Leveraging Machine Learning for Dynamic Prediction and Mitigation of Acute Kidney Injury after Cardiac Surgery

Baylor College Of Medicine, Houston TX

Investigators

Abstract

Acute kidney injury (AKI) is the most frequent complication after cardiac surgery and is associated with a multitude of adverse outcomes including increased short- and long-term mortality, length of stay, and hospital cost. Early prediction of AKI can enable targeted interventions involving volume, inotrope, and pressor management that have been shown to improve outcomes and mitigate injury. Presently, AKI is identified via clinical parameters such as urine output and serum creatinine, however these represent late findings often manifesting after the treatment window. Furthermore, current methods for estimating AKI risk provide static predictions that are of limited clinical value in the dynamic postcardiac surgery critical care environment. In our preliminary work, our team recently demonstrated that ensemble machine learning (ML) analysis of electronic medical record (EMR) intensive care unit (ICU) data enables the early and accurate prediction of AKI after cardiac surgery. The objective of this proposal is to leverage ML based analysis of EMR data to develop new clinical decision support (CDS) tools to monitor AKI risk and suggest personalized, timely, data-driven interventions that prevent or mitigate morbidity. Our central hypothesis is that ML models can facilitate clinical decision making by accurately and dynamically predict AKI from routinely collected EMR data and enhancing clinical decision making by quantifying risk reduction of therapeutic interventions. The rationale underling this proposal is that completion will yield a clinically translatable ML-enabled CDS tool that can identify AKI earlier than clinical detection and effectively support clinical decision making to mitigate AKI after cardiac surgery. The central hypothesis will be tested by pursuing three specific aims: 1) develop personalized ML algorithms for real-time early detection of AKI; 2) develop interpretable ML algorithms for therapeutic interventions to mitigate AKI; 3) prospectively validate model predictive performance and efficacy as CDS tool. The proposed project is innovative in the application of novel ensemble ML techniques to analyze our unique Baylor College of Medicine Cardiac Surgery Database that combines validated clinical registry data with time series EMR data (hemodynamic, medication, and laboratory) during the pre-, intra- and post-operative phases of care. The proposed project is significant because early identification and optimized management of AKI will improve outcomes for patients with this highly prevalent and morbid complication. Furthermore, successful measurement of model efficacy in clinical practice will legitimize ML-enabled CDS systems thereby improving clinician buy in, as heretofore few ML-enabled CDS tools have been evaluated in real-world conditions and even fewer clinically implemented. This work will provide a framework for development and translation of ML-enabled digital health solutions that can be applied to a variety of clinical scenarios. The expected outcomes of this work are 1) the early and accurate identification of AKI; 2) the data-driven identification of optimal management strategies to mitigate AKI; 3) validation of ML-enabled clinical decision support systems in the clinical environment.