Leveraging Explainable Artificial Intelligence to Identify Key Features required for Differentiating Clinical Neurocognitive Disorder Diagnoses using Toronto Cognitive Assessment

TL;DR

This study uses explainable AI to identify key cognitive features from a clinical test that help distinguish between different types of neurocognitive disorders.

Contribution

The novel use of XAI with SVM to identify critical cognitive features for differentiating neurocognitive disorder diagnoses using TorCA.

Findings

Classification accuracy for AD vs NC was excellent using top 5 TorCA features (0.96±0.04).

MCI or naMCI vs NC classification was very good using top 5 features (0.86±0.03 to 0.89±0.04).

Delayed Recall and Immediate Recall Trials were among the most important features identified.

Abstract

Accurate clinical diagnosis for Alzheimer's disease dementia (AD), amnestic mild cognitive impairment (aMCI), and non‐amnestic MCI (naMCI) is essential for timely management. The diagnosis is made using a range of factors including cognitive testing. Explainable artificial intelligence (XAI)‐based SHAP (SHapley Additive exPlanations) is a machine learning interpretability tool that can provide insights into specific features that drive classification decisions. We used XAI with support vector machines (SVM) to identify key cognitive features of Toronto Cognitive Assessment (TorCA), a user‐friendly cognitive assessment administered by frontline clinicians, for differentiating neurocognitive disorder diagnoses.

We used data from the Toronto Dementia Research Alliance (TDRA) database, comprising of participants with AD, aMCI, naMCI, or normal cognition (NC) seen in memory clinics across Toronto. An SVM model with radial basis function (RBF) kernel was configured with 10‐fold cross‐validation. XAI was integrated using SHAP values to identify the most important critical features contributing to the model predictions. Classification accuracies, defined as the proportion of correct classifications for each pairwise comparison, were calculated using TorCA total scores and specific features from subtests.

We included 695 participants (149 AD, 189 aMCI, 304 naMCI, and 53 NC). Classification accuracy for distinguishing AD vs NC was excellent, whether using all TorCA subtests (0.97±0.03), or the top 5 features (0.96±0.04) (Delayed Recall, Immediate Recall Trials 1 and 2, Sentence Comprehension, Benson Figure Recall), but lower (0.93±0.06) with only TorCA total score. Classification accuracy for MCI or naMCI vs NC was also very good (0.86±0.03 to 0.89±0.04) using the top 5 features.

TorCA combined with XAI can accurately differentiate common clinical neurocognitive disorder phenotypes in ambulatory settings. These findings point to specific cognitive subtests important for diagnosis and may help improve the efficiency of cognitive testing. Future studies should investigate the differentiation of other neurocognitive disorders using these tools and further validate these findings using formal neuropsychological testing.

Figures