Advances in sensor miniaturization, pervasive connectivity, and scalable cloud architectures have accelerated the adoption of Internet-of-Things solutions in healthcare, enabling continuous physiological monitoring, early disease detection, and remote clinical interventions. Yet, the complexity of heterogeneous sensor data, variable patient contexts, and unpredictable network conditions still limit reliability and predictive accuracy in real-world deployments. This study develops a hybrid deep-learning pipeline that integrates convolutional neural networks, bidirectional recurrent architectures, and attention-based temporal encoders with cloud-native event stream processing to enable real-time interpretation of multimodal physiological signals. The research examines how edge-assisted inference, micro-batch stream analytics, and distributed message brokers collectively enhance detection latency, anomaly classification, and model robustness. A mixed-method methodology combines simulation-driven performance evaluation with empirical analysis of IoT device logs and consumable EHR-derived datasets. Results demonstrate significant improvements in prediction accuracy, event-processing throughput, alert precision, and resilience against noisy sensor streams. The findings highlight the potential of hybrid AI pipelines to strengthen remote patient monitoring, chronic disease management, and population-health surveillance while addressing operational barriers tied to privacy, scalability, and interoperability.