Machine learning and metabolic signatures of drowning: A pathway to uncovering cause of death in aquatic environments

Abstract

The postmortem diagnosis of drowning is challenging due to the nonspecific and transient nature of classical autopsy findings. This study aimed to investigate whether postmortem metabolomics can differentiate drowning from other causes of death, offering a potential biochemical tool to support forensic diagnosis in water-related deaths. A total of 503 drowning cases and four control groups were included, matched on sex, age, BMI, and postmortem interval. Three control groups represented alternative causes of death relevant to bodies found in water; chronic heart disease (n = 510), intoxication (n = 516), and trauma (n = 497). Hangings (n = 511) were included as a fourth "positive" control group to see how well the model can separate two different asphyxiation processes. Using multivariate modeling, we investigated whether drowning could be discriminated from these competing causes of death based on metabolomic signatures. Four binary OPLS-DA models comparing drowning to each control group showed good performance (R2 = 0.61-0.76; Q2 = 0.40-0.56), with sensitivities and specificities ranging from 83 to 87 % and 78-89 %, respectively. Metabolite and pathway analyses identified 52 differentiating features and multiple significantly enriched pathways, including glycerophospholipid metabolism, steroid hormone biosynthesis, and cytochrome P450-related drug metabolism. In conclusion, postmortem metabolomics show promising accuracy for forensic cause-of-death determination of drowning cases, with minimal impact from postmortem submersion times, though further research is needed to fully evaluate the applicability.

Keywords: Drowning; Forensic diagnostics; Machine learning; Metabolomics.