An IoT-Based Healthcare Monitoring and Surveillance Framework for People with Special Needs

People with special needs, such as frail older adults and people living with Parkinson’s disease (PD), heart-rhythm disorders or dementia, are often alone when a critical event happens. We present IoT-CARES, a four-layer Internet of Things (IoT) framework with one lightweight edge pipeline for five services: fall detection, activity monitoring of adults aged 70+, freezing-of-gait (FoG) detection, single-lead electrocardiogram (ECG) arrhythmia monitoring and camera-free cognitive-health surveillance. A dual-view ensemble (a fixed average of gradient-boosted trees on physiological descriptors and a linear model on random-convolution features) runs without a deep-learning runtime. A subject-level conformal calibration bounds a new person’s false-alarm rate by a carer-chosen budget with probability ≥90%, provided that new and calibration people are exchangeable, and federated learning trains shared models without moving raw data. On five open datasets (438 people), evaluated on held-out people, the framework reached an area under the receiver-operating-characteristic curve (AUROC) of 0.9998 for simulated falls and 0.908 for FoG, and a macro-F1 of 0.778 for activity; federated training matched centralised training. Personalisation helped only when the onboarding data contained the target events: five labelled ECG minutes raised arrhythmia macro-F1 from 0.667 (best global model) to 0.932, whereas fine-tuning on routine activity lowered activity macro-F1 (0.766 to 0.683) unless only the output layer was adapted (0.795). Conformal calibration kept 16 of 17 young volunteers (94.1%; 95% confidence interval 71.3–99.9%) free of false fall alarms during their activity protocol while detecting 94.6% of simulated falls (window-optimal threshold: 76.5% free of false alarms), and kept 9 of 10 people with PD within four false cues per hour; with so few people, these estimates are imprecise. In free-living data of 18 adults aged 70–95 years, a fall threshold calibrated on the volunteers raised no false alarms in 12.6 h, but calibration on the older adults’ own data detected more of the volunteers’ simulated falls (99.1% vs. 92.1%). Ambient sensors detected dementia (AUROC 0.962) but not reliably mild cognitive impairment (0.523). Inference took ≤ 24.3 ms per window on one processor core, and data leaving the home fell 219–759-fold.