Background A cardiopulmonary exercise test (CPET) provides the estimated lactate threshold (θLT) and respiratory compensation point (RCP) through visual identification of multivariate gas exchange and ventilatory profiles. Artificial intelligence tools, such as deep neural networks, can learn, replicate, and classify these patterns and potentially aid in θLT and RCP identification, removing the subjectivity of threshold detection. This study evaluated a set of deep learning models (Oxynet) pre-trained with more than 1200 CPET files and tested its performance against visual inspection of experts. Methods Evaluation included three phases: In phase I, 50 simulated ventilatory and gas exchange CPET files were generated, mixed with 50 authentic files, presented sequentially and in randomized order to three independent evaluators, and judged to be real or fake. In phase II, a new set of 50 files were generated, θLT and RCP were identified by both Oxynet and the consensus of three experts, and these estimates were compared with known values. In phase III, a subset of 163 CPETs were used to fine-tune Oxynet, and its evaluation of 50 independent authentic ramp CPET files were compared with those of the three experts. Results Experts correctly discriminated simulated from authentic data in 44% of cases (phase I). One-way ANOVA revealed no main effect of identified <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\dot{V}{\text{O}}_{{2}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mover accent="true"> <mi>V</mi> <mo>˙</mo> </mover> <msub> <mtext>O</mtext> <mn>2</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> (known vs Oxynet vs human evaluators) for both θLT (p = 0.41) and RCP (p = 0.39) with ~ zero effect size for both θLT (ω2 = 0.00) and RCP (ω2 = 0.00) (phase II). Using real ramp-incremental data (phase III), the fine-tuned Oxynet identified the <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\dot{V}{\text{O}}_{{2}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mover accent="true"> <mi>V</mi> <mo>˙</mo> </mover> <msub> <mtext>O</mtext> <mn>2</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> at 1944 ± 401 and 2555 ± 602 mL min−1 for θLT and RCP, respectively. Expert evaluators identified these at 1900 ± 469 and 2581 ± 625 mL min−1 with mean between-method biases of 45 mL min−1 (p = 0.087) and − 26 mL min−1 (p = 0.118). Conclusions Oxynet can be used as an accurate, reliable, and objective tool to identify or aid in the identification of exercise thresholds from gas exchange and ventilatory CPET data in healthy individuals.

AI-Driven Analysis of Cardiopulmonary Exercise Tests to Identify Gas Exchange and Ventilatory Thresholds

期刊论文
发表日期：

2026年3月5日

查看全文

Daniel A. Keir,
Andrea Zignoli,
Danilo Iannetta,
Felipe Mattioni Maturana,
Juan M. Murias

摘要

Background

A cardiopulmonary exercise test (CPET) provides the estimated lactate threshold (θ_LT) and respiratory compensation point (RCP) through visual identification of multivariate gas exchange and ventilatory profiles. Artificial intelligence tools, such as deep neural networks, can learn, replicate, and classify these patterns and potentially aid in θ_LT and RCP identification, removing the subjectivity of threshold detection. This study evaluated a set of deep learning models (Oxynet) pre-trained with more than 1200 CPET files and tested its performance against visual inspection of experts.

Methods

Evaluation included three phases: In phase I, 50 simulated ventilatory and gas exchange CPET files were generated, mixed with 50 authentic files, presented sequentially and in randomized order to three independent evaluators, and judged to be real or fake. In phase II, a new set of 50 files were generated, θ_LT and RCP were identified by both Oxynet and the consensus of three experts, and these estimates were compared with known values. In phase III, a subset of 163 CPETs were used to fine-tune Oxynet, and its evaluation of 50 independent authentic ramp CPET files were compared with those of the three experts.

Results

Experts correctly discriminated simulated from authentic data in 44% of cases (phase I). One-way ANOVA revealed no main effect of identified $\dot{V}{\text{O}}_{{2}}$ $\dot{V} O_{2}$ (known vs Oxynet vs human evaluators) for both θ_LT (p = 0.41) and RCP (p = 0.39) with ~ zero effect size for both θ_LT (ω² = 0.00) and RCP (ω² = 0.00) (phase II). Using real ramp-incremental data (phase III), the fine-tuned Oxynet identified the $\dot{V}{\text{O}}_{{2}}$ $\dot{V} O_{2}$ at 1944 ± 401 and 2555 ± 602 mL min⁻¹ for θ_LT and RCP, respectively. Expert evaluators identified these at 1900 ± 469 and 2581 ± 625 mL min⁻¹ with mean between-method biases of 45 mL min⁻¹ (p = 0.087) and − 26 mL min⁻¹ (p = 0.118).

Conclusions

Oxynet can be used as an accurate, reliable, and objective tool to identify or aid in the identification of exercise thresholds from gas exchange and ventilatory CPET data in healthy individuals.