Abstract
Computerized training has demonstrated efficacy in people with mild cognitive impairment (MCI), but benefits vary across participants. Identifying risk for poor outcomes could inform personalization. We previously identified participants who mastered this 6-task digital training program with a baseline algorithm. Here we report on prediction of durability of training gains using machine learning.
MCI (n = 84) and NC (n = 68) participants received remote, computerized functional skills training (FUNSAT™), twice a week for up to 12 weeks. Half of the MCI participants had a burst of cognitive training before skills training. Training gains for both groups had effect sizes >0.75. 30 days post-training, participants were reassessed for durability of gains (≥80 %). Machine Learning analyses, predicted durability using individual characteristics (e.g., cognitive performance), baseline task performance, or training features (e.g., amount of training, training gains per session). We examined easiest (Ticket Purchase) and hardest (Internet Banking) tasks.
Durability was 78 % for ticket and 65 % for banking. Most informative predictors of durability were training process variables: improvement in completion time for ticket and completion time end of training for banking. For ticket, three baseline measures were also informative.
FUNSAT training results were durable for most participants and process variables were the best predictors. Variables that previously identified failure in mastery did not predict durability of gains. Computerized cognitive training added to FUNSAT training did add to durability for MCI participants These results suggest that identifying predictors of different types of poor training outcomes could personalize computerized training to increase mastery and durability.