A Review of Edge Intelligence–Multimodal Data Fusion Methodologies for AI-IoT-Based Patient's Predictive Early Warning Systems

Clinical deterioration outpaces recognition in real workflows: threshold-based monitoring floods nurses with low-value alarms, while cloud-centric analytics impose latency and privacy costs that obstruct real-time response. This review synthesizes the ten most recently published advances (2026) in AI-driven IoT-based patient monitoring, categorizing multimodal fusion architectures, edge inference strategies, early warning and alarm-management designs, and deployment-validation practice. Multimodal deep architectures reach AUROC 0.7857 over 5.7 million prediction samples (Sadanandan, 2026) and ICU mortality AUC 0.96 (Abuhamad et al., 2026); transformer-based IoT-EHR fusion cuts sensor-forecasting error by 16.1% and raises clinical-risk AUC by 3.9% (Liu et al., 2026); yet every high-accuracy system remains cloud-resident and retrospectively validated. At the bedside, machine learning alarms (0.032 versus 0.132 alarms per patient-hour; PPV 0.181 versus 0.073) (Rosario et al., 2026) reduce but do not remove noise, and continuous monitoring devices still present usability barriers in noncritical care (Pan et al., 2026). A comparative gap analysis of the ten works shows no system simultaneously delivering edge-resident inference, quality-gated multimodal fusion, alarm-aware alerting, and prospective multi-site validation; the design space this review formalizes into a reference architecture.