Interpretable Sub-phenotype Identification in Acute Kidney Injury.

Area of Science:

• Nephrology
• Computational Biology
• Data Mining

Background:

• Acute kidney injury (AKI) is a complex, life-threatening condition requiring personalized treatment.
• Effective AKI sub-phenotyping is essential for understanding pathophysiology and developing targeted interventions.
• Current clustering methods for AKI sub-phenotyping lack interpretability and specificity.

Purpose of the Study:

• To introduce a pattern mining approach using a multiobjective evolutionary algorithm (MOEA) for AKI sub-phenotyping.
• To generate explicit, interpretable rules for AKI sub-phenotypes.
• To discover specific sub-phenotypes for AKI vs. non-AKI and mild vs. moderate-to-severe AKI.

Main Methods:

• Utilized a pattern mining approach combined with a multiobjective evolutionary algorithm (MOEA).
• Focused on generating explicit rules for AKI sub-phenotypes, eliminating the need for post-hoc explanations.
• Employed the method to efficiently search feature subspaces for minor and highly specific sub-phenotypes.

Main Results:

• Identified 174 (178) significant sub-phenotypes for AKI vs. non-AKI and AKI-2/3 vs. AKI-1 classifications.
• Achieved an average confidence of 0.33 (0.33) for the discovered sub-phenotypes.
• Demonstrated superior patient assignment confidence compared to k-means clustering, with an average improvement of 0.20 (0.23).

Conclusions:

• The proposed pattern mining and MOEA approach offers a more interpretable and specific method for AKI sub-phenotyping.
• This technique facilitates the discovery of clinically relevant patient subgroups for improved AKI management.
• The enhanced confidence in patient assignment suggests potential for more precise and effective personalized AKI treatments.