Predicting Alzheimer's Disease from Spoken and Written Language Using Fusion-Based Stacked Generalization

The importance of automating the diagnosis of Alzheimer disease (AD) towards facilitating its early prediction has long been emphasized, hampered in part by lack of empirical support. Given the evident association of AD with age and the increasing aging population owing to the general well-being of individuals, there have been unprecedented estimated economic complications. Consequently, many recent studies have attempted to employ the language deficiency caused by cognitive decline in automating the diagnostic task via training machine learning (ML) algorithms with linguistic patterns and deficits. In this study, we aim to develop multiple heterogeneous stacked fusion models that harness the advantages of several base learning algorithms to improve the overall generalizability and robustness of AD diagnostic ML models, where we parallelly utilized two different written and spoken-based datasets to train our stacked fusion models. Further, we examined the effect of linking these two datasets to develop a hybrid stacked fusion model that can predict AD from written and spoken languages. Our feature spaces involved two widely used linguistic patterns: lexicosyntactics and character n-gram spaces. We firstly investigated lexicosyntactics of AD alongside healthy controls (HC), where we explored a few new lexicosyntactic features, then optimized the lexicosyntactic feature space by proposing a correlation feature selection technique that eliminates features based on their feature-feature inter-correlations and feature-target correlations according to a certain threshold. Our stacked fusion models establish benchmarks on both datasets with AUC of 98.1% and 99.47% for the spoken and written-based datasets, respectively, and corresponding accuracy and F1 score values around 95% on spoken-based dataset and around 97% on the written-based dataset. Likewise, the hybrid stacked fusion model on linked data presents an optimal performance with 99.2% AUC as well as accuracy and F1 score falling around 97%. In view of the achieved performance and enhanced generalizability of such fusion models over single classifiers, this study suggests replacing the initial traditional screening test with such models that can be embedded into an online format for a fully automated remote diagnosis.