Точность пульсометров умных часов резко падает при плавании: итальянские ученые создали новый открытый набор данных

An open-access dataset reveals how sharply smartwatch heart rate accuracy can decline underwater, providing researchers with a new foundation for improving wearable monitoring in swimming and aquatic health. Study: Wearable Heart Rate Monitoring during Swimming. Image Credit: Dean Drobot / Shutterstock In a recent study published in the journal Scientific Data, researchers in Italy described the development and metrological characterization of an open-access dataset designed to evaluate the performance of commercial wearable heart rate monitors in aquatic environments. The study systematically evaluated three commercial sensors (a Polar chest strap and Polar and Garmin smartwatches) in competitive swimmers under both dry (metrological control) and wet conditions. Study findings revealed that the wrist-worn sensors performed substantially better on land, while their metrological performance degraded underwater. The authors noted that water can attenuate photoplethysmography (PPG) optical signals and introduce hydrodynamic noise, providing potential mechanisms for this decline in performance. This dataset provides an open-access benchmark that enables researchers to investigate machine learning (ML) approaches, sensor fusion, signal-quality assessment, and improved wearable monitoring in aquatic environments. Background Decades of cardiovascular and sports science research have established continuous heart rate (HR) monitoring as an essential tool for assessing athletes' baselines and tracking changes in several key physiological measures, including physical exertion, nervous system activity, and psychophysical stress. The standard approach to HR measurement is electrocardiography (ECG). However, conventional ECG monitoring is often limited to short measurement periods, and prolonged monitoring requires specialized portable systems such as Holter monitors, thereby restricting its routine use during many sporting activities. The recent rapid proliferation of consumer wearables, particularly smartwatches and HR-monitoring chest straps, has significantly mitigated this limitation, enabling prolonged heart rate monitoring during daily routines and athletic training. However, aquatic environments create distinctive challenges for wrist-based PPG monitoring because water acts as a dense medium that attenuates photoplethysmography (PPG) optical signals and introduces distinct hydrodynamic noise that differs substantially from dry-land motion artifacts. These challenges highlight the need for well-characterized aquatic datasets that allow researchers to distinguish the effects of water from motion artifacts and investigate data-processing approaches specifically under water-based conditions. About the study The present study aimed to address this limitation and inform future wearable sensor design and software development by developing a synchronized multi-sensor dataset across dry and aquatic experimental conditions. The study sample cohort comprised 10 healthy female competitive swimmers (17 ± 3 years old; Fitzpatrick skin types I and II) who were required to simultaneously wear three commercial water-resistant devices: 1. Polar H10 chest strap (ECG-based reference device), 2. Polar Vantage V2 smartwatch (worn on the left wrist; PPG-based), and 3. Garmin Venu Sq smartwatch (worn on the right wrist; PPG-based). Data collection was divided into a “dry condition,” which served as the metrological control, and a “wet condition” representing the aquatic test environment. During the dry condition, researchers measured baseline (resting state), followed by three treadmill walking trials (4 km/h) and three running trials (6 km/h) to help distinguish motion-related effects from those introduced by water. In the wet condition, researchers measured resting state in water followed by four swimming activities (freestyle, butterfly stroke, backstroke, and breaststroke; 4 laps each in a 25-meter pool) performed at self-perceived normal and fast paces. The final dataset compiled 18,622 synchronized 1 Hz timestamps, representing 97.61% of the 19,078 expected samples, which were used to statistically evaluate the accuracy and precision of the wrist wearable devices against those of the reference chest strap using Bland-Altman agreement plots, Mean Absolute Percentage Error (MAPE), and Pearson’s correlation coefficients (r). Study findings Statistical analyses revealed that in dry conditions, both smartwatches demonstrated substantially better agreement with the reference chest strap. The Garmin Venu Sq recorded a mean residual of 1 beat per minute (bpm; MAPE = 4.05%, r = 0.95), while the Polar Vantage V2 exhibited a mean residual of -5 bpm (MAPE = 8.00%, r = 0.85). In stark contrast, in wet conditions, the Garmin Venu Sq exhibited substantial heart rate underestimation, with a mean residual of -44 bpm (MAPE = 29.95%, r = 0.26). While the Polar Vantage V2 demonstrated notably superior accuracy (mean residual = -14 bpm; MAPE = 17.17%; r = 0.59), its performance was still substantially poorer than under dry conditions. Furthermore, Bland-Altman plots revealed distinct "looping" error patterns in the analyses of PPG-based devices under dry and wet conditions, which the authors attributed to physiological and algorithmic lag in PPG optical sensors during rapid transitions in physical exertion. Conclusions The authors describe the dataset as a distinctive resource for characterizing how aquatic environments affect the metrological performance of wrist-worn consumer optical HR monitors and note that, to their knowledge, no comparable dataset with these characteristics is currently available. The study was limited by a demographically homogeneous sample of 10 young, light-skinned female athletes. The Polar H10 reference was not medical-grade, and the dataset includes manufacturer-generated heart-rate series rather than raw ECG or PPG waveforms, which limits signal-quality assessment and the development of custom algorithms. Even so, the open-access dataset is a useful resource for preliminary machine-learning studies, sensor-fusion algorithm development, and refining wearable technology for athletic performance and potentially aquatic rehabilitation.