Semi-supervised learning with dynamic classifier selection for PMI estimation: An animal study

The application of machine learning in analyzing postmortem molecular alterations represents a promising strategy for estimating the Postmortem Interval (PMI), yet samples with undetermined PMI significantly compromise model efficacy and compound the existing scarcity of forensically viable samples. The generalization ability of the model is significantly impaired by the insufficient effective sample size, which greatly hinders the widespread application of PMI estimation based on machine learning. Semi-supervised learning (SSL) methods can effectively leverage both labeled samples with known PMI and unlabeled samples with unknown PMI to jointly train models. This approach is crucial for enhancing sample utilization and improving the accuracy of PMI estimation. In this study, we examined skeletal muscle from rats at different PMI. Using both labeled and unmarked PMI samples, we evaluated the efficacy of Supervised Learning (SL) and SSL models in addressing the prevalent challenge of PMI estimation within forensic science. On this basis, we employed the dynamic classifier selection (DCS) strategy to enhance the construction process of both SL and SSL models. The findings demonstrate that SSL-DCS significantly enhances the predictive efficacy of PMI compared to SL-DCS, exhibiting an area under receiver operating characteristic curve (AUC) of 0.89 and an R² of 0.93 for time to death from 0 to 9 days. The outcomes of this study suggest that SSL-DCS not only enhances prediction efficacy but also improves the utilization rate of scarce samples, which would serve as a paradigm that can be extended and applied to other biomedical domains.