Explainable artificial intelligence in forensic DNA analysis: Alleles identification in challenging electropherograms using supervised machine learning methods

Abstract

Challenging samples in capillary electrophoresis (CE)-based short tandem repeat (STR) analysis often produce artefactual signals that cannot be completely filtered out by expert electropherogram (EPG) reading systems, complicating allele interpretation. Previous studies have demonstrated the potential of artificial intelligence (AI) to address this issue by accurately distinguishing allele signals from artefacts in EPGs. Traditional machine learning models offer significant advantages in enhancing the interpretability and transparency of AI models used in DNA analysis, particularly in criminal investigations and legal contexts. In this study, five traditional machine learning algorithms were employed to train and construct models using EPG signal datasets from single-source low-template EPGs, mixture EPGs, and combined datasets. Performance evaluation and validation with additional datasets demonstrated the feasibility of these models in improving the reportability of potential information in EPGs. However, further optimization is needed for mixture EPGs to enhance classification accuracy. Implementing Receiver Operating Characteristic (ROC) curve analysis and prediction probability thresholds effectively reduced false positive classifications. Additionally, a user-friendly platform was developed for EPG signal classification based on machine learning and ensemble learning, allowing for the classification of any signal datasets using traditional machine learning models and combining the prediction results of multiple models. This platform will provide analysts with more optimal and robust results. This study shows that machine-learning-based EPG signal classification models can significantly enhance the efficiency of sample analysis and interpretation, providing a solid foundation for future research.