Expert-level classification of ventilatory thresholds from cardiopulmonary exercising test data with recurrent neural networks

MDMNet: Multi-dimensional multi-modal network to identify organ system limitation in cardiopulmonary exercise testing

BACKGROUND AND OBJECTIVE

Cardiopulmonary exercise testing (CPET) serves as an integrative and comprehensive assessment tool for cardiorespiratory fitness. In this paper, we present a novel multi-dimensional multi-modal network (MDMNet) to identify functional limitation of organ systems via CPET, which is of great importance in clinical practice and yet a challenging task due to (1) the intricate intra-variable associations, and (2) the significant inter-individual variability.

METHODS

The proposed model has three compelling characteristics. First, we employ a dedicated embedding strategy for CPET data to map raw inputs into the learned embedding space, facilitating the detection of latent features of physiological variables. Second, we devise a novel multi-dimensional feature extraction module to capture rich features of physiological inputs at different dimensions, which consists of a one-dimensional feature extraction branch unfolding both temporal and spatial patterns of the entire data, and a two-dimensional feature extraction branch based on Gramian Angular Field (GAF) encoding to reveal the complicated temporal correlation relationships between time points within a variable. Third, we integrate these techniques with clinically significant demographic information to establish our MDMNet incorporating multi-dimensional with multi-modal learning, thereby further addressing the issues of complex intra-variable associations and inter-individual variability simultaneously.

RESULTS

We evaluated the proposed method on the publicly available CPET dataset, achieving AUC scores of 0.948, 0.949 and 0.931 for three tasks respectively.

CONCLUSIONS

The superiority of our method in discerning inter-individual differences was further demonstrated through partial least squares discriminant analysis, which holds significant potential for automated clinical application of CPET.

Dynamic Prediction of Physical Exertion: Leveraging AI Models and Wearable Sensor Data During Cycling Exercise

Background/Objectives: This study aimed to explore machine learning approaches for predicting physical exertion using physiological signals collected from wearable devices. Methods: Both traditional machine learning and deep learning methods for classification and regression were assessed. The research involved 27 healthy participants engaged in controlled cycling exercises. Physiological data, including ECG, heart rate, oxygen saturation, and pedal speed (RPM), were collected during these sessions, which were divided into eight two-minute segments. Heart rate variability (HRV) was also calculated to serve as a predictive indicator. We employed two feature selection algorithms to identify the most relevant features for model training: Minimum Redundancy Maximum Relevance (MRMR) for both classification and regression, and Univariate Feature Ranking for Classification. A total of 34 traditional models were developed using MATLAB's Classification Learner App, utilizing 20% of the data for testing. In addition, Long Short-Term Memory (LSTM) networks were trained on the top features selected by the MRMR and Univariate Feature Ranking algorithms to enhance model performance. Finally, the MRMR-selected features were used for regression to train the LSTM model for predicting continuous outcomes. Results: The LSTM model for regression demonstrated robust predictive capabilities, achieving a mean squared error (MSE) of 0.8493 and an R-squared value of 0.7757. The classification models also showed promising results, with the highest testing accuracy reaching 89.2% and an F1 score of 91.7%. Conclusions: These results underscore the effectiveness of combining feature selection algorithms with advanced machine learning (ML) and deep learning techniques for predicting physical exertion levels using wearable sensor data.

From data to decision: Machine learning determination of aerobic and anaerobic thresholds in athletes

