Predicting Exercise Intensity from SmartwatchData: Device Capabilities and Limitations

Sammanfattning

Wearable sensors and machine learning are increasingly used to monitor exercise intensity for remote health applications, yet most existing research relies on expensive, proprietary clinical hardware that restricts access to raw data and algorithms. A significant gap remains in understanding whether low-cost, fully open-source consumer wearables can produce sufficiently accurate predictions of exercise intensity to serve as viable alternatives for accessible health monitoring. This thesis addresses two problems: first, the data engineering challenge of temporally aligning asynchronous sensor streams from devices with fundamentally different sampling rates and clock configurations, and second, the machine learning feasibility of predicting Metabolic Equivalents of Task (METs) from open-source hardware benchmarked against a clinical gold standard. Using a multi-sensor dataset of 24 participants from the DIWAH study at Linnaeus University, we developed an end-to-end pipeline comprising manual temporal alignment validated by cross-correlation, standardised FLIRT feature extraction, and Leave-One-Subject-Out cross-validation with Ridge Regression and Random Forest models across three device scenarios (Reference, EmotiBit, and Bangle.js). The cross-correlation analysis confirmed successful synchronisation across all device pairs (𝑟 ≥ 0.93). The Random Forest models achieved mean absolute errors between 1.49 and 1.73 METs across all devices, while Ridge Regression failed to produce viable predictions. Furthermore, a Wilcoxon signed-rank test confirmed there is no statistically significant difference (𝑝 > 0.39) in predictive accuracy between the open-source consumer devices and the clinical reference setup. These results suggest that expensive research equipment may not be strictly necessary for broad exercise intensity classification. Instead, open-source consumer wearables, when combined with robust data processing methodologies, can achieve comparable predictive accuracy to proprietary clinical hardware.