Early detection of antimicrobial drug resistance for salmonella typhistrains using machine learning techniques in Rwanda

Abstract

Introduction: Antimicrobial drug resistance to Salmonella Typhi is among complex risk factors for morbidity and mortality, thus of global public health concern. Antimicrobial resistance patterns for S.Typhi suggest that currently, it presents a growing problem for developing countries. Early treatment of infectious diseases like Salmonella Typhi is key to combating high morbidity and mortality rates. Antimicrobial drug susceptibility test takes more than 24 hours, which is time consuming and inefficient. More recently, predictive models have been used elsewhere to predict antimicrobial drug resistance patterns using machine-learning techniques for quick turn-around time and more efficient especially for patients in acute care conditions. Objective: The main aim of this study was to predict a patient’s antimicrobial drug resistance to Salmonella Typhi using Machine Learning Techniques. Methodology: A cross-sectional study (2015 -2019) was conducted and of the 152 Salmonella Typhi isolates included in the study, 140 (92.1%) were from blood while 12 (7.9 %) were from stool. The Kirby-Bauer testing method was used for antimicrobial susceptibility. This study also predicted a patient’s antimicrobial drug resistance to Salmonella Typhi using four machine learning techniques namely; Support Vector Machine, Decision tree, Random Forest, and Logistic Regression using Antimicrobial Resistance data from a national reference laboratory on 765 cases in Rwanda. 5-fold cross-validation, classification report and confusion matrix metrics were used for performance measurement of the models. Results: From 2015 to 2019, Cotrimoxazole resistance (86.2%) was highest compared to other first-line drugs: Ampicillin (85.5%) and chloramphenicol (80.9%). Nalidixic acid resistance (59.9%) and ciprofloxacin (20.4%) were high. There was lower resistance ceftazidime (32.9%), Tetracycline (9.9 %), and Cefotaxime (7.2%). All the built models had high predictions of antimicrobial drug resistance to Salmonella Typhi. Decision tree gave f1-score [0.89], accuracy [0.85] and AUC [0.82], Random forest gave f1-score [0.86], accuracy [0.90] and AUC [0.83], logistic regression gave f1-score [0.86], accuracy [0.88] and AUC [0.87] while Support Vector Machine f1-score [0.86], accuracy [0.89] and AUC [0.88]. However, a comparison that is based on the detailed performance measures suggests that the Support Vector Machine performs best. Conclusion: There are significant antimicrobial resistance patterns in S.Typhi isolates to commonly used antibiotics. Applying machine-learning techniques can predict antimicrobial drug resistance for Salmonella Typhi with high accuracy without clinical information. This approach may be extrapolated to predict antimicrobial drug resistance for any other organism. Further studies are recommended to determine the actual cost of predictive models on drug resistance in other clinical settings.