An ML-powered, web-interface tool for Robust, Efficient, Affordable Diagnosis of Dementia
READi-Dem is a web-based application to aid in dementia screening, using cutting-edge machine learning techniques on a unique population-level dataset. We first identified the most strongly-predictive features for dementia diagnosis from our dataset, and further refined these to a set of 27 minimal features. These can be easily obtained by a patient and their caregiver, without the need for any training or time consuming cognitive tests. We then tested a range of ML models, and found that the support vector machine provided the highest classification accuracy. The trained model was incorporated into our app, which uses questionnaire answers to predict the likelihood of a dementia diagnosis for the individual.
Dementia screening tools typically involve face-to-face cognitive testing. In the face of growing demand, there is an increasing burden on clinical staff, particularly in low-resource settings. The objective of our study was to build an integrated online platform for efficient dementia screening, using a brief and cost-effective assessment.
We used the Longitudinal Ageing Study in India dataset (LASI-DAD, n = 2528) to predict dementia diagnosis based on the Clinical Dementia Rating (CDR). Using feature selection algorithms and principal component analysis (PCA), we identified key predictive features. We compared performance of six machine learning classifiers when trained on the 42 selected features (full model) and the two components identified by PCA (minimal model). The best-performing model was selected for our web platform.
Selected features mapped onto two distinct, interpretable domains: a cognitive domain and an informant domain. The first two principal components cumulatively explained 90.2\% of the variance and included questions from the Mini-Mental State Exam (MMSE) and the Informant Questionnaire on Cognitive Decline in the Elderly (IQCODE). Classifiers trained on the minimal model performed on par with the full model, with Support Vector Machine performing best (93.4\%). The model did not reliably predict Parkinson’s disease (67\% accuracy) or stroke (53.1\% accuracy), suggesting dementia specificity. The respective questions from MMSE and IQCODE (27 items) were incorporated into our online platform.
We built an online platform enabling end-to-end screening for dementia from assessment to prediction, based on patient and caregiver reports. The web link to access the app is available on the top of this page. This research was published by MedRxiv available at the below link.
Verena Klar, Elinor Thompson, Melvin Selim Atay, Peter Owoade, Sofia Toniolo, Amir Dehsarvi, Sanjay Rathee, "READi-Dem: ML-powered, web-interface tool for Robust, Efficient, Affordable Diagnosis of Dementia", medRxiv 2023.10.23.23297405,doi: https://doi.org/10.1101/2023.10.23.23297405