Explainable Machine Learning-Based Classification of Clinical Dementia Ratings Using a Tablet-Based Cognitive Assessment

Dementia is a substantial public health issue with increasing prevalence globally. Fast and accurate diagnosis is vital for successful management and intervention. This article presents a new machine-learning model that uses data from a tablet computer cognitive test to label clinical dementia ratings (CDR). Traditional cognitive assessments often depend on subjective evaluations, which may introduce bias and limit applicability across diverse populations. We evaluated five different machine-learning algorithms for dementia classification, integrating advanced feature extraction techniques, and to enhance the interpretability of our model, we employed explainable artificial intelligence (XAI) techniques, specifically local interpretable model-agnostic explanations (LIME) and Shapley Additive exPlanations (SHAP) with ensemble methods to capture subtle cognitive deficits and these methods elucidate the contributions of individual features to the model’s predictions, fostering trust and transparency in clinical applications. Area Under the Receiver-Operating Characteristic (AUROC) is used to measure the model’s ability to distinguish between classes across various threshold settings. Among different ML models, the deep learning model demonstrated superior predictive performance, achieving an accuracy of 100% and 95.8% for training and testing, respectively. Additionally, it attained area under the receiver-operating characteristic (AUROC) scores of 1.00 (95% CI 1.00–1.00) for training and 0.98 (95% CI 0.92–1.00) for testing, indicating strong discrimination between different CDR levels. Our analysis identified features, such as delay, mini-mental state examination (MMSE) score, gender, yong_overtimes, and mouse_times as the top five predictors. On the other hand, angle degree, spiral tremble, and yong tremble were the least predictive features. These findings are consistent with the previous literature, highlighting the importance of these features in the assessment of dementia. Our approach illustrates the promise of fusing state-of-the-art machine-learning algorithms with explainability methods to enhance the accuracy and interpretability of dementia classification. Such integration is central to the development of trustworthy digital tools that can support early diagnosis and management of dementia, eventually leading to improved patient outcomes.