Meta-learning for predictive modeling of antimicrobial resistance in dynamic healthcare environments

Abstract

Antimicrobial Drug Resistance (AMR) is a growing global health concern char- acterized by microbes developing mutations to survive antimicrobial drugs. Al- though mutation among organisms is a natural evolutionary process, it can be en- hanced by human activity, such as antibiotic misuse. Existing AMR determination methods present several challenges, including high costs, laborious nature, and a turnaround time that usually runs up to 48 hours. During this delay, patients are often prescribed broad-spectrum antibiotics that target both Gram-positive and Gram-negative bacteria as a precautionary measure. Broad-spectrum antibiotics have been found to contribute significantly to antimicrobial resistance. This work presents a machine learning approach to the timely prediction and interpretation of AMR from genomic sequences. It describes a two-part intelligent clinical decision support tool comprised of an Amino Acid analyzer that converts resistant bacterial genomes into amino acid sequences to establish a relationship between amino acid frequency and a machine learning model that predicts resistance patterns in micro- bial genomes. It starts with Exploratory data analysis which is done with the help of a Python algorithm that drops the least represented datasets and uses augmenta- tion techniques to generate new sequences with characteristics similar to the target genome without affecting model predictions. Different AI models are developed, an RNN with K-Mer encoding, a combined LSTM and RNN and a combined LSTM, RNN and Meta Learner with accuracies of 66.4%, 92.9% and 89.9% respectively. Model 3 which makes use of metalearning is selected as the best model because of its ability to make accurate predictions from partial sequences. The meta-learner is then packaged into an API that is embedded into the rest of the intelligent clinical decision support tool together with Explainable AI tools Alibi, Lime, and SHAP to generate clinically relevant and interpretable predictions. Results from the amino acid analyzer indicate a close correlation between Valine, Glycine, Leucine and Ala- nine and AMR. This work addresses the computational problem of large-scale ge- nomic data efficient analysis, genomic data imbalances, and the inability to provide insights into underlying resistance mechanisms, such as the potential relationship between amino acid frequency and AMR by developing accurate, generalizable, and explainable models capable of identifying resistance patterns from raw genomic data and helping healthcare providers make timely data-driven decisions while uncovering potential biochemical resistance markers through amino acid frequency analysis.