Oxynet: A collective intelligence that detects ventilatory thresholds in cardiopulmonary exercise tests

Area of Science:

• Exercise Physiology
• Artificial Intelligence in Medicine
• Biomedical Data Analysis

Background:

• The accurate determination of ventilatory thresholds (VT1 and VT2) from cardiopulmonary exercise testing (CPET) remains debated, with current expert visual inspection methods lacking consistent reliability.
• There is a clinical need for objective, automated methods to determine ventilatory thresholds, improving the consistency and accessibility of CPET data analysis.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) framework for the automatic detection of VT1 and VT2 using a large, crowd-sourced CPET dataset.
• To assess the performance of a convolutional neural network (CNN) algorithm against expert opinions for ventilatory threshold determination.
• To explore the potential of collective intelligence approaches in improving the accuracy of CPET data analysis.

Main Methods:

• A web-application was utilized to crowdsource 1245 CPET datasets from individuals with varying aerobic fitness levels.
• A convolutional neural network (CNN) algorithm was trained and tested on this database for automatic detection of VT1 and VT2.
• The CNN's performance was evaluated against expert visual inspection using 206 CPETs, analyzing accuracy via mean absolute error and correlation coefficients.

Main Results:

• The CNN achieved high accuracy in detecting VT2 (MAE 144 mlO2/min, 6.1%, r=0.99) and VT1 (MAE 178 mlO2/min, 11.1%, r=0.97).
• Performance for VT1 detection was less accurate in individuals with lower aerobic fitness.
• The AI system demonstrated potential to surpass the accuracy of isolated experts in identifying ventilatory thresholds.

Conclusions:

• A collaborative intelligence system, leveraging AI and crowd-sourced data, shows significant potential for accurate ventilatory threshold detection in CPET.
• Further development, including a larger expert-validated VT1 dataset, is necessary to fully translate these findings into routine clinical practice.
• Automated methods offer a promising avenue to address the current controversies surrounding ventilatory threshold determination in CPET.