence (ML/AI) techniques challenging because initial treatment decisions are often based on few pieces textual data rely heavily experience prior knowledge. ML/AI techniques, moreover, do not represent in way transparent physician, raising barriers usability. Logic-based systems more approaches, but generalize poorly require expert curation implement maintain. We constructed probabilistic logic network model how toxicologist recognizes toxidrome, using only physical exam findings. Our approach transparently mimics representation decision-making practicing clinicians. created library 300 synthetic cases varying complexity. Each case contained 5 findings drawn from mixture 1 2 toxidromes. used this evaluate performance our network, dubbed Tak, against toxicologists, decision tree model, well its recover actual diagnosis. inter-rater reliability between Tak consensus human raters was ? = 0.8432 straightforward cases, 0.4396 moderately complex 0.3331 cases. classifier was, 0.2522 0.1963 0.0331 software, performs comparably humans intermediate difficulty outperformed by outperforms at all levels difficulty. results proof-of-concept that, restricted domain, networks can perform reasoning humans.