Development and Validation of a Personalized Model With Transfer Learning for Acute Kidney Injury Risk Estimation Using Electronic Health Records

Abstract

Importance: Acute kidney injury (AKI) is a heterogeneous syndrome prevalent among hospitalized patients. Personalized risk estimation and risk factor identification may allow effective intervention and improved outcomes.

Objective: To develop and validate personalized AKI risk estimation models using electronic health records (EHRs), examine whether personalized models were beneficial in comparison with global and subgroup models, and assess the heterogeneity of risk factors and their outcomes in different subpopulations.

Design, setting, and participants: This diagnostic study analyzed EHR data from 1 tertiary care hospital and used machine learning and logistic regression to develop and validate global, subgroup, and personalized risk estimation models. Transfer learning was implemented to enhance the personalized model. Predictor outcomes across subpopulations were analyzed, and metaregression was used to explore predictor interactions. Adults who were hospitalized for 2 or more days from November 1, 2007, to December 31, 2016, were included in the analysis. Patients with moderate or severe kidney dysfunction at admission were excluded. Data were analyzed between August 28, 2019, and May 8, 2022.

Exposures: Clinical and laboratory variables in the EHR.

Main outcomes and measures: The main outcome was AKI of any severity, and AKI was defined using the Kidney Disease: Improving Global Outcomes serum creatinine criteria. Performance of the models was measured with area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve, and calibration.

Results: The study cohort comprised 76 957 inpatient encounters. Patients had a mean (SD) age of 55.5 (17.4) years and included 42 159 men (54.8%). The personalized model with transfer learning outperformed the global model for AKI estimation in terms of AUROC among general inpatients (0.78 [95% CI, 0.77-0.79] vs 0.76 [95% CI, 0.75-0.76]; P < .001) and across the high-risk subgroups (0.79 [95% CI, 0.78-0.80] vs 0.75 [95% CI, 0.74-0.77]; P < .001) and low-risk subgroups (0.74 [95% CI, 0.73-0.75] vs 0.71 [95% CI, 0.70-0.72]; P < .001). The AUROC improvement reached 0.13 for the high-risk subgroups, such as those undergoing liver transplant and cardiac surgery. Moreover, the personalized model with transfer learning performed better than or comparably with the best published models in well-studied AKI subgroups. Predictor outcomes varied significantly between patients, and interaction analysis uncovered modifiers of the predictor outcomes.

Conclusions and relevance: Results of this study demonstrated that a personalized modeling with transfer learning is an improved AKI risk estimation approach that can be used across diverse patient subgroups. Risk factor heterogeneity and interactions at the individual level highlighted the need for agile, personalized care.

Figure 1.. Comparison of Model Performance in General Inpatients

In panels B and C, personalized models used 10% of overall training sample as the threshold for number of similar patients. Global model used 100% of training samples. AKI indicates acute kidney injury; AUROC, area under the receiver operating characteristic curve.

Figure 2.. Radar Chart of the Area Under the Receiver Operating Characteristic Curve (AUROC) for Personalized and Subgroup Models Across 20 High-Risk Subgroups

OR indicates operating room.

Figure 3.. Heatmaps of Outcomes of Top 20 Global Model Predictors Across 20 High-Risk Subgroups

A, Relative effect was calculated as follows: (area under the curve [AUC] gain of predictor when global model was used in subgroup − AUC gain of predictor when global model was used in whole population) / (AUC gain of predictor when global model was used in whole population). Red represents increased and blue represents decreased predictive effect in subgroups vs whole population. B, Relative effect was calculated as follows: (AUC gain of predictor when personalized model with transfer learning was used in subgroup − AUC gain of predictor when global model was used in subgroup) / (AUC gain of predictor when global model was used in general patients). Red represents increased and blue represents decreased predictive effect in personalized model with transfer learning vs global model. Other race and ethnicity included American Indian or Alaskan Native, Native Hawaiian or Other Pacific Islander, 2 races, and unreported race. AST indicates aspartate aminotransferase; BMI, body mass index; CHF, congestive heart failure; DMV, durative mechanical ventilation; WBC, white blood cell.

Figure 4.. Outcomes of Top 20 Personalized Model With Transfer Learning Predictors Across 15 Large Diagnostic Subgroups

To improve the stability of results, the 63 subgroups in Figure 4B and C were further abstracted into 15 large subgroups, both full name of the 63 diagnostic subgroups and their classification are provided in eTable 15 in the Supplement. Predictors outcome in each of the 63 subgroups are presented in eFigures 16 and 17 in the Supplement. AMI indicates acute myocardial infarction; AST, aspartate aminotransferase; AUC, area under the curve; BMI, body mass index; DMV, durative mechanical ventilation; WBC, white blood cell.