Machine Learning-Based Uncertainty Quantification for Postmortem Interval Prediction from Metabolomics Data

Sammanfattning

Determining the postmortem interval (PMI)–the time since death–is critical in forensic science. Yet, traditional methods often become unreliable due to the variable nature of biological processes after death. Recent advances in machine learning (ML) and metabolomics offer new opportunities for more accurate and interpretable PMI estimation. However, most ML approaches provide only point estimates without conveying the model’s confidence. This study uses Monte Carlo Dropout (MCD) to predict PMI from metabolomics data, providing a corresponding prediction interval, and investigates Variational Autoencoder(VAE) for capturing uncertainty in the data structure, reflecting variability in the metabolomics data. The best-performing MCD model, a multilayer perceptron with dropout active during inference, achieved a mean absolute error (MAE) of 1.57 days and showed a moderate correlation between Mean Prediction Interval Width (MPIW) and model error, indicating that uncertainty estimates can reflect predictive confidence. Increasing the dropout rate improved Prediction Interval Coverage Probability (PICP) but also widened MPIW, highlighting a trade-off between reliability and precision. The VAE, while effective at reconstructing data, showed limited utility for detecting out-of-distribution samples. These findings highlight the importance of incorporating uncertainty estimation into PMI prediction, as it enhances the interpretability and applicability of the results, making them more defensible in forensic contexts.