Causally informed explainable artificial intelligence for cognitive risk stratification in hearing-impaired adults

Abstract Background Hearing loss has been increasingly recognized as a modifiable risk factor for cognitive decline and dementia. Within otorhinolaryngology and audiology clinics, early identification of patients at elevated cognitive risk remains challenging. Although artificial intelligence (AI) is increasingly applied in ENT diagnostics, most existing models emphasize predictive accuracy without mechanistic validation or interpretability, thereby limiting clinical integration. Objective To develop a causally informed, explainable AI framework for identifying cognitive vulnerability in hearing-impaired adults using real-world otologic data. Methods We analyzed 581 subjects from the Oldenburg Hearing Health Repository (OHHR), comprising 520 cognitively normal individuals and 61 individuals with cognitive impairment. Audiometric, cognitive, demographic, and health-related variables were incorporated for analysis. Multi-modal harmonization was performed prior to analysis. Causal structure learning was conducted using Fast Greedy Equivalence Search (FGES) with bootstrap validation to identify stable directional relationships. A LightGBM classifier was subsequently trained for cognitive impairment prediction. Model performance was evaluated using the area under the receiver operating characteristic curve (ROC-AUC), Matthews Correlation Coefficient (MCC), calibration analysis, and subgroup performance by sex. Model interpretability was assessed using Shapley Additive Explanations (SHAP). Results Bootstrap-validated causal discovery identified age as the most consistent upstream determinant of cognitive vulnerability, with relationships observed among age, auditory burden measured by pure-tone average (PTA), and systemic health indicators. The predictive LightGBM model demonstrated good discriminative performance (ROC-AUC = 0.89; MCC = 0.30; F1 = 0.33). SHAP explainability highlighted age, pure-tone average (PTA), and physical health indices (PCS and MCS) as dominant contributors to prediction. Calibration analysis indicated reasonable reliability within low-risk probability ranges. Subgroup analysis showed broadly comparable predictive behavior across sex groups, although class imbalance limited stable estimation of certain subgroup metrics. Conclusion A causally informed and explainable AI framework can enhance understanding of cognitive vulnerability in hearing-impaired adults. By integrating structural inference with predictive modeling, this approach supports development of transparent decision-support tools in audiology and otorhinolaryngology practice.