Time Series AI Model for Acute Kidney Injury Detection Based on a Multicenter Distributed Research Network:...

Area of Science:

• Medical Informatics
• Nephrology
• Artificial Intelligence

Background:

• Acute kidney injury (AKI) is a critical indicator of patient deterioration and a significant cause of renal toxicity.
• Existing prediction models for AKI are limited, especially those utilizing time-series data from distributed research networks (DRNs).

Purpose of the Study:

• To develop and validate an interpretable long short-term memory (LSTM)-based model for early detection of AKI.
• To apply the model to hospital electronic health record (EHR)-based time-series data in patients administered nephrotoxic drugs within a DRN.

Main Methods:

• A multi-institutional retrospective cohort study involving 6 hospitals and a DRN.
• Development of an interpretable multivariable LSTM (IMV-LSTM) model using patient data and propensity score matching.
• Analysis of temporal attention values to identify key predictive variables for AKI.

Main Results:

• The study analyzed 8,643 AKI patients and 31,012 controls across 6 hospitals.
• The IMV-LSTM model demonstrated strong predictive performance, with the highest AUC for acyclovir (0.94).
• Lymphocytes and calcium showed high attention values, while other variables like albumin and hemoglobin decreased over time, indicating potential AKI markers.

Conclusions:

• An interpretable LSTM model applied to EHR data via a DRN enables early surveillance of AKI.
• This approach aids in identifying risk factors and facilitates early detection of adverse drug reactions related to nephrotoxic medications.