Decoding the renal-cochlear axis: explainable machine learning and phenotype clustering reveal high-risk hearing loss...

Area of Science:

• Nephrology and Audiology
• Artificial Intelligence in Healthcare
• Biostatistics

Background:

• Chronic kidney disease (CKD) is associated with a high prevalence of hearing loss (HL).
• Accurate risk stratification and phenotyping are crucial for managing HL in CKD patients.
• Existing methods lack precision in identifying at-risk CKD populations for hearing impairment.

Purpose of the Study:

• To develop and validate a dual-level machine learning (ML) framework for hearing loss risk stratification and phenotyping in chronic kidney disease patients.
• To identify key predictors of hearing loss in the CKD population.
• To create a clinically applicable tool for early detection and intervention.

Main Methods:

• Utilized data from 3,402 chronic kidney disease patients from the National Health and Nutrition Examination Survey (NHANES).
• Employed feature selection via logistic regression, constructing predictive models using nine ML algorithms, with eXtreme Gradient Boosting (XGBoost) showing superior performance.
• Applied SHapley Additive exPlanations (SHAP) for model interpretability and Gaussian Mixture Modeling (GMM) for patient subtyping.

Main Results:

• The XGBoost model achieved high discrimination (AUC = 0.984 training, 0.939 testing).
• Age was identified as the primary risk determinant for hearing loss.
• Two distinct patient subtypes were identified: low-risk (1.58% HL prevalence) and high-risk (48.2% HL prevalence), characterized by specific clinical factors.
• A web-based tool incorporating six key features was developed.

Conclusions:

• The developed dual-level ML framework effectively integrates explainable AI and unsupervised clustering for hearing loss risk assessment in CKD.
• This approach enables precise screening of high-risk CKD subpopulations for hearing loss.
• The findings support integrating hearing assessments into routine CKD care for proactive management.