CardioSentinel-AI: An Explainable, Quality-Aware Edge Framework for Personalised Multimodal Cardiorespiratory Abnormality Detection

Low-cost monitors usually compare isolated vital signs with fixed population limits, which raises false alarms in people whose normal physiology sits near a limit, misses change in people whose baseline sits far from it, and fails silently when a sensor loses contact. This paper presents CardioSentinel-AI, an edge framework built around a Raspberry Pi gateway that combines electrocardiogram, photoplethysmogram, oxygen saturation, respiration and temperature channels. Each modality has its own gradient-boosted expert whose log-odds are weighted by a signal-quality index and summed, so every alert decomposes exactly into per-modality contributions. Features are also expressed as deviations from a personal baseline learned during a short enrolment, shrunk toward population statistics. The framework was evaluated on a physiologically grounded synthetic cohort of 80 virtual subjects (6,400 windows of 30 s) with atrial-fibrillation-like rhythm, desaturation, tachycardia and ectopy episodes, motion artefact and sensor dropout, under subject-independent cross-validation. Personalisation dominated: at 90% sensitivity it cut false alarms from 17.5 per hour for an equivalent population model to between 1.6 and 3.2 per hour, against 28.9 per hour for a fixed-threshold rule of the NEWS2 type. The quality-aware additive fusion did not outperform a personalised early-fusion model under nominal conditions (AUROC 0.989 against 0.993), but it degraded more slowly when each of the ECG, PPG and oximeter channels was dropped with probability 0.5 (0.917 against 0.890) and was better calibrated. Continuous quality weighting added little beyond masking absent channels. All results are simulated and must next be confirmed on MIT-BIH, BIDMC and prospective recordings.