Abstract
BACKGROUND: Voice biomarkers hold potential for early cognitive disorder detection, but variations in recording conditions across different environments present challenges for accurate diagnosis using artificial intelligence (AI) models. This study aims to develop a robust, generalizable model for reliably diagnosing cognitive impairments across varied datasets. METHODS: We implemented a domain generalization approach using an adapted Deep Domain-Adversarial Image Generation (DDAIG) framework. This method transforms input data to reduce center-specific characteristics and emphasizes domain-invariant features, allowing the model to focus on cognitive impairment indicators. RESULTS: Before applying domain generalization, both cognitive impairment (CI) and center classification models achieved accuracies of 0.96. After implementing domain generalization, the CI classification accuracy decreased to 0.90, while the center classification model's accuracy dropped to 0.64. This reduction in the center classification metrics reflects the model's reduced dependence on center-specific features, indicating effective domain generalization. CONCLUSION: The adapted DDAIG framework effectively reduced center-specific learning, enhancing the model's ability to generalize cognitive impairment classifications across different centers. These findings suggest the role of domain generalization in developing reliable AI diagnostic tools for cognitive disorder detection.