Machine Learning Multimodality Fusion for Antimicrobial Resistance Prediction Using Electronic Health Records

Document Type



The spread of antimicrobial resistance (AMR) among patients in the healthcare systems causes major challenges and losses of human lives and medical resources, with an expectation of more deterioration occurring in the future if the problem is not controlled. This effect requires global and local healthcare organizations and researchers to combine efforts to take the appropriate actions to combat the issue. From a machine learning perspective, models could aid clinicians and microbiologists by anticipating the resistance beforehand. One of the ways is by predicting the susceptibility of a patient towards antibiotics using historical patients' records. Electronic health records (EHR) are platforms used by healthcare providers to document the history of patients. EHR systems are a rich resource of clinical notes, time-dependent records, and time-independent records. The variety of modalities enables implementing several machine learning approaches according to the type of data used. In our work, we utilize the MIMIC-IV dataset which includes information of 315,460 patients collected from 2008-2019 in a hospital in the United States, that is then filtered according to the medical task. Through it, we provide an analysis of the EHR data processing methods that aims towards finding suitable practices to handle the patient history to predict their resistance against antibiotics. We use the FIDDLE framework to create datasets customized to the AMR task with different settings of prediction times and time granularities. FIDDLE produces separate matrices for time-invariant and time-series data. According to the quantity of the data, we consider the AMR prediction for the antibiotic Gentamicin. We apply a multimodality approach to take advantage of the three modalities in the MIMIC database: time-series, time-invariant, and clinical notes. We implement various combinations of encoders to the modalities and fuse them. We use four fusion mechanisms and compare their efficiency according to the area under the receiving operating characteristic curve (AUROC) and the area under the precision recall curve (AUPR). Among them, the MultiModal Infomax fusion mechanism achieves the best performance for the AMR task. The highest performance is achieved on the Gentamicin dataset with a prediction time of 4 hours and time granularity of 1 hour, reaching an AUROC of 0.75 and an AUPR of 0.1825.

Publication Date



Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Machine Learning

Advisors: Dr. Mohammad Yaqub, Dr. Min Xu

with 2 year embargo period

This document is currently not available here.