Data Processing Methods And Their Effects On The Limits Of Agreement And Reliability Of Automated Submaximal Threshold Calculations

Abstract

Cardiovascular exercise intensity constitutes an important component of exercise prescription, but its individualization poses challenges. Previous research underscores the greater efficacy of physiological threshold-based cardiovascular exercise intensity prescription, when compared to standardized percentages of maximum heart rate, VO2, or workload. However, identifying these thresholds usually entails pre-processing steps like outlier elimination, interpolation, and data averaging. Although diverse algorithms exist to pinpoint these thresholds, the influence of prior data processing steps on algorithm-derived thresholds remains unclear. Through a scoping review, we gathered articles from studies that collected breath-by-breath gas exchange data during exercise in humans. We assessed the reporting prevalence and the nature of outlier removal, interpolation, and data-averaging methods. Approximately 5% of articles described outlier removal and interpolation details in their methods, while 2/3 reported data averaging. We developed an open-source R package, ”gasExchangeR,” to assess the effects of data processing choices on algorithm-derived thresholds. We included multiple threshold-detection algorithms from previous research in this package and validated them against simulated and human exercise tests. Most algorithms performed well under low simulated noise conditions but had higher relative and absolute error than visual detection. Leveraging the gasExchangeR package, algorithm-derived thresholds were computed across varied outlier removal limits, averaging durations, and algorithms using 350 exercise tests. A similar analysis was performed with 17 participants to assess the effect of these parameters on the test-retest reliability of algorithm-derived thresholds. The outcomes exhibited generally negligible main effects and interactions between outlier removal limit, averaging duration, and algorithm selection on average threshold values. Nevertheless, some statistically significant differences were observed. The 95% limits of agreement (LOA) among diverse data processing and algorithm combinations exceeded the expected measurement error in VO2. Linear regression highlighted algorithmic comparisons as the primary contributor to LOA variance. Specific algorithm types yielded statistically significant ICC values more frequently. These findings indicate that manipulating the data appearance despite constant underlying fitness can unveil inherent variability in distinct algorithms. In aggregate, these investigations underscore the potential for enhanced reproducibility through improved method documentation and the use of open-source software.