Identifying acute kidney injury subphenotypes using an outcome-driven deep-learning approach.

Area of Science:

• Nephrology
• Data Science
• Intensive Care Medicine

Background:

• Acute kidney injury (AKI) is a common and serious condition in intensive care units (ICUs).
• Current AKI management overlooks patient heterogeneity, hindering targeted interventions.
• Identifying distinct AKI subphenotypes is crucial for understanding pathophysiology and improving patient outcomes.

Purpose of the Study:

• To develop and validate a data-driven deep-learning approach for identifying AKI subphenotypes.
• To analyze the prognostic and therapeutic implications of identified AKI subphenotypes.
• To leverage time-series electronic health record (EHR) data for AKI subphenotype discovery.

Main Methods:

• Developed a supervised long short-term memory (LSTM) autoencoder (AE) to extract relevant representations from time-series EHR data.
• Utilized K-means clustering on extracted representations to identify distinct AKI subphenotypes.
• Validated the approach on two independent, publicly available datasets.

Main Results:

• Identified three distinct AKI subphenotypes across both datasets.
• Observed significantly different mortality rates among the identified subphenotypes (e.g., 11.3%, 17.3%, 96.2% in one dataset).
• Demonstrated statistical significance of identified subphenotypes across various clinical characteristics and outcomes.

Conclusions:

• The proposed deep-learning approach successfully clusters ICU patients with AKI into three distinct subphenotypes.
• This subtyping has the potential to enhance risk assessment and guide personalized treatment strategies for AKI patients.
• Further research can build upon this methodology to improve AKI patient management and outcomes in critical care settings.