Cardiorespiratory Fitness Estimation utilising Wearable ECG Data

Cardiorespiratory fitness (CRF), expressed as maximal oxygen uptake (VO₂max), is a key indicator of perioperative risk and long-term health outcomes. However, gold-standard cardiopulmonary exercise testing is resource-intensive and not always practical in preoperative pathways. Wearable sensors offer a scalable alternative for estimating VO₂max, but challenges remain around the quality of raw signals, transparency of pre-processing, and integration of advanced features such as heart rate variability (HRV). This thesis aimed to develop and evaluate a methodological framework for predicting VO₂max from wearable sensor data in a preoperative cohort. It (i) evaluated existing signal quality indices (SQIs) for ECG, (ii) assessed the contribution of HRV features, and (iii) developing a task-specific SQI tailored to accurate heart rate estimation.
Data were collected from the REMOTES clinical study which recorded 72-hour wearable-ECG and accelerometer signals from patients scheduled for major abdominal surgery. Multiple open-source SQIs were first compared using annotated synthetic ECG data, and the best-performing approach was applied to the REMOTES dataset. From these SQI-filtered signals, wearable-derived features were used to train a range of machine learning models; regression-based models performed best and were selected for further analysis. HRV features were then incorporated to evaluate their added predictive value, and a new HR-specific SQI was developed using open-access datasets to assess its impact on HR extraction and predictive accuracy.
Integrating HRV features improved VO₂max prediction (R² = 0.47 vs 0.42). Implementing the HR-specific SQI further enhanced model performance (R² = 0.51; correlation = 0.73), confirming that aligning signal-quality assessment with analytical goals improves HR-derived features and downstream model accuracy. Findings highlight how successive optimisation of feature extraction (via HRV) and signal processing (via a task-specific SQI) can enhance predictive performance. This work presents an analytical pipeline from raw wearable ECG data to VO₂max prediction, providing a foundation for scalable, data-driven cardiorespiratory fitness assessment.