Objective: This study uses a temporal convolutional network with contrastive pretraining (TCN-CL) on smart home ambient sensor data to classify older adults into healthy and mild cognitive impairment (MCI) categories. We aim to overcome limitations of traditional machine learning (ML) methods, logistic regression, and decision tree, by employing deep learning to capture complex temporal dependencies and extract high-level features. We -hypothesize that a pre-trained deep network will improve diagnostic accuracy across diverse, multi-resident environments.
Method: Participants were 137 community-dwelling older adults, classified as healthy older adults (HOA, n = 76) or individuals with MCI (n = 61). A set of 34 digital markers related to sleep, time out of home, activity level, and behavior regularity were derived over a 30-day period from ambient sensors installed in individuals' homes. Diagnosis predictions were examined with traditional ML classifiers and a deep network with contrastive loss (TCN-CL).
Results: TCN-CL significantly outperformed baseline classifiers (logistic regression and decision tree) in predicting cognitive diagnoses. TCN-CL achieved high accuracy (85%), sensitivity (.77), -specificity (.92), and a Matthews correlation coefficient (.71), demonstrating its effectiveness in classifying MCI using smart home sensor data. Baseline models performed poorly, with logistic regression showing marginal improvement over random guessing, and decision tree performing worse.
Conclusion: The TCN-CL, pretrained on a larger dataset, offered a robust approach for early cognitive decline detection using smart home ambient sensors. Continuous monitoring of subtle behavioral patterns through ambient sensor data holds potential to complement and enhance clinical decision-making, enabling accurate and timely diagnoses.
Keywords: Mild cognitive impairment; ambient sensors; cognitive decline; digital markers; smart home; temporal convolutional networks.