Wearable sensing is increasingly used in smart healthcare for continuous and non-intrusive monitoring, yet many studies emphasize predictive accuracy without sufficiently considering edge–cloud deployment constraints such as inference latency, model footprint, feature compactness, and communication overhead. This paper presents a resource-aware edge–cloud continuum framework for lightweight wearable smart-healthcare analytics based on windowed processing, handcrafted features, feature reduction, and CPU-efficient classifiers within a sensing–gateway–edge–cloud architecture. The contribution is positioned as an integrative deployment-evaluation framework rather than as a new learning algorithm, scheduler, offloading protocol, or resource-adaptive optimizer. The primary benchmark uses the original 12-class MHEALTH taxonomy under subject-independent evaluation because it provides a more discriminative assessment than the grouped three-state formulation, which is retained as a simplified monitoring setting. External robustness is evaluated on PPG-DaLiA under its own LOSO protocol, while direct \(MHEALTH\rightarrow PPG-DaLiA\) transfer is treated as a stringent stress test. The evaluation reports predictive performance, per-class results, confusion matrices, inference latency, model size, feature compactness, and payload-level communication cost, together with implementation details needed for reproducibility. Results show that Random Forest achieves the strongest class-balanced performance on the primary original 12-class MHEALTH benchmark, while Decision Tree provides the lightest model-level runtime profile. The grouped three-state task reaches near-ceiling performance and is therefore interpreted only as a complementary coarse monitoring scenario. In this grouped monitoring setting, the compactness analysis identifies a robust operating region around 60 retained features, while the executed primary benchmark uses the predefined reduced representation. Payload-level communication accounting shows that feature-based transmission substantially reduces application-level payload size relative to raw-window transmission. Under dataset-specific LOSO evaluation on PPG-DaLiA, the framework retains meaningful performance, whereas direct cross-dataset transfer degrades severely. Overall, the results support the framework as a compact and communication-aware wearable-analytics approach from an edge–cloud architectural perspective, while full real-device execution, adaptive scheduling, resource-aware offloading, and end-to-end network validation remain necessary future work.
