Explainable and Subject-Independent VO₂ Estimation Using a Single IMU: A Lightweight Ensemble Framework Under LOSO...

Area of Science:

• Biomechanics
• Sports Science
• Wearable Technology

Background:

• Continuous estimation of oxygen uptake (VO2) using wearable sensors is a practical alternative to laboratory testing.
• Challenges exist due to the indirect relationship between kinematics and physiological demand.

Purpose of the Study:

• To present a lightweight, two-stage pipeline for simultaneous heel-strike (HS) detection and VO2 estimation using a single calf-mounted IMU.
• To evaluate the pipeline's performance using machine learning models and leave-one-subject-out validation.

Main Methods:

• Stage 1: Heel-strike detection using an Extreme Learning Machine (ELM) + Random Forest (RF) ensemble.
• Stage 2: VO2 estimation using RF and ensemble regression with kinematic and gait-derived features from 30s windows.
• Leave-one-subject-out (LOSO) cross-validation across 24 participants.

Main Results:

• The ELM+RF ensemble achieved the highest HS detection F1-score (0.818), outperforming a TCN baseline (F1=0.674).
• RF model achieved a median R² of 0.687 using predicted HS events for VO2 estimation, with no significant difference compared to ground-truth.
• Accelerometer variability and gyroscope-derived features were dominant predictors for VO2 estimation.

Conclusions:

• VO2 estimation is feasible using automatically detected gait events without manual annotation.
• The proposed pipeline is computationally efficient and shows promise for controlled conditions, pending further validation.