Development of a Machine Learning Model for Rapid Identification of Antibiotic-Resistant Bacteria

Thesis Abstract

The rapid and precise identification of antibiotic-resistant bacteria remains a critical challenge in clinical microbiology, contributing significantly to the escalation of treatment failures, increased healthcare costs, and the propagation of antimicrobial resistance globally. Traditional laboratory methods such as phenotypic testing and molecular diagnostics, while accurate, are often time-consuming, labor-intensive, and resource-intensive, underscoring the urgent need for innovative, rapid diagnostic tools that can facilitate timely clinical decision-making. This study aims to develop and evaluate a machine learning-based predictive model capable of swiftly identifying antibiotic-resistant bacterial strains from microbiological data, thereby enhancing diagnostic efficiency and supporting antimicrobial stewardship initiatives. The specific objectives include (1) to compile and preprocess a comprehensive dataset comprising genomic sequences, phenotypic resistance profiles, and clinical metadata of bacterial isolates; (2) to explore and select optimal features influencing resistance patterns using feature importance and correlation analyses; (3) to design, implement, and train various supervised machine learning classifiers—including Random Forest, Support Vector Machine (SVM), and Gradient Boosting algorithms—aiming to identify the most accurate and robust model for resistance prediction; (4) to validate model performance through cross-validation and independent testing, assessing metrics such as accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC-ROC); and (5) to interpret predictive features within the context of resistance mechanisms guided by the Health Belief Model and pathogen-specific resistance theories. The study employed a quantitative research design, utilizing a retrospective dataset of 1,200 bacterial isolates collected from hospital microbiology laboratories over a three-year period. The dataset included genomic data obtained via whole-genome sequencing, phenotypic antibiotic susceptibility testing results, and relevant patient demographic and clinical information. Data collection instruments incorporated automated data extraction from laboratory information management systems (LIMS), supplemented by manual curation to ensure data integrity. The dataset was randomly divided into training (70%) and testing (30%) subsets to evaluate model generalizability. Preprocessing steps involved normalization, encoding categorical variables, and handling missing data through multiple imputation techniques. Dimensionality reduction was performed using principal component analysis (PCA), while feature selection was guided by recursive feature elimination (RFE). The study compared multiple classifiers—specifically Random Forest, SVM with radial basis function kernel, and Gradient Boosting Machines—using scikit-learn in Python. Model optimization was conducted via grid search with stratified k-fold cross-validation to prevent overfitting. The best-performing model was selected based on comprehensive evaluation of the aforementioned metrics, with emphasis on precision and recall to minimize false negatives in resistance detection. Expected findings include the identification of a machine learning model with an accuracy exceeding 85% and an AUC-ROC greater than 0.90, capable of differentiating resistant and susceptible bacterial strains rapidly within a timeframe of under one hour. It is anticipated that genomic features—particularly resistance gene markers and mutations—will emerge as significant predictors, reinforcing the biological plausibility of the model. These findings are expected to demonstrate that machine learning can significantly reduce diagnostic turnaround time, enabling earlier targeted therapy and curtailing inappropriate antibiotic use. This research contributes to the expanding field of data-driven microbiology and precision medicine by providing an efficient, scalable tool for resistance detection that complements existing laboratory methods. It advances the theoretical understanding of resistance prediction through the integration of genomics, machine learning, and behavioral health models. By elucidating key genomic predictors and optimizing classifier performance, this thesis offers practical insights for implementing rapid diagnostic tools across healthcare settings. In conclusion, the study recommends the adoption of machine learning algorithms in routine microbiological workflows, further validation with larger and diverse datasets across different regions, and integration with clinical decision support systems to improve patient outcomes and combat antimicrobial resistance effectively.