AI for Systematic Reviews, Gait, and Activity Recognition

Abstract

Background: As the global population ages, the imperative to address challenges related to healthy aging, healthcare burdens, and continuous vital sign monitoring intensifies. Despite numerous digital healthcare solutions, gaps persist in their effective implementation for diagnostic and monitoring applications. This thesis presents innovative tools designed to bridge these gaps, enhancing both the diagnostic capabilities and efficiency of health monitoring technologies. Objectives: The primary goals of this thesis are threefold: (1) to establish the diagnostic utility of gait signatures for identifying different subtypes of dementia; (2) to enhance the efficiency of systematic literature reviews through automated tools; and (3) to develop and validate models for human activity recognition using smartphone sensor data, addressing challenges such as manual annotation and sensor heterogeneity. Methods: We conducted a systematic review and meta-analysis to investigate specific gait patterns among individuals with dementia and healthy controls, thereby uncovering subtype-specific gait signatures. We also designed software, leveraging ChatGPT and GPT-3.5 Turbo, to automate critical steps of systematic reviews, optimizing for efficiency and screening performance via different prompting strategies. Finally, we created a smartphone application for gait and activity monitoring, training machine learning models on multi-sensor data to classify daily activities. Results: Our meta-analysis confirmed that gait parameters reliably distinguish dementia subtypes, including Alzheimer’s disease and vascular dementia. The AI-based systematic review tools significantly reduced screening time while maintaining acceptable accuracy, demonstrating the potential of automated evidence synthesis. In the realm of smartphone-based HAR, our models performed robustly in controlled datasets, yet encountered generalization challenges on new data due to sensor heterogeneity and domain shift. These findings highlight both the promise and the practical complexities of scaling smartphone-based monitoring solutions. Conclusion: Integrating advanced computational approaches and AI into healthcare can enhance diagnostic precision, streamline systematic reviews, and expand smartphone-based health monitoring capabilities. While our models demonstrated strong initial performance, real-world applications require careful consideration of data variability and domain shift. Future research should focus on refining domain adaptation techniques, ensuring more diverse data coverage, and further validating these methods in clinical and community settings.