A Counterfactual Verified Semi-Supervised Learning Framework for Older Adults' Functional and Cognitive Health Assessment

Author/Creator ORCID




Information Systems


Information Systems

Citation of Original Publication


This item may be protected under Title 17 of the U.S. Copyright Law. It is made available by UMBC for non-commercial research and education. For permission to publish or reproduce, please see http://aok.lib.umbc.edu/specoll/repro.php or contact Special Collections at speccoll(at)umbc.edu
Access limited to the UMBC community. Item may possibly be obtained via Interlibrary Loan through a local library, pending author/copyright holder's permission.


Recent advances in Internet-of-Things (IoT) devices in conjunction with artificial intelligence (AI) and machine learning (ML) algorithms have accelerated research in the early detection of an underlying neurodegenerative disorder by monitoring daily activities using accelerometry. Therefore, this thesis investigates the underpinning relationships between daily activities and older adults' functional and cognitive health impairments while addressing various challenges arising from sensing, modeling, and inferencing micro and macro activities using wearable accelerometry and deep learning algorithms. To support this effort, we performed an extensive data collection drive to gather sensor-based activity data and survey-based clinical assessment on functional, behavioral, and cognitive health from 25 older adults residing in their homes in a retirement community center. To address sensing-related challenges (abundant unlabeled and scarcity of labeled data), particularly the availability of an abundance of unlabeled data in a practical setting, we posit a self-taught learning approach that leverages a pre-training phase to learn representations from unlabeled data. The labeled data representations are then projected onto the newly learned representation space, which helps reduce the missed detection of activities by 20%. Furthermore, we leverage the similarities and disparities between those representations using supervised and non-parametric contrastive learning approaches to make our activity recognition model user-invariant and scalable (F1 score 92%). Next, we investigate the interrelation between activities and underlying cognitive health impairments. We propose a novel contrastive multi-task learning framework to concurrently estimate activity labels and underlying cognitive (Dementia, Mild Cognitive Impairment, Normal) or activity performance scores that indicate the health impairment status. This contrastive multi-task learning approach improves the detection of activities, activity performance score, and stage of dementia to 92%, 97%, and 98% (F1-score), respectively. Besides, we defined a novel impairment indicator, computed based on the model error capable of alerting when a change in cognitive or functional health status is at an early stage. Since the impairment indicator relies on the model error, it is imperative to explain that the model error is caused due to health impairment such as Dementia or Mild Cognitive Impairment. Therefore, we rely on computing domain-specific counterfactuals to confuse the model, observe the model performance, and explain the causation problem. Upon introducing 50% counterfactual representations during inferencing, we observed a 14% improvement in the classification performance, which proves that the error is not due to the model's inability but the underlying health impairment.