An Adaptive Machine Learning Framework for Real-Time Health Monitoring using IoT-based Physiological Data

• Published in: 2025 3rd International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS) pp. 728 - 733
• Main Authors: Pasha, MD Sajit, Shaik, Mohammed Ali
• Format: Conference Proceeding
• Language: English
• Published: IEEE 11.06.2025
• Subjects: Accuracy; Adaptation models; Adaptive machine learning; Adaptive systems; Biological system modeling; Biomedical monitoring; Concept drift detection; Ensemble learning; IoT-based healthcare; Monitoring; Random forests; Real-time health monitoring; Real-time systems; Sensitivity; Wearable devices
• DOI: 10.1109/ICSSAS66150.2025.11080865

In view of the advancement brought by IoT-healthcare technology, positive feedback including the ability to check constantly wearing apparels such as wearable devices has provided real-time, early assessment on the state of health. However, the more classic approaches of machine learning are not suitable to work on the high-frequency real-time data flows as they are not ready to learn on the alterations of a patient's condition. This paper proposes WPLFS-ELR, a weighted late fusion streaming data analysis framework that has the capability to work with physiological data for intelligent health applications. The overall proposed model combines a number of non-complex online learners who are updated periodically by the Exponential Moving Average of the learner's prediction accuracy and current performance feedback. Late fusion of different physiological signals is assigned with independence of each signal model, and the Page-Hinkley and ADWIN methods help to keep adaptivity based on the subject's condition. Verifications conducted on the real biological dataset lead to the observation that the WPLFS-ELR is more accurate than base learners including Hoeffding Trees, Adaptive Random Forests and More classifiers with reference to both, accuracy and drift sensitivity. In particular, WPLFS-ELR to score 96% of the accuracy, the general evaluation's absolute maximum out of all examined models, with the adaptation of 0.55s, which is also acceptable in comparison to Random Forests (0.16s) or SVMs (0.18s). It also offers better adaptability when it comes to real-time inference ahead of the defined model.