Background: Available therapies for Alzheimer's disease (AD) can only alleviate and delay the advance of symptoms, with the greatest impact eventually achieved when provided at an early stage. Thus, early identification of which subjects at high risk, e.g., with MCI, will later develop AD is of key importance. Currently available machine learning algorithms achieve only limited predictive accuracy or they are based on expensive and hard-to-collect information.
Objective: The current study aims to develop an algorithm for a 3-year prediction of conversion to AD in MCI and PreMCI subjects based only on non-invasively and effectively collectable predictors.
Methods: A dataset of 123 MCI/PreMCI subjects was used to train different machine learning techniques. Baseline information regarding sociodemographic characteristics, clinical and neuropsychological test scores, cardiovascular risk indexes, and a visual rating scale for brain atrophy was used to extract 36 predictors. Leave-pair-out-cross-validation was employed as validation strategy and a recursive feature elimination procedure was applied to identify a relevant subset of predictors.
Results: 16 predictors were selected from all domains excluding sociodemographic information. The best model resulted a support vector machine with radial-basis function kernel (whole sample: AUC = 0.962, best balanced accuracy = 0.913; MCI sub-group alone: AUC = 0.914, best balanced accuracy = 0.874).
Conclusions: Our algorithm shows very high cross-validated performances that outperform the vast majority of the currently available algorithms, and all those which use only non-invasive and effectively assessable predictors. Further testing and optimization in independent samples will warrant its application in both clinical practice and clinical trials.
Keywords: Alzheimer’s disease; clinical prediction rule; machine learning; mild cognitive impairment; personalized medicine.