Lactate analysis plays an important role in sports science and training decisions for optimising performance, endurance, and overall success in sports. Two parameters are widely used for these goals: aerobic (AeT) and anaerobic (AnT) thresholds. However, determining AeT proves more challenging than AnT threshold due to both physiological intricacies and practical considerations. Thus, the aim of this study was to determine AeT and AnT thresholds using machine learning modelling (ML) and to compare ML-obtained results with the parameters' values determined using conventional methods. ML seems to be highly useful due to its ability to handle complex, personalised data, identify nonlinear relationships, and provide accurate predictions. The 183 results of CardioPulmonary Exercise Test (CPET) accompanied by lactate and heart ratio analyses from amateur athletes were enrolled to the study and ML models using the following algorithms: Random Forest, XGBoost (Extreme Gradient Boosting), and LightGBM (Light Gradient Boosting Machine) and metrics: R2, mean absolute error (MAE), mean squared error (MSE) and root mean square error (RMSE). The regressors used belong to the group of ensemble learning algorithms that combine the predictions of multiple base models to improve overall performance and counteract overfitting to training data. Based on evaluation metrics, the following models give the best predictions: for AeT: Random Forest has an R2 value of 0.645, MAE of 4.630, MSE of 44.450, RMSE of 6.667; and for AnT: LightGBM has an R2 of 0.803, the highest among the models, MAE of 3.439, the lowest among the models, MSE of 20.953, and RMSE of 4.577. Outlined research experiments, a comprehensive review of existing literature in the field, and obtained results suggest that ML models can be trained to make personalised predictions based on an individual athlete's unique physiological response to exercise. Athletes exhibit significant variation in their AeT and AT, and ML can capture these individual differences, allowing for tailored training recommendations and performance optimization.

Estimation of ventilatory thresholds during exercise using respiratory wearable sensors

Ventilatory thresholds (VTs) are key physiological parameters used to evaluate physical performance and determine aerobic and anaerobic transitions during exercise. Current assessment of these parameters requires ergospirometry, limiting evaluation to laboratory or clinical settings. In this work, we introduce a wearable respiratory system that continuously tracks breathing during exercise and estimates VTs during ramp tests. We validate the respiratory rate and VTs predictions in 17 healthy adults using ergospirometry analysis. In addition, we use the wearable system to evaluate VTs in 107 recreational athletes during ramp tests outside the laboratory and show that the mean population values agree with physiological variables traditionally used to exercise prescription. We envision that respiratory wearables can be useful in determining aerobic and anaerobic parameters with promising applications in health telemonitoring and human performance.

Deep learning-based dynamic ventilatory threshold estimation from electrocardiograms

BACKGROUND AND OBJECTIVE

The ventilatory threshold (VT) marks the transition from aerobic to anaerobic metabolism and is used to assess cardiorespiratory endurance. A conventional way to assess VT is cardiopulmonary exercise testing, which requires a gas analyzer. Another method for measuring VT involves calculating the heart rate variability (HRV) from an electrocardiogram (ECG) by computing the variability of heartbeats. However, the HRV method has some limitations. ECGs should be recorded for at least 5 minutes to calculate the HRV, and the result may depend on the utilized ECG preprocessing algorithms.

METHODS

To overcome these problems, we developed a deep learning-based model consisting of long short-term memory (LSTM) and convolutional neural network (CNN) for a lead II ECG. Variables reflecting subjects' physical characteristics, as well as ECG signals, were input into the model to estimate VT. We applied joint optimization to the CNN layers to generate an informative latent space, which was fed to the LSTM layers. The model was trained and evaluated on two datasets, one from the Bruce protocol and the other from a protocol including multiple tasks (MT).

RESULTS

Acceptable performances (mean and 95% CI) were obtained on the datasets from the Bruce protocol (-0.28[-1.91,1.34] ml/min/kg) and the MT protocol (0.07[-3.14,3.28] ml/min/kg) regarding the differences between the predictions and labels. The coefficient of determination, Pearson correlation coefficient, and root mean square error were 0.84, 0.93, and 0.868 for the Bruce protocol and 0.73, 0.97, and 3.373 for the MT protocol, respectively.

CONCLUSIONS

The results indicated that it is possible for the proposed model to simultaneously assess VT with the inputs of successive ECGs. In addition, from ablation studies concerning the physical variables and the joint optimization process, it was demonstrated that their use could boost the VT assessment performance of the model. The proposed model enables dynamic VT estimation with ECGs, which could help with managing cardiorespiratory fitness in daily life and cardiovascular rehabilitation in patients.

Machine Learning Models for the Automatic Detection of Exercise Thresholds in Cardiopulmonary Exercising Tests: From Regression to Generation to Explanation

The cardiopulmonary exercise test (CPET) constitutes a gold standard for the assessment of an individual's cardiovascular fitness. A trend is emerging for the development of new machine-learning techniques applied to the automatic process of CPET data. Some of these focus on the precise task of detecting the exercise thresholds, which represent important physiological parameters. Three are the major challenges tackled by this contribution: (A) regression (i.e., the process of correctly identifying the exercise intensity domains and their crossing points); (B) generation (i.e., the process of artificially creating a CPET data file ex-novo); and (C) explanation (i.e., proving an interpretable explanation about the output of the machine learning model). The following methods were used for each challenge: (A) a convolutional neural network adapted for multi-variable time series; (B) a conditional generative adversarial neural network; and (C) visual explanations and calculations of model decisions have been conducted using cooperative game theory (Shapley's values). The results for the regression, generation, and explanatory techniques for AI-assisted CPET interpretation are presented here in a unique framework for the first time: (A) machine learning techniques reported an expert-level accuracy in the classification of exercise intensity domains; (B) experts are not able to substantially differentiate between a real vs an artificially generated CPET; and (C) Shapley's values can provide an explanation about the choices of the algorithms in terms of ventilatory variables. With the aim to increase their technology-readiness level, all the models discussed in this contribution have been incorporated into a free-to-use Python package called pyoxynet (ver. 12.1). This contribution should therefore be of interest to major players operating in the CPET device market and engineering.

Using Machine Learning to Identify Organ System Specific Limitations to Exercise via Cardiopulmonary Exercise Testing

Cardiopulmonary Exer cise Testing (CPET) is a unique physiologic medical test used to evaluate human response to progressive maximal exercise stress. Depending on the degree and type of deviation from the normal physiologic response, CPET can help identify a patient's specific limitations to exercise to guide clinical care without the need for other expensive and invasive diagnostic tests. However, given the amount and complexity of data obtained from CPET, interpretation and visualization of test results is challenging. CPET data currently require dedicated training and significant experience for proper clinician interpretation. To make CPET more accessible to clinicians, we investigated a simplified data interpretation and visualization tool using machine learning algorithms. The visualization shows three types of limitations (cardiac, pulmonary and others); values are defined based on the results of three independent random forest classifiers. To display the models' scores and make them interpretable to the clinicians, an interactive dashboard with the scores and interpretability plots was developed. This machine learning platform has the potential to augment existing diagnostic procedures and provide a tool to make CPET more accessible to clinicians.

A Reappraisal of Ventilatory Thresholds in Wheelchair Athletes With a Spinal Cord Injury: Do They Really Exist?

The ventilatory threshold (VT) separates low- from moderate-intensity exercise, the respiratory compensation point (RCP) moderate- from high-intensity exercise. Both concepts assume breakpoints in respiratory data. However, the objective determination of the VT and RCP using breakpoint models during upper-body modality exercise in wheelchair athletes with spinal cord injury (SCI) has received little attention. Therefore, the aim of this study was to compare the fit of breakpoint models (i.e., two linear regression lines) with continuous no-breakpoint models (i.e., exponential curve/second-order polynomial) to respiratory data obtained during a graded wheelchair exercise test to exhaustion. These fits were compared employing adjusted R², and blocked bootstrapping was used to derive estimates of a median and 95% confidence intervals (CI). V̇O₂-V̇CO₂ and V̇E/V̇O₂-time data were assessed for the determination of the VT, and V̇CO₂-V̇E and V̇E/V̇CO₂-time data for the determination of the RCP. Data of 9 wheelchair athletes with tetraplegia and 8 with paraplegia were evaluated. On an overall group-level, there was an overlap in the adjusted R² median ± 95% CI between the breakpoint and the no-breakpoint models for determining the VT (V̇O₂-V̇CO₂: 0.991 ± 0.003 vs. 0.990 ± 0.003; V̇E/V̇O₂-time: 0.792 ± 0.101 vs. 0.782 ± 0.104, respectively) and RCP (V̇E-V̇CO₂: 0.984 ± 0.004 vs. 0.984 ± 0.004; V̇E/V̇CO₂-time: 0.729 ± 0.064 vs. 0.691 ± 0.063, respectively), indicating similar model fit. We offer two lines of reasoning: (1) breakpoints in these respiratory data exist but are too subtle to result in a significant difference in adjusted R² between the investigated breakpoint and no-breakpoint models; (2) breakpoints do not exist, as has been argued previously.

Making Cardiopulmonary Exercise Testing Interpretable for Clinicians

ABSTRACT

Cardiopulmonary exercise testing (CPET) is a dynamic clinical tool for determining the cause for a person's exercise limitation. CPET provides clinicians with fundamental knowledge of the coupling of external to internal respiration (oxygen and carbon dioxide) during exercise. Subtle perturbations in CPET parameters can differentiate exercise responses among individual patients and disease states. However, perhaps because of the challenges in interpretation given the amount and complexity of data obtained, CPET is underused. In this article, we review fundamental concepts in CPET data interpretation and visualization. We also discuss future directions for how to best use CPET results to guide clinical care. Finally, we share a novel three-dimensional graphical platform for CPET data that simplifies conceptualization of organ system-specific (cardiac, pulmonary, and skeletal muscle) exercise limitations. Our goal is to make CPET testing more accessible to the general medical provider and make the test of greater use in the medical toolbox.

The importance of ventilatory thresholds to define aerobic exercise intensity in cardiac patients and healthy subjects

BACKGROUND

Although structured exercise training is strongly recommended in cardiac patients, uncertainties exist about the methods for determining exercise intensity (EI) and their correspondence with effective EI obtained by ventilatory thresholds. We aimed to determine the first (VT1 ) and second ventilatory thresholds (VT2 ) in cardiac patients, sedentary subjects, and athletes comparing VT1 and VT2 with EI defined by recommendations.

METHODS

We prospectively enrolled 350 subjects (mean age: 50.7±12.9 years; 167 cardiac patients, 150 healthy sedentary subjects, and 33 competitive endurance athletes). Each subject underwent ECG, echocardiography, and cardiopulmonary exercise testing. The percentages of peak VO2 , peak heart rate (HR), and HR reserve were obtained at VT1 and VT2 and compared with the EI definition proposed by the recommendations.

RESULTS

VO2 at VT1 corresponded to high rather than moderate EI in 67.1% and 79.6% of cardiac patients, applying the definition of moderate exercise by the previous recommendations and the 2020 guidelines, respectively. Most cardiac patients had VO2  values at VT2 corresponding to very-high rather than high EI (59.9% and 50.3%, by previous recommendations and 2020 guidelines, respectively). A better correspondence between ventilatory thresholds and recommended EI domains was observed in healthy subjects and athletes (90% and 93.9%, respectively).

CONCLUSIONS

EI definition based on percentages of peak HR and peak VO2  may misclassify the effective EI, and the discrepancy between the individually determined and the recommended EI is particularly relevant in cardiac patients. A ventilatory threshold-based rather than a range-based approach is advisable to define an appropriate level of EI.

Sensitivity of movement features to fatigue during an exhaustive treadmill run

HIGHLIGHTS

Trunk frontal plane kinematics is the most sensitive parameter to fatigue. Practitioners should consider this finding during endurance training.Kinetics exhibit a stable linear increase in mean values but a non-linear increase in variance during an exhaustive incremental treadmill run. This may affect training at a submaximal fatigued state.Specific areas in the joint distributions of kinetics and kinematics during treadmill running exhibit increased sensitivity in predