Background. Interpreting toxicology results at the level of an individual case is hard. The clinical meaning of a measured drug (or metabolite) concentration depends on numerous case-specific factors (e.g. individual’s body mass, tolerance, co-ingestants). In practice, practitioners often rely on therapeutic/toxic/fatal “thresholds” and professional judgement to bridge information gaps — an approach that can hide assumptions, treat uncertainty poorly, and lead to inconsistent conclusions. A transparent framework that integrates imperfect evidence for the individual case is still missing.
Novelty & importance. This project will create a hybrid AI framework uniting object-oriented Bayesian networks (OOBNs) with modern machine learning to support case-level conclusions. OOBNs encode causal structure and scale via modular building blocks, with strong precedents in medical diagnostics and forensic evaluation. What is new here is (i) learning from data to update parameters (and, where appropriate, sub-structures) rather than fixing them a priori; (ii) coupling to physics-informed pharmacokinetics/pharmacodynamics (PK/PD) surrogates (e.g., neural-ODE models) to infer patient-specific dynamics from sparse measurements; and (iii) producing quantified outputs with calibrated uncertainty and provenance. Together, these advances will enable a more robust and transparent interpretation of toxicological findings and blueprint next generation expert systems.
Aims & objectives. The project aims to build and validate a decision-support system for transparent, uncertainty-calibrated, case-level interpretation of toxicology findings. Concretely, it will: (1) design modular OOBN components that capture key PK/PD processes and contextual factors; (2) learn network parameters/structure from curated case datasets while retaining expert priors; (3) integrate neural-ODE models to represent patient-specific dynamics; (4) implement a citation-tracked, LLM-assisted knowledge-ingestion pipeline to keep priors and ranges current; (5) calibrate uncertainty and assess fairness/robustness; and (6) deliver a clinician-centred interface with “why–because” explanations aligned to routine reporting. Initial development will focus on well-characterised substances (e.g., ethanol), with later extension to additional drugs.