Machine Learning Model To Automate Ventilatory Threshold Determination During Vo2max Assessments

First (VT1) and second (VT2) ventilatory thresholds represent the intensities of distinct changes in oxygen intake during exercise. Identifying VT1 and VT2 pose a challenge when analyzing graded maximal exercise test data, as tests are often adjudicated manually. Automating the determination of VT1 and VT2 using machine learning models may provide an objective, unbiased method to aid in the process of identifying these thresholds. PURPOSE: Develop a machine learning model to identify VT1 and VT2 during graded maximal cardiopulmonary exercise tests (CPET). METHODS: 106 recreational runners (age 34 ± 11 yrs; BMI: 23.4 ± 2.9 kg/m2) completed a graded maximal CPET. Test protocol was dictated by self-selected half marathon race pace with 1 km/hr increments every 3 min until volitional fatigue. Breath-by-breath data were used to identify VT1 and VT2 using the v-slope method to produce ground truth measures. Subsequently, data were used by the Long Short-Term Memory (LSTM) recurrent neural network model with inputs of absolute (VO2) and relative (VO2/kg) oxygen uptake, carbon dioxide production (VCO2), and minute ventilation (VE). The network output produced a 1x3 vector specifying the probability of the point to lie in the first region (input<VT1), second region (VT1 < input<VT2), or the third region (input>VT2). Categorical cross entropy was used as loss and accuracy as the metric to measure network performance. The point where the probability of input changes from first region to second region was marked as VT1, and from second to third region was marked as VT2. The model’s identified thresholds were compared to ground truth thresholds using mean absolute error (MAE). RESULTS: The LSTM model produced an accuracy of 99.7% and a categorical loss of 0.002 for the classification of time series data. Prediction accuracy of VT1 was 92% and for VT2 was 97%. The MAE between the predicted VT1 and ground truth VT1 was 3.22 ml/kg/min and between predicted VT2 and ground truth VT2 was 3.12 ml/kg/min. CONCLUSION: A machine learning approach provided an accurate method to identify VT1 and VT2 from CPET data. Further model iterations are needed to refine these methods and increase accuracy. This study received funding from Apple, Inc. The funding source had no role in the analysis andinterpretation of the data or in the submission of this abstract. This study received funding from Apple, Inc. The funding source had no role in the analysis and interpretation of the data or in the submission of this abstract.