Development and Validation of Transportable, Clinically Applicable and Scalable Machine Learning Models for Acute Kidney Injury

Abstract

Acute kidney injury (AKI), a frequent complication in hospitalized patients, poses significant challenges due to its high incidence, short-term mortality, and substantial economic burden. Current AKI models utilizing electronic health records (EHR) and machine learning (ML) confront limitations in external validation, the exclusion of urine output as a predictor, and a predominant reliance on single-center data. In this dissertation, I present a comprehensive exploration of ML applications for AKI, with a focus on crucial dimensions such as transportability, clinical applicability, and scalability. In Chapter II, I reproduce and evaluate the transportability of a leading AKI model originally developed by DeepMind for the veterans. Despite the model's high performance in predicting AKI, the predominantly male population on which it is trained have led to questions about its generalizability in other cohorts. I reproduce key aspects of their GBDT model and assess its performance in a sex-balanced patient population at the UM, revealing suboptimal discrimination and calibration in females. A continued training approach at UM partially addresses model differential performance in sex. An exploration of potential reasons for this model discrepancy by sex reveals that it is complex and cannot be simply explained by a low sample size or difference in patient characteristics. This study demonstrates that local fine-tuning may be a promising solution for mitigating sex and gender inequalities in healthcare ML models. In Chapter III, I investigate the urine output (UO) documentation pattern in the EHR and assess the role of UO as an AKI predictor. Analysis of a five-year inpatient cohort at UM reveals frequent and diverse UO documentation for non-ICU patients. Despite its value, the inclusion of UO as a predictor minimally improves the ability to predict AKI over a comprehensive model without UO. This study emphasizes the ongoing need for refining UO documentation practices to augment its clinical utility. In Chapter IV, I introduce a novel Federated Stacked Learning (FSL) framework to enhance the scalability of AKI models in multicenter settings where data sharing may not be permitted. Focusing on predicting cardiac surgery-associated AKI within a national perioperative research network, the study compares the performance of single-center models with both a pooled model and the proposed FSL approach. The single-center models perform worse than the multicenter approaches. The FSL approach demonstrates comparable performance with pooled models, suggesting that it is a practical alternative when patient-level data sharing is not an option. The study underscores the significance of collaborative research networks and illustrates how the size of both the hospital and the network can influence the optimal modeling strategy. Collectively, this dissertation contributes valuable insights into AKI prediction, advocating for a pragmatic model development approach encompassing transportability, clinical applicability, and scalability. The findings pave the way for future advancements in ML applications for AKI, promoting the development of models that are not only accurate but also accessible, generalizable, and adaptable across diverse healthcare settings.