From continuous wearable heart rate to fitness forecasting: an interpretable proof-of-concept model to predict change in maximal oxygen uptake
P Vermunicht, C Buyck, S Naessens, W Hens, E Van Craenenbroeck, J S Piedrahita Giraldo, K Laukens, J Roeykens, K De Deckere, L Desteghe, H HeidbuchelAbstract
Background/Introduction
Cardiorespiratory fitness predicts cardiovascular and all-cause outcomes. Improving fitness through physical activity is a core aim of cardiac rehabilitation, yet tracking fitness change outside supervised settings is difficult. Wrist photoplethysmography (PPG) provides continuous heart rate in daily life. This proof-of-concept study developed and evaluated the smart Antwerp Activity Index (smart-AAI), an interpretable machine learning model that estimates relative VO2max change using wearable heart rate and clinical features.
Purpose
To assess whether a simple model combining a heart rate-based activity score and baseline characteristics can predict relative percentage change in VO2max between cardiopulmonary exercise tests (CPETs).
Methods
Participants (n=149; 46 following cardiac rehabilitation, 103 healthy) wore a wrist-based PPG heart rate monitor continuously for 12 weeks and underwent CPETs. A subset of 78 participants was classified as PPG-compatible (mean absolute percentage error <10% during ≥70% of training data) and included in the training (n=60) and independent testing (n=18) datasets. Inputs were the Antwerp Activity Index (AAI), baseline VO2max, age, sex, body mass index, smoking status and days between CPETs. The AAI is an individualised heart rate based physical activity score that reflects both volume and intensity of activity. A multivariable linear regression predicted relative VO2max change. Performance was evaluated with correlation, mean absolute error (MAE), root mean squared error (RMSE) and Bland Altman agreement. Feature contributions were examined with SHapley Additive ExPlanations (SHAP).
Results
Across all PPG-compatible participants, mean VO2max rose from 33.8±12.5 to 35.6±11.3 mL/min/kg, corresponding to a relative increase of 7.2±16.7%. In the testing set (18 participants, 26 CPET intervals), predicted and observed relative VO2max change correlated moderately (r=0.45, p=0.02, Figure 1A). MAE was 9.6±7.4 percentage points and RMSE was 12.1±7.4 percentage points. Bland Altman analysis showed minimal bias of +1.3% and no systematic directional error (Figure 1B). As illustrated by the feature contribution SHAP plot (Figure 2), baseline VO2max exerted the largest influence on predictions, while the AAI, and to a lesser extent body mass index, were the main modifiable contributors, with the AAI being the most dynamic day to day driver for improvement. All feature impacts aligned with expected physiological directions.
Conclusion(s)
This small proof-of-concept study demonstrated that a lightweight model using wrist-derived heart rate activity and routine clinical data can estimate relative VO2max change with moderate accuracy in daily life. Smart-AAI may support remote cardiac rehabilitation follow-up and personalised feedback on expected fitness gains. Optimisation and validation in larger and more diverse cohorts is required before clinical deployment.