A computational predictor of the anaerobic mechanical power outputs from a clinical exercise stress test

**Background:** Exercise training programs aim to improve cardiorespiratory fitness. While anaerobic metabolic capacity is important for athletic performance and health, it typically requires specialized tests like the Wingate anaerobic test (WAnT). Standard graded exercise tests (GXT) with ECG and blood pressure monitoring are more common in clinical settings but lack gas exchange measurements. The authors previously predicted anaerobic power outputs from cardiopulmonary exercise testing (CPET) with gas exchange data. This study investigates whether features from a standard GXT (without gas exchange) can similarly predict anaerobic mechanical power outputs.

**Methods:** Data from 93 healthy subjects (53 males, 40 females; mean age 29±6 and 26±4 years, respectively) who previously completed both a maximal incremental CPET on a treadmill and a WAnT were re-analyzed. CPET data were adjusted to simulate standard GXT conditions by calculating age-predicted maximal heart rate (HRmax = 208 – 0.7 × age) and deriving features at 85%, 90%, and 100% of HRmax. Twenty-five features were extracted, including predicted VO2max (using ACSM metabolic equations), exercise time, treadmill speed, and slope at each HR level. A greedy heuristic multiple linear regression algorithm was used, randomly dividing data into calibration (40%), test (30%), and validation (30%) sets. The algorithm selected optimal feature combinations to predict peak power (PP) and mean power (MP) from the WAnT. The model was applied to 100 randomly selected validation sets to determine mean ± SD percent error.

**Key Results:** For submaximal GXT at 85% age-predicted HRmax, a combination of 3 variables predicted PP with r = 0.93 (p < 0.001, RMSE = 38.9) and 4 variables predicted MP with r = 0.92 (p < 0.001, RMSE = 79). Percent errors were 15±3 for PP and 16±3 for MP. For maximal GXT at 100% age-predicted HRmax, 4 variables predicted PP with r = 0.92 (p < 0.001, RMSE = 90) and 2 variables predicted MP with r = 0.94 (p < 0.001, RMSE = 66). Percent errors were 12±2 for PP and 14±3 for MP. At 90% HRmax, predictions were slightly lower: r = 0.90 for PP and r = 0.93 for MP. Compared to the authors' prior CPET-based model (r = 0.95 for PP, r = 0.91 for MP), the GXT-based predictions were only 2-3% lower for PP and actually higher for MP. The percent error was also lower using GXT features (12-15% vs. 19% for PP with CPET).

**Clinical Implications:** This study provides evidence that anaerobic mechanical power outputs can be accurately predicted from a standard GXT without gas exchange measurements, which is more widely available and commonly used in clinical settings than CPET. This could allow clinicians, exercise physiologists, and coaches to assess both aerobic and anaerobic capacities from a single test, simplifying fitness evaluation and training prescription. The information is particularly valuable for athletes and for monitoring training adaptations. However, the study population was limited to healthy individuals aged 20-40 years, and the data were derived from CPET rather than an actual GXT. Validation in broader populations (including different athletic phenotypes, older adults, and clinical populations) and in prospective studies using real GXT protocols is needed before clinical implementation. The authors also note that larger sample sizes could support more complex models (e.g., neural networks) that might improve prediction accuracy.