Interpretable machine learning models to predict antimicrobial resistance

Antimicrobial resistance (AMR) is a growing global public health concern due to the rapid emergence and spread of resistant bacteria and other pathogens. Machine learning (ML) and genomics-based approaches are becoming more common compared to lab-based phenotypic methods for AMR detection as they bypass the laboratory culture process and offer an understanding of AMR-associated factors. The simultaneous prediction of multiple AMRs is a developing area since current multi-label prediction models are not well-suited to deal with a significant number of missing labels. A Rectified Classifier Chain (RCC) model was proposed and showed that it could effectively handle large numbers of missing labels in multi-label genomic sequence data and explain the results. Developing Next-Generation Sequencing (NGS) technologies and increasing Graphical Processing Unit capabilities have enabled rapid prediction of AMR with Deep Learning (DL) models. However, many genomics sequence datasets lack complete labelling, making applying a DL modelling approach challenging. Therefore, a Masked Loss Convolution Neural(ML-ConvNet) network was developed to improve the performance of the AMR prediction, particularly when there are large numbers of missing labels as the second contribution. Additionally, different Explainable AI (XAI) models were explored, and an XAI pipeline was integrated and validated, which showed that the model could predict AMR effectively even with a limited number of features. Finally, a 1-D Transformer based interpretable model was proposed to predict AMR from antimicrobial administration data for real-time antimicrobial stewardship. The results were interpreted and verified using quantitative validation metrics through an Integrated Gradient with rectified baseline pipeline. The proposed methods showed significant improvements in simultaneous AMR prediction with missing labels and provided interpretable results, which can be applied in real-world scenarios to improve antimicrobial stewardship. These contributions will pave the way for personalised antibiotic treatments and the introduction of AMR-aware food in the long term to overcome AMR challenges.