Real-Time Predictive Analytics for Healthcare Monitoring and Terrain Classification Using Data Streaming and Online Learning

This thesis develops a scalable, real-time predictive framework aimed at improving outcomes in healthcare monitoring and gait classification. Specifically, it addresses two primary applications: early detection of cardiovascular events, myocardial infarction (MI) and arrhythmia, through continuous analysis of ECG data streams; and adaptive terrain classification from gait data. Utilizing data streaming technologies such as Apache Kafka and Apache Spark, coupled with machine learning models, this framework enables near real-time and precise predictions from high-throughput, real-time data. Furthermore, the integration of online learning frameworks, including the River API and TensorFlow I/O, allows for continuous model adaptation, enhancing the system’s responsiveness and predictive accuracy in dynamic, evolving environments.In the healthcare domain, a heart attack and arrhythmia prediction pipeline is established using logistic regression, multilayer perception (MLP), and LSTM-based models to assess real-time ECG data for early indications of cardiovascular risk. Findings demonstrate that the online learning models incorporated into the system outperformed static models by continuously adapting to new data patterns, maintaining high levels of predictive relevance. In the context of terrain classification, machine learning models such as Random Forest, XGBoost, BiLSTMs, and MLPs are applied to gait data from foot-mounted inertial measurement units (IMUs) across various surfaces. Adaptive real-time classification is achieved through online learning, allowing the system to dynamically adjust as new terrain data is received. A comparative analysis between River API-based and TensorFlow-based models reveals trade-offs between high static accuracy and the flexibility of incremental learning, suggesting that each approach has unique advantages depending on the stability or variability of the data environment. Results from the real-time health monitoring and gait classification applications underscore the efficacy of combining data streaming platforms with adaptive learning frameworks. The study indicates that while deep learning models excel in stable environments requiring static accuracy, online learning frameworks provide superior adaptability, making them well-suited for dynamic, continuously evolving data contexts. Furthermore, the research addresses challenges such as class adaptation, dynamic data scaling, and missing value management in real-time streaming applications, contributing solutions to improve efficiency. This research advances the field of real-time predictive analytics, offering robust, adaptive system architecture that supports proactive healthcare monitoring and terrain classification for enhanced safety and functionality in assistive technologies and autonomous systems.