A-87: Investigating a Boolean Approach to the Diagnosis of Neurodegenerative Diseases

Abstract

Objective: Given the success of machine-learning techniques at improving diagnostic practice in other branches of medicine, the current study evaluated the utility of three decision-tree models—CART, CHAID, and Random Forests (RF)—to redesign neuropsychological assessment batteries using an if–then approach to assessment administration and interpretation. The goal of the study was to shorten testing procedures and improve diagnostic accuracy for neurodegenerative diseases. Method/Data Selection: Data were derived from the Unified Dataset of the National Alzheimer’s Coordinating Center, funded by NIA/NIH Grant U24AG072122. Analyses were restricted to participants with normal cognition, Alzheimer’s disease, vascular cognitive impairment, frontotemporal dementias, or Lewy body diseases. CART, CHAID, and RF models were trained to identify etiological class using assessment measures as predictors. Models separated data using an 80/20 stratified split. Ten-fold cross-validation was used to tune hyperparameters and prevent overfitting. Cohen’s kappa coefficients were averaged to quantify predicted and actual classification congruence, and model discriminative ability was examined via multiclass AUC. Results/Data Synthesis: An RF model outperformed all other models, achieving a classification accuracy of 81%, significantly higher than the chance rate of 57% (p < 0.001). The mean Cohen’s kappa indicated substantial agreement beyond chance (k = 0.62), and multiclass AUC (0.79) revealed good discriminative ability. Conclusions: Findings support the feasibility of using RF models to improve the design and interpretation of assessment batteries. Results of the current study cast RF as a useful tool for developing abbreviated batteries, as importance values and cutoff scores are derived from whole-battery analysis, thereby accounting for potential inter-test effects that impact sensitivity, specificity, and predictive value of assessment measures.