EXPLAINABLE RECURRENT NEURAL NETWORKS FOR PREDICTING ANTIMICROBIAL MULTIDRUG RESISTANCE IN INTENSIVE CARE UNIT

Abstract

Antimicrobial Resistance (AMR) is the natural phenomenon that allows pathogens to withstand antimicrobial treatments, reducing the efficacy of these medications. Human activities, especially the misuse and overuse of antibiotics, significantly contribute to the acceleration of AMR, and the issue has been further exacerbated by the COVID-19 pandemic that began in 2019. As a result, many microorganisms have developed resistance to multiple antimicrobials they were once susceptible to, a condition known as Multidrug Resistance (MDR). This development complicates patient care due to limited effective treatments, increasing hospital stays and mortality rates. MDR pathogens pose a major concern, especially in Intensive Care Units (ICUs), given the compromised health status of ICU patients. The current method for detecting MDR pathogens requires 24 to 48 hours, which delays the initiation of appropriate treatments. Therefore, accelerating MDR detection is critical, and advancements in data science offer hope for predicting MDR cases and improving early preventive measures.

This bachelor's thesis focuses on interpretable data science techniques to predict the temporal development of MDR in ICU settings and identify potential risk factors over time. The study analyzes Electronic Health Record data from ICU patients admitted between January 2004 and February 2020 at the University Hospital of Fuenlabrada in Madrid, Spain. This clinical data provides information on the health status of the patient over time. Therefore, the data is modeled as an irregular Multivariate Time Series (MTS), to classify whether a patient will acquire MDR on each day of the first 14 days of their ICU stay. To ensure clarity for clinicians regarding influential factors, the study develops an interpretable framework using various Explainable Artificial Intelligence (XAI) techniques adapted to temporal outputs and irregular MTS. The performance and interpretability of these methodologies are then compared.

From a data science and literature perspective, this study introduces several significant contributions to the field. Firstly, it pioneers the development of a Gated Recurrent Unit (GRU) model tailored for temporal classification to predict MDR. Secondly, it adapts XAI methods for irregular MTS with a GRU designed for temporal output. Thirdly, it involves the adaptation and application of Shapley Additive Explanations for irregular MTS and a Recurrent Neural Network with temporal output.

Clinicians view the performance outcomes of this study as a notable advancement in addressing MDR. The study achieved a Receiver Operating Characteristic Area Under the Curve of 78.27\%, demonstrating an improvement over results reported in the literature. Additionally, the interpretability analysis reinforces clinical knowledge and enhances understanding of the underlying factors contributing to MDR.