Differentiating patients with dementia from patients with depression and healthy controls using the Cognitive Functions Dementia (CFD) test set and machine learning.

Objective: Due to similarities in cognitive impairments shown in early states of dementia (DEM) and depression (DEP), accurate differentiation between the two in elderly remains challenging in clinical practice. Using Machine Learning (ML) algorithms in addition to the gold standard of neuropsychological assessment could help the differentiation between healthy controls (HC) and DEM, as well as between DEM and DEP by providing a diagnostic rule for clinicians. Methods: We used four different ML algorithms (SVM: Support Vector Machine, NB: Gaussian Naïve Bayes, RF: Random Forest, GLMnet: Lasso and Elastic-Net Regularized Generalized Linear Models) and logistic regression (LR) to differentiate between HC (n = 407), patients with DEM (n = 131) and patients with DEP (n = 145) using features from the tablet-based neuropsychological test battery Cognitive Functions Dementia (CFD). We also investigated whether the type of input data (i.e. raw scores vs. sociodemographically adjusted raw scores or T-scores) influences the classification accuracy. Results: Using raw data from the CFD and the GLMnet algorithm, we could accurately differentiate between DEM vs. HC with accuracies ranging up to 94.0%. Similarly, we could classify DEM and DEP with accuracies up to 80.8% using the Naïve Bayes algorithm and raw scores. Measures for verbal memory, word fluency and processing speed showed the highest feature importance within these classifications, highlighting their importance for differential diagnosis. Conclusions: We provide preliminary evidence that ML algorithms in combination with the CFD can aid clinicians in the differential diagnosis of HC and DEM, as well as DEM and DEP by providing a decision-making aid.