Design, implementation and evaluation of a wearable metabolic monitoring system for exercise physiology

Thesis Abstract

Cardiovascular and metabolic responses during high-intensity and endurance exercise vary markedly among individuals, yet practical real-time assessment tools that can be deployed in field settings remain limited, hindering precision training and clinical decision-making. This study addresses the gap by designing, implementing, and evaluating a wearable metabolic monitoring system capable of real-time measurement of energy expenditure, substrate utilization, and ventilatory efficiency during diverse exercise modalities. The aim is to develop a low-cost, user-friendly platform that integrates multi-sensor data (gas exchange, accelerometry, heart rate, and dermal biosignals) with an on-device processing unit and cloud-based analytics to provide individualized metabolic profiling. Specific objectives are to (1) design and prototype a wearable device capable of accurate breath-by-breath gas analysis and motion sensing; (2) validate the system against a criterion instrument (Douglas bag/indirect calorimetry) in a laboratory setting across cycling, running, and resistance training protocols; (3) implement real-time data fusion and machine learning algorithms to estimate energy expenditure, respiratory exchange ratio, and substrate oxidation ( carbohydrate vs fat); (4) evaluate the system’s reliability and robustness during field tests in recreational and trained populations; and (5) assess user acceptability and practical utility for exercise prescription and clinical monitoring. Methodologically, the study adopts a mixed-methods design with a predominantly quantitative focus. The population comprises adults aged 18–55 years, including 40 healthy volunteers and 20 endurance-trained athletes, recruited from university campuses and local clubs. A randomized testing sequence exposes participants to three exercise modalities (graded treadmill running, cycle ergometry, and resistance circuit) at varying intensities (50–95% peak VO2). The sample size provides adequate power (>80%) to detect mean differences within 95% confidence for energy expenditure estimates (? = 0.5 kcal·min?1) and respiratory exchange ratio estimates (? = 0.02). Instruments include a custom wearable metabolics module integrated with a portable gas analyzer, tri-axial accelerometer, photoplethysmography sensor, skin temperature sensor, and a microcontroller-based data logger. Indirect calorimetry serves as the gold standard for validation, while standardized metabolic carts capture energy expenditure, RER, VO2, and VCO2. Data collection involves simultaneous wearable and reference measurements during each protocol, with repeated trials to assess intra-individual variability. Data analysis follows a rigorous, multi-layered approach. First, descriptive statistics summarize sensor performance and measurement concordance relative to indirect calorimetry. Second, agreement analyses employ Bland-Altman plots and intraclass correlation coefficients to quantify accuracy and repeatability for energy expenditure and RER estimates. Third, regression analyses (ordinary least squares and mixed-effects models) examine the relationship between wearable-derived metrics and criterion measures, accounting for modality, intensity, and participant characteristics. Fourth, machine learning models, including random forest and gradient boosting, are trained to predict substrate oxidation rates (carbohydrate vs fat) from fused sensor features, with cross-validation to prevent overfitting. The theoretical framework leverages the best-practice metabolic physiology concepts and the Information Processing Theory to interpret real-time data integration as a perceptual-motor aid for exercise regulation. Ethical considerations include informed consent, data anonymity, and compliance with institutional review board protocols. Expected findings anticipate high correlation (ICC > 0.85) between wearable estimates and criterion measures for energy expenditure and RER across activities, with root-mean-square errors within acceptable bounds for practical use (?0.25 kcal·min?1 for energy expenditure; ?0.03 for RER). Real-time data fusion is expected to yield reliable substrate oxidation estimates in real-world settings, enhanced by machine learning models capturing inter-individual variability. Qualitative feedback from participants and coaches is anticipated to indicate strong usability and perceived utility for tailoring training intensity and monitoring recovery. The study contributes to knowledge by advancing a validated, scalable wearable platform that bridges laboratory-grade metabolic assessment and field applicability, enabling personalized exercise prescriptions and remote monitoring in sports science and clinical physiology. Practical implications include data-driven training optimization, injury risk reduction through precise exertion control, and the potential integration with telehealth platforms for chronic disease management. Limitations include potential sensor drift under extreme conditions and the need for ongoing calibration. Recommendations emphasize iterative hardware refinement, broader demographic validation, integration with nutrition tracking, and exploration of longitudinal applications in rehabilitation and performance enhancement.