From Lab to Wrist: Bridging Metabolic Monitoring and Consumer Wearables for Heart Rate and Oxygen Consumption Modeling

TL;DR

This work addresses the barrier of laboratory-only VO$_2$ measurement by predicting instantaneous VO$_2$ trajectories from consumer wearables. It couples a physiologically informed HR dynamics model (via a neural ODE or neural Kalman filter) with a sequence-to-sequence VO$_2$ predictor that is calibrated from the initial VO$_2$ second, validated on a synchronized multimodal dataset against Cosmed K5 measurements. The approach yields a 1-second HR MAE of $2.81$ bpm (correlation $0.87$) and VO$_2$ MAPEs around $11$–$13\%$, demonstrating robust cross-subject performance and the potential to identify metabolic zones from noninvasive data. By embedding physiology-informed constraints into modern ML, this framework democratizes advanced metabolic monitoring for both elite athletes and recreational exercisers, bridging laboratory-grade accuracy with everyday accessibility.

Abstract

Understanding physiological responses during running is critical for performance optimization, tailored training prescriptions, and athlete health management. We introduce a comprehensive framework -- what we believe to be the first capable of predicting instantaneous oxygen consumption (VO$_{2}$) trajectories exclusively from consumer-grade wearable data. Our approach employs two complementary physiological models: (1) accurate modeling of heart rate (HR) dynamics via a physiologically constrained ordinary differential equation (ODE) and neural Kalman filter, trained on over 3 million HR observations, achieving 1-second interval predictions with mean absolute errors as low as 2.81\,bpm (correlation 0.87); and (2) leveraging the principles of precise HR modeling, a novel VO$_{2}$ prediction architecture requiring only the initial second of VO$_{2}$ data for calibration, enabling robust, sequence-to-sequence metabolic demand estimation. Despite relying solely on smartwatch and chest-strap data, our method achieves mean absolute percentage errors of approximately 13\%, effectively capturing rapid physiological transitions and steady-state conditions across diverse running intensities. Our synchronized dataset, complemented by blood lactate measurements, further lays the foundation for future noninvasive metabolic zone identification. By embedding physiological constraints within modern machine learning, this framework democratizes advanced metabolic monitoring, bridging laboratory-grade accuracy and everyday accessibility, thus empowering both elite athletes and recreational fitness enthusiasts.

Figures (8)

• Figure 1: Runners wearing Garmin 965 smartwatches, chest strap and Cosmed K5 portable metabolic systems during experimental sessions on an athletic track
• Figure 2: Schematic of HR prediction framework. Wearable data is encoded into latent states $\mathbf{s}_t$ and processed via either: (1) neural ODE solver with equation $\frac{dg_t}{dt} = d_t-g_t$, or (2) Kalman filter with learnable updates. Both use estimated moments $\hat{\mu}_t$ and $\hat{\sigma}_t$ for denormalization to produce HR predictions $\hat{h}_t$. Gray indicates learnable parameters $\bm{\mathrm{\vartheta}}_\ast$.
• Figure 3: Example HR prediction from the ODE based model (128, 2) for a high intensity workout. Blue line represents true HR measurements, and red line shows model predictions.
• Figure 4: The VO$_{2}$ prediction framework. The pipeline processes wearable sensor inputs through normalization and encoding steps before implementing a dual-stream architecture with a neural Kalman filter and direct estimation pathway, which are adaptively blended to produce the final VO$_{2}$ predictions.
• Figure 5: VO$_{2}$ predictions from first-second data during an incremental testing session with periodic lactate measurements. Blue: ground truth; orange: model predictions.