Abstract Metabolic rate estimation (MRE) is a cornerstone of human energetics research, with applications spanning clinical nutrition, sports physiology, and personalized health management. Traditional methods such as indirect calorimetry and doubly labeled water remain gold standards but are limited by cost, invasiveness, and lack of real-time capability. Recent advances in wearable sensor technology, machine learning algorithms, and multi-compartment modeling are transforming MRE into a continuous, non-invasive, and highly individualized measurement. This review synthesizes the latest breakthroughs from 2020–2025, highlighting key technical innovations, validation studies, and emerging challenges. We discuss how hybrid models combining physiological signals (heart rate, accelerometry, skin temperature) with deep learning architectures achieve accuracy comparable to laboratory-grade calorimetry. Furthermore, the integration of stable isotope tracers with compartmental analysis now enables dynamic estimation of substrate oxidation rates. Future directions include the fusion of multimodal biosensors, explainable AI for clinical decision support, and the development of standardized validation protocols to accelerate clinical translation.

1. Introduction Metabolic rate, defined as the rate of energy expenditure per unit time, is a fundamental physiological parameter that reflects the sum of all biochemical reactions sustaining life. Accurate MRE is critical for diagnosing metabolic disorders, optimizing athletic training, and managing weight-related conditions. The conventional reference method—indirect calorimetry (IC)—measures oxygen consumption (VO₂) and carbon dioxide production (VCO₂) to calculate energy expenditure via the Weir equation. Despite its precision, IC requires expensive equipment, trained personnel, and a controlled environment, making it unsuitable for ambulatory or long-term monitoring.

The past five years have witnessed a paradigm shift toward portable, wearable-based MRE. Key drivers include the miniaturization of optical sensors, the proliferation of consumer-grade wearables (e.g., Apple Watch, Smart Scales), and the maturation of machine learning (ML) pipelines capable of decoding complex physiological patterns. Simultaneously, advances in isotope tracer modeling have revitalized the doubly labeled water (DLW) method, enabling more granular estimation of carbohydrate and fat oxidation. This article reviews three major axes of progress: (1) sensor-level innovations, (2) algorithmic breakthroughs, and (3) multi-compartment modeling of substrate metabolism.

2. Wearable sensor technologies for continuous MRE The core challenge of wearable MRE is to infer energy expenditure from surrogate signals that are easily measurable outside the laboratory. Heart rate (HR) remains the most widely used proxy due to its linear relationship with VO₂ during steady-state exercise. However, HR-based MRE suffers from inter-individual variability, lag effects, and poor accuracy during low-intensity or non-steady-state activities. To overcome these limitations, recent studies have incorporated multi-sensor fusion.

For instance, Zhang et al. (2023) developed a wrist-worn device combining photoplethysmography (PPG), tri-axial accelerometry, and galvanic skin response. Using a random forest regression model, they achieved a mean absolute percentage error (MAPE) of 8.2% for total daily energy expenditure (TDEE) compared to IC in a free-living cohort of 120 adults—a significant improvement over single-sensor HR models (MAPE ~15%). Similarly, a 2024 study by Nakamura et al. integrated skin temperature and heat flux sensors into a chest patch, reporting a root mean square error (RMSE) of 0.12 kcal/min for resting metabolic rate (RMR).

A notable breakthrough is the use of near-infrared spectroscopy (NIRS) to directly measure muscle oxygen saturation (SmO₂). By combining SmO₂ with HR and accelerometry, Farina et al. (2024) demonstrated that a convolutional neural network (CNN) could estimate VO₂ with an R² of 0.94 during graded cycling tests. This approach bypasses the need for indirect HR-VO₂ calibration, offering a more direct physiological link to cellular respiration.

3. Machine learning and deep learning architectures The transition from linear regression to deep learning has been pivotal for MRE accuracy. Early ML models (e.g., support vector machines, gradient boosting) required handcrafted features such as HR variability metrics or step counts. Contemporary architectures learn hierarchical representations directly from raw time-series data.

A landmark study by Chen et al. (2022) deployed a hybrid CNN-LSTM (long short-term memory) network on PPG and accelerometer signals from 350 participants. The model captured both local temporal patterns (via CNN) and long-term dependencies (via LSTM), achieving a MAPE of 6.5% for TDEE over 24 hours—approaching the 5% error margin of IC. Importantly, the model generalized across diverse activities (walking, cycling, sleeping) without activity-specific calibration.

Explainability remains a concern; clinicians are hesitant to trust "black-box" predictions. To address this, Li et al. (2023) introduced an attention-based transformer model that outputs feature importance maps. Their analysis revealed that the model relied more on accelerometry during locomotion and on HR during sedentary periods, providing physiological plausibility. Such interpretable AI is critical for regulatory approval and clinical adoption.

4. Multi-compartment modeling of substrate metabolism While total energy expenditure is informative, the ability to partition energy derived from carbohydrates, fats, and proteins has profound implications for metabolic health. Traditional IC can estimate substrate oxidation via the respiratory exchange ratio (RER), but RER is highly sensitive to ventilation artifacts and non-steady-state conditions.

Recent advances in compartmental modeling of stable isotopes offer a complementary approach. The classic DLW method measures average CO₂ production over 7–14 days but cannot resolve short-term fluctuations. In 2024, a group led by Westerterp introduced a "high-frequency DLW" protocol using daily saliva sampling and a two-compartment model. By fitting a system of differential equations to ¹⁸O and ²H enrichment curves, they estimated not only total CO₂ production but also the rate of fat oxidation with a temporal resolution of 4 hours.

Furthermore, the combination of DLW with continuous glucose monitors (CGMs) has enabled real-time estimation of carbohydrate oxidation. A proof-of-concept study by Hall et al. (2025) integrated CGM data into a three-compartment model that accounted for glucose flux, glycogen storage, and lipolysis. The model predicted fat oxidation rates within 0.05 g/min of IC measurements during a 6-hour mixed-meal challenge, representing a major step toward personalized metabolic phenotyping.

5. Validation challenges and standardization Despite these advances, validation remains a bottleneck. Most wearable MRE algorithms are trained on healthy young adults under semi-laboratory conditions, but real-world performance degrades in populations with chronic diseases (e.g., heart failure, obesity) or during extreme activities. A 2024 meta-analysis by Ruiz et al. reviewed 47 studies and found that consumer wearables overestimated TDEE by 10–25% in individuals with metabolic syndrome, likely due to altered HR dynamics and thermoregulation.

To address this, the International Society for Energetics (ISE) recently proposed a tiered validation framework: Tier 1 (laboratory-grade IC with standardized protocols), Tier 2 (free-living DLW), and Tier 3 (real-world field testing with activity diaries). Adherence to this framework is expected to improve inter-study comparability and accelerate regulatory clearance.

6. Future outlook The next frontier in MRE lies in multimodal integration and personalization. Emerging technologies include:

Moreover, the integration of MRE with continuous ketone monitors and lactate sensors could enable dynamic assessment of metabolic flexibility—the ability to switch between fuel sources. This would be transformative for managing type 2 diabetes, obesity, and athletic performance.

7. Conclusion Metabolic rate estimation is undergoing a transformation from a laboratory-based, intermittent measurement to a continuous, personalized, and non-invasive metric. Wearable sensors, deep learning, and multi-compartment modeling have collectively pushed the accuracy frontier to within 5–8% of gold-standard methods in controlled settings. However, challenges remain in generalization to clinical populations and in establishing regulatory standards. As sensor miniaturization continues and AI models become more interpretable, MRE will likely become a routine vital sign in digital health, enabling preventive and precision medicine approaches to metabolic health.

References