Machine Learning-Based Identification of Poison Type from Blood Biochemical Profiles in Forensic Toxicology: A Proof-of-Concept Study Using Literature-Parameterised Synthetic Data
Litsenziya: cc-by
Annotatsiya
Identification of the causative agent in fatal poisoning is a central task of forensic toxicology. Gas chromatography–mass spectrometry (GC-MS) and liquid chromatography–tandem mass spectrometry (LC-MS/MS) remain the definitive, legally required standard; modern methods have short run times, but their use depends on instrumentation, trained personnel and funding that are not uniformly available. As a proof-of-concept, we examined whether machine learning (ML) can recover toxicant-specific patterns from 20 routine blood biochemical markers. A synthetic dataset of 3,000 records (150 per category) was generated by Gaussian sampling parameterised from published clinical toxicology data for 20 poison categories, each linked to one of 14 mechanistic cause-of-death categories. Random Forest (RF), Gradient Boosting (GB) and a Multi-Layer Perceptron (MLP) were trained on a stratified 80/20 split and evaluated with stratified five-fold cross-validation (CV); 95% confidence intervals (CIs) were computed for CV and test-set estimates. Feature importance was assessed with Gini and permutation methods. RF performed best, with a test weighted F1-score of 0.9850 (accuracy 95% CI 0.972–0.992) and a CV F1 of 0.9850 (95% CI 0.978–0.991) for poison-category classification, and a weighted F1 of 0.9870 for cause-of-death assignment; GB and MLP reached test F1-scores of 0.9769 and 0.9717. Creatine kinase, aspartate and alanine aminotransferases, lactate and bilirubin were the most discriminative markers, consistent with known toxicodynamic mechanisms. Because the data are synthetic, these results measure how well the algorithms recover literature-encoded biochemical signatures (internal validity), not diagnostic accuracy in real casework; polysubstance exposure, inter-individual variability, comorbidities and post-mortem changes were not modelled. The framework is intended as a complementary, hypothesis-generating triage aid, not a replacement for chromatographic–mass spectrometric confirmation, and requires external validation on real antemortem and post-mortem data.
Hali tarjima qilinmagan