The current methods for determining the cause of death in cases of suspected intoxication oroverdose require improvement. Complex cases involving multiple substances, a developed drugtolerance, or the emergence of New Psychoactive Substances (NPS), pose significant challenges toexisting approaches. Therefore, more advanced methods are necessary to allow for more robust andreliable conclusions. This master’s thesis explores an alternative approach that utilizes post-mortemmetabolomics to identify the lethal substance, in the absence of detectable drug concentrations inthe blood. Each drug belongs to a toxidrome based on the physiological response it triggers. Theunderlying hypothesis is that these distinct physiological reactions leave specific metabolic signaturesin the blood, which can be analyzed to determine the toxidrome that contributed to the fatal outcome.This method has the potential to determine the cause of death without directly detecting the drug, aidinterpretation in cases involving multiple substances, and account for drug tolerance effects.Using multiple datasets from the National Board of Forensic Medicine (RMV, Rättsmedicinalver-ket) containing in total 582 individuals with approximately 3500 metabolites each, several modelswere developed to predict toxidromes and determine the presence of drugs in individual cases. Thiswas achieved by using positive and negative control cases, as well as intoxication cases. Random forestand logistic regression were employed to predict the toxidrome of each individual, while principalcomponent analysis (PCA) was conducted to assess the influence of variables such as age, sex, BMI,and post-mortem interval (PMI). Additionally, the internet platform MetaboAnalyst was utilized tofurther analyze metabolite differences between toxidromes and to identify specific metabolites recog-nized as important by machine learning models for predictingThe results suggest that age, sex, BMI, PMI do not significantly influence post-mortem metabolomicswhen analyzed using PCA. However, a slight separation of data points was observed when color-coded by intoxication, positive control group, and negative control group. Regarding classificationperformance, the logistic regression model consistently outperformed the random forest model innearly all area under the curve (AUC) comparisons, regardless of whether negative and positive con-trol cases were included. The logistic regression model achieved AUC scores ranging from 0.74 to 0.99,while the random forest model produced scores between 0.62 and 0.98. Furthermore, both machinelearning models and MetaboAnalyst presented challenges in metabolite identification, suggesting thatthis approach using these tools may not be optimal for this purpose.When the random forest and logistic regression models were applied to cases influenced by drugsnot included for model training, notable results emerged. This dataset consisted of NPSs, whichbelong to the same toxidromes as the original drugs and are expected to trigger similar physiologicalresponses. Since the machine learning models were designed to predict toxidromes based on post-mortem metabolomic data, the models should, in theory, be capable of classifying the NPS accordingly.However, the models demonstrated poor performance, with most AUC scores falling below 0.5, indi-cating that their predictions were less accurate than random guessing. However, the random forestperformed well for accurately predicting cases that belonged to the NPS opioid group.In conclusion, it was possible to develop models to accurately predict the correct toxidrome basedon the post-mortem metabolomics, but only if the models were trained on the same drugs as testedon. The results from the NPS dataset could indicate that the models do not predict the toxidromebased on the endogenous metabolites found post-mortem, but has instead based the predictions onsomething else, possibly the metabolites from the consumed drug. In order to further investigatewhich metabolites play a role in the toxidrome predictions and the responses from the body, othermethods needs to be implemented and include additional cases.