HIDDEN ACUTE KIDNEY DISEASE IN OUTPATIENT SETTINGS: MACHINE LEARNING-BASED PREDICTION AND ONE-YEAR DIALYSIS RISK

Outpatient acute kidney disease (AKDOPT) often goes undiagnosed due to fragmented kidney function data. In 2017, we launched the Acute Kidney Injury Detection System (AKIDS) to automatically integrate national and local electronic medical records to screen for AKDOPT. However, up to 40% of outpatients lack prior kidney function data, limiting AKIDS’s detection capability. This study aims to develop a machine learning (ML) model to predict AKDOPT and evaluate its clinical impact.

This retrospective cohort study utilized real-world data from adult outpatients (18-90 years) screened by AKIDS between 2017 and 2020 at a tertiary medical center in central Taiwan. Patients with prior end-stage kidney disease (ESKD), routine kidney care, or cancer history were excluded. AKDOPT was defined as a >50% increase in serum creatinine (S-Cre) or a >35% decline in eGFR within 180 days preceding the outpatient visit. For model development, Cohort 1 comprised 218,274 visits from 116,893 patients with at least two S-Cre measurements during the prior 180 days, eligible for AKDOPT detection. Repeated visits within 180 days were considered as the same episode, and each episode was treated as an independent observation. Cohort 1 was randomly divided into training, testing, and validation subsets (6:2:2). An XGBoost model was built using 177 clinicodemographic features, excluding S-Cre and eGFR. Model performance was assessed using area under the receiver operating characteristic curve (AUROC) and Bangdiwala’s B-statistic as primary metrics. For outcome validation, Cohort 2 included 490,314 patients whose first outpatient visits lacked sufficient S-Cre measurement for AKDOPT detection and were not part of Cohort 1. Logistic regression was used to estimate the 1-year risk of composite kidney outcome (CKO: ESKD, S-Cre doubling, or >40% eGFR decline) and all-cause mortality for true and predicted AKDOPT in the validation set in Cohort 1 and in Cohort 2.

About 11.8% of outpatient visits in Cohort 1 had AKDOPT. The ML model demonstrated robust predictive performance, achieving an AUROC of 0.85 and B statistic of 0.83 in the validation set (Table 1). The model showed high accuracy (0.86) and specificity (0.89) but moderate sensitivity (0.65). The top 10 predictors included prior inpatient or emergency visit, previous medical exams (chest X-ray, kidney sonography), latest serum albumin level, prior labs for albumin, calcium, BUN, or hs-CRP (yes/no), and prior AKDOPT (yes/no). Within one year for Cohort 1, patients with true AKDOPT had higher risks of CKO (odds ratio [OR], 8.98) and mortality (9.32) compared with those without AKDOPT (Table 2). When applying the ML model to predict AKDOPT, the predicted AKDOPT showed significant 1-year risks of CKO (7.57) and mortality (6.12). In Cohort 2, where true AKDOPT was unavailable, the predicted AKDOPT continued to show clinical relevance, with significant associations with CKO (5.66) and mortality (2.07).

Our findings support the utility of ML model in predicting AKDOPT even without prior kidney function data, enabling proactively risk assessment in outpatient settings. Further model tuning is needed to reduce feature complexity and enhance robustness against data drift, thereby improving generalizability.