Development and validation of a machine learning predictive model for acute kidney injury nonrecovery at hospital discharge (Preprint)

Abstract

Some kidney functions may not be recovered at hospital discharge after acute kidney injury (AKI) exposure resulting in subsequent kidney replacement therapy and mortality. However, there is a lack of evidence about prognosticators of AKI nonrecovery to identify high-risk patients requiring post-discharge AKI care to improve long-term outcomes.

To develop and validate a machine learning model for predicting AKI nonrecovery at hospital discharge from AKI hospitalization using electronic health record data.

Data for hospitalized patients in the Acute Kidney Injury Recovery Evaluation Study were derived from a large healthcare delivery system in Taiwan between January 2011 and December 2017. Patients alive with AKI nonrecovery were used to derive and validate multiple feature predictive model. AKI nonrecovery was defined by a serum creatinine (SCr) level at hospital discharge of ≥1.5 times of pre-hospitalization baseline level. Sixty-four candidate features, including demographic characteristics, comorbidity, healthcare services utilization, laboratory values, and nephrotoxic medication use, were measured within 1 year before the index admission and during hospitalization for AKI.

The risk predictive model was derived from 8600 patients with AKI in 2010–2015 and validated with data from 2866 patients with AKI in 2016–2017. The proportion of AKI nonrecovery was 45% in both the derivation and validated cohorts. Among the top 20 important features in the predictive model, eight features had a positive effect on AKI nonrecovery prediction: AKI during hospitalization, SCr level at admission, receipt of dialysis during hospitalization, baseline comorbidity of cancer, AKI at admission, and baseline proportions of lymphocyte count, potassium, and low-density lipoprotein cholesterol. The predicted AKI nonrecovery risk model using the XGBoost algorithm achieved an area under the receiver operating characteristic curve statistic of 0.81 and discrimination with a sensitivity of 0.73 and a specificity of 0.72 in the temporal validation cohort.

The machine learning model approach can accurately predict AKI nonrecovery using routinely collected health data in practice. These results suggest multifactorial risk factors involved in AKI nonrecovery, requiring patient-centered risk assessment and promotion of post-discharge AKI care to prevent AKI complications.

Not applied