Deep Learning Based Time-to-Event Prediction of MCI Transition Leveraging Sensor-captured Daily Activities Data
Rajaram Narasimhan, Muthukumaran Gopalan, Mathangi Damal Chandrasekhar · International Journal of Engineering Trends and Technology · 2023
Mild Cognitive Impairment (MCI) is known to be a condition in older adults presenting cognitive impairment symptoms in the absence of functional impairment. This could be a transitional stage to developing Alzheimer’s disease (AD), but it does not always lead to AD. Early detection of MCI symptoms is vital in determining personalized interventions to slow down the MCI progression. This study proposes a survival analysis approach based on deep learning techniques to predict the probability of an individual transitioning from a cognitively healthy stage to MCI at a given time point by utilizing activity data captured through unobtrusive sensors by continuously monitoring the older adults’ daily routines. The performance of the proposed models, Neural Multi-Task Logistic Regression and Non-Linear Cox, in predicting the probability of time-to-transition are examined and compared against the Standard Cox PH model. The two well-known metrics, Concordance Index (CI) and Integrated Brier Score (IBS), are used to evaluate the model performance. Additionally, the features are ranked based on the model-learned weights and results are interpreted. Deep learning-based models perform better than the standard Cox PH model, with the best average CI of 0.714 and IBS of 0.119. The results suggest that the proposed models can accommodate the nonlinear elements from the data and account for the fact that the rate of progression of two individuals will vary with time. Feature ranking reveals the age and years of education to be in the top 5, in addition to features from sleep and mobility domains which are clinically meaningful. This study demonstrates that a practical, less expensive, and non-invasive way of observing older adults’ activity routines coupled with computing advancements such as deep learning techniques offer phenomenal opportunities for early detection of MCI transition.