Diagnosis of Acute Poisoning using explainable artificial intelligence.

Area of Science:

• Medical Toxicology
• Artificial Intelligence in Medicine
• Clinical Decision Support Systems

Background:

• Medical toxicology faces challenges due to the vast and rapidly expanding volume of knowledge.
• Current machine learning/artificial intelligence (ML/AI) techniques in toxicology lack transparency and physician usability.
• Logic-based systems offer transparency but often generalize poorly and require extensive expert curation.

Purpose of the Study:

• To develop a transparent ML/AI approach for medical toxicology that models clinician decision-making.
• To evaluate a probabilistic logic network (PLN) for recognizing toxidromes based on physical exam findings.
• To compare the PLN's performance against human experts and a decision tree classifier.

Main Methods:

• Constructed a probabilistic logic network, named Tak, to model toxicologist recognition of toxidromes using physical exam findings.
• Created a library of 300 synthetic cases with varying complexity, each featuring physical exam findings from one or two toxidromes.
• Evaluated Tak's performance against two medical toxicologists and a decision tree classifier using inter-rater reliability metrics (Cohen's kappa).

Main Results:

• Tak demonstrated strong inter-rater reliability with human consensus on straightforward cases (κ = 0.8432).
• Performance decreased with complexity, showing moderate (κ = 0.4396) and challenging (κ = 0.3331) case reliability.
• Tak consistently outperformed a decision tree classifier across all case difficulty levels.

Conclusions:

• The probabilistic logic network (Tak) performs comparably to human experts in straightforward and moderately complex medical toxicology cases.
• While outperformed by humans on challenging cases, Tak represents a proof-of-concept for PLNs in medical reasoning within a restricted domain.
• This approach offers a transparent alternative to traditional ML/AI methods, potentially improving usability for clinicians.