LSTM machine learning of longitudinal clinical data accurately predicts acute kidney injury onset in COVID-19: a two-center study.

ABSTRACT:

Background

To employ long-short-term-memory (LSTM) artificial-intelligence method to model multiple time points of clinical laboratory data, along with demographics and comorbidities, to predict hospital-acquired acute kidney injury (AKI) onset in COVID-19 patients.

Methods

Montefiore Health System data consisted of 1,982 AKI and 2,857 no-AKI (NAKI) hospitalized COVID-19 patients and Stony Brook Hospital validation data consisted of 308 AKI and 721 NAKI hospitalized COVID-19 patients. Demographic, comorbidities, and longitudinal (3 days before AKI onset) laboratory tests were analyzed. LSTM was used to predict AKI with five-fold cross-validation (80%/20% for training/validation).

Results

The top predictors of AKI onset were glomerular filtration rate, lactate dehydrogenase, alanine aminotransferase, aspartate aminotransferase and C-reactive protein. Longitudinal data yielded marked improvement in prediction accuracy over individual time points. Inclusion of comorbidities and demographics further improves prediction accuracy. The best model yielded an AUC, accuracy, sensitivity and specificity, respectively, to be 0.965±0.003, 89.57±1.64%, 0.95±0.03, and 0.84±0.05 for the Montefiore validation dataset, and 0.86±0.01, 83.66±2.53%, 0.66±0.10, 0.89±0.03 for the Stony Brook Hospital validation dataset.

Conclusions

LSTM model of longitudinal clinical data accurately predicted AKI onset in COVID-19 patients. This approach could help heighten awareness for AKI complications and identify patients for early interventions to prevent long-term renal complications.

SUBMITTER: Lu J 

PROVIDER: S-EPMC9303068 | biostudies-literature | 2022 Jul

REPOSITORIES: biostudies-literature