Predicting in-hospital outcomes of patients with acute kidney injury

Abstract

Acute kidney injury (AKI) is prevalent and a leading cause of in-hospital death worldwide. Early prediction of AKI-related clinical events and timely intervention for high-risk patients could improve outcomes. We develop a deep learning model based on a nationwide multicenter cooperative network across China that includes 7,084,339 hospitalized patients, to dynamically predict the risk of in-hospital death (primary outcome) and dialysis (secondary outcome) for patients who developed AKI during hospitalization. A total of 137,084 eligible patients with AKI constitute the analysis set. In the derivation cohort, the area under the receiver operator curve (AUROC) for 24-h, 48-h, 72-h, and 7-day death are 95·05%, 94·23%, 93·53%, and 93·09%, respectively. For dialysis outcome, the AUROC of each time span are 88·32%, 83·31%, 83·20%, and 77·99%, respectively. The predictive performance is consistent in both internal and external validation cohorts. The model can predict important outcomes of patients with AKI, which could be helpful for the early management of AKI.

Fig. 1. Flow chart of the study population selection.

We selected patients who developed AKI during hospitalization for further screening. The exclusion criteria were as follows: (1) patients who had less than two SCr results during hospitalization; (2) patients <18 years old; (3) patients who had HIV or AIDS; and (4) patients who had end-stage kidney disease (ESKD, defined as maintenance dialysis, kidney transplantation, or eGFR <15 ml/min per 1.73m2). SCr, serum creatinine. HIV human immunodeficiency virus, AIDS acquired immunodeficiency syndrome.

Fig. 2. Various evaluation indicators of the prediction model of death.

It showed the AUROC curves (a), accuracy (b), F-score (c), precision (d), and recall (e) for predicting 24-h, 28-h, 72-h, and 7-day mortality in derivation, internal, and external validation cohorts. In this study, we train the deep learning model in derivation cohort and test in internal validation cohort with 100 epochs and validate in external validation cohort to obtain the results.

Fig. 3. Various evaluation indicators of the prediction model of dialysis.

It showed the AUROC curves (a), accuracy (b), F-score (c), precision (d), and recall (e) for predicting 24-h, 28-h, 72-h, and 7-day dialysis in derivation, internal, and external validation cohorts. In this study, we train the deep learning model in derivation cohort and test in internal validation cohort with 100 epochs and validate in external validation cohort to obtain the results.

Fig. 4. Prediction of death and dialysis in Subgroup cohorts.

We conducted subgroup analyze of age, gender, hypertension, diabetes, AKI stage, baseline SCr, length of ICU stay, and major surgery. X-axis was the value of AUROC, and Y-axis was model predicting death and dialysis at 24-h, 28-h, 72-h, and 7-day.

Fig. 5. Conceptual model of continuously predicting AKI in-hospital outcomes.

First, we collect the patient information, such as laboratory, procedure, medication, etc. Second, we construct a sequential representation of electronic health records by merging patient data in 24 h. Third, we propose AKIEPM based on deep learning. Fourth, we predict the occurrence of death or need for dialysis at 24 h, 48 h, 72 h, and 7d.

Fig. 6. Flow chart of the construction of derivation, internal validation, and external validation cohorts.

To train and validate the performance of AKIEPM, we divided the patients into the derivation, internal validation, and external validation cohorts with a 7:2:1 ratio. For the external validation cohort, we chose three hospitals wherein the number of patients was the closest to 10% of the overall cohort (14,610 patients). From patients in the remaining hospitals, 27,217 patients (20% of the total) were randomly selected as the internal validation cohort, then all other patients (~70% of the total) were selected as the derivation cohort.

Fig. 7. Schematic of the merge rules in the AKI event prediction model AKIEPM.

It showed the situations of no SCr result within 24 h (1), one SCr result within 24 h (2), and two SCr results within 24 h (3), and (4) for constructing a sequential representation of electronic health records by merging patient data in 24 h.