Anomaly Detection and Pattern Recognition in Cognitive Therapy for Parkinsonian Patients

Anomaly detection is essential in healthcare, particularly for monitoring cognitive therapy in Parkinson’s patients. This thesis develops an unsupervised anomaly detection system within TheraBrain, a cloud-based platform for cognitive training. The system analyzes multivariate time-series data to identify deviations in user performance—such as reaction time, distraction, and accuracy—using a sliding window approach to capture temporal dependencies and aid early detection of cognitive decline. Given the lack of labeled anomalies, synthetic anomaly injection techniques were employed. Various detection methods were tested, including Isolation Forest, Gaussian Mixture Models, Autoencoders, Variational Autoencoders (VAE), and Generative Adversarial Networks (GANs). The best-performing models, VAE-LSTM and VAE-GAN, effectively captured temporal patterns while maintaining generalizability. A scalable deployment pipeline was implemented using AWS cloud services, enabling real-time anomaly detection. Experimental results highlight the potential of deep learning in enhancing remote cognitive therapy. This research demonstrates AI’s role in improving accessibility and early intervention for Parkinson’s patients. Future work will focus on refining detection thresholds, integrating self-reported player feedback to improve anomaly contextualization, and allowing clinicians to fine-tune the model for greater accuracy and reliability.