Date of Award

January 2024

Document Type

Thesis

Degree Name

Master of Public Health (MPH)

Department

School of Public Health

First Advisor

Francis P. Wilson

Second Advisor

Dennis Shung

Abstract

Acute kidney injury (AKI) increases the risk of death in hospitalized adults, and early medical chart review and involvement by a nephrologist may lead to improved patient outcomes. A Kidney Action Team—comprised of a pharmacist and a board-certified nephrologist or nephrology fellow—can serve as a remote consultation service to deliver recommendations for AKI patient diagnosis, initial work up, and care. We aimed to evaluate whether an AI-automated clinical decision support system for AKI could demonstrate predictive power comparable to that of human specialists, with regards to AKI evaluation and treatment. This study uses data from KAT-AKI (NCT04040296), a triple-masked, randomized controlled trial being conducted at two major US hospital systems. From November 2021 to December 2023, 3,000 participants were enrolled from six hospitals in Yale New Haven Health System. The median (IQR) age was 73.4 (62.2, 82.8) years, 1,411 (47.0%) were women, and 560 (18.7%) self-identified as Non-Hispanic Black. With two-thirds of this data, we trained a neural network with three hidden layers of 1,168 neurons each to jointly predict 70 recommendations of varying coarseness, collected from Kidney Action Teams. Following training and validation of the data, the final third was used as a test set, from which AUCs were calculated for all recommendations. The median (IQR) across all recommendations was 0.78 (0.72–0.89). This work demonstrates the ability of AI to generate recommendations that are fairly concordant with expert recommendations on AKI. This algorithm may be evaluated in a future trial to demonstrate benefit for patients in terms of clinical outcomes, potentially reducing resource burden and expanding the ability to provide expert-level consultation on AKI in resource-constrained settings.

Comments

This thesis is restricted to Yale network users only. It will be made publicly available on 05/07/2026

Share

COinS