NeuroCareIoT for reliable multimodal edge based Alzheimer’s home monitoring using NeuroFuseNet

Continuous, reliable assessment of falls and night-time wandering, routine deviations and physiological instability must be provided for people with Alzheimer’s disease at home. Current methods, however, typically address only a single modality of sensing, are grounded in fixed decision thresholds, and fail to appropriately consider probability calibration and predictive uncertainty, resulting in too many false alarms at night, delayed detection of subtle changes, and diminished trust by caregivers. To fill these gaps, we present NeuroCareIoT, an edge-first IoT framework that integrates motion, location, routines, and vitals while accounting for reliability and uncertainty. It consists of four task-specific backbones (IMU micro-dynamics: SafeFallNet; room-level localisation with path entropy: WanderSenseNet; temporal deviation modelling: DailyRoutineNet; motion-aware HR/HRV: VitalRhythmNet), coupled by NeuroFuseNet, a reliability-aware fusion layer that weighs modalities based on quality scores ( $$qi$$ ) and outputs a calibrated risk. On top of this, NeuroFuseNet applies Platt scaling at the internal node level for calibration and includes an entropy-based gate to suppress low-confidence windows and drive tiered escalation. NeuroCareIoT was shown to provide better performance in terms of precision (86.9%) and recall (90.3%) across five random seed data sets than both the heuristic baseline, the best-performing single-modality model and naïve average fusion. It also achieved better probability calibration (Brier score = 0.061), lower expected calibration error (2.6%), and more stable performance across decision thresholds. Ablation analysis revealed that reliability weighting and uncertainty gating had the greatest effect on precision, and SafeFallNet had the greatest effect on sensitivity, while maintaining a 95th-percentile decision latency under 200 ms on both the Raspberry Pi 4B and Jetson Nano platforms. Personalisation reduced the number of false alarms by 0.54 per night and shortened the time to detect routine deviations by 1.7 minutes. Quiet-hours alert burden was further lowered, and the precision of night-time alerts was further improved by tuning the policy for caregivers. In addition, case studies of SHAP-based care notes were given to facilitate transparent decision-making, along with risk and entropy trajectories. All results show that NeuroCareIoT is a viable prototype for implementing a multimodal monitoring framework that can be realised through an edge-first approach and tested in controlled, synchronised replay experiments built on publicly available datasets. The suggested framework serves as a basis for future validation through longitudinal real-home studies with Alzheimer’s patients, caregivers, and clinician-adjudicated incident records. Not applicable.