Dementia screening tools typically involve resource-costly face-to-face cognitive testing. The objective of our study was to build an integrated online platform for efficient dementia screening. Key criteria were the use of information that would be easy and cost-effective to obtain in clinical settings and not require cognitive testing, while also being specific to dementia.
We used the Longitudinal Ageing Study in India dataset (LASI-DAD) to predict general dementia measured by the Clinical Dementia Rating (CDR).
First, we identified key predictive features for dementia diagnosis, prioritising information that can be easily obtained using multiple feature selection algorithms such as Boruta and Elastic Net. The 42 chosen features mapped onto two distinct cognitive and informant domains which we further reduced to two compound scores using PCA. We trained 6 standard machine learning classifiers (Gradient Boosting Machine, Neural Net, Random Forest, Support Vector Machine, Logistic Regression, Cartesian Genetic Programming) on the 42 selected features (full model) and the two compound features (minimal model) and compared accuracies. Using RShiny, we created a web platform to apply the model to new datapoints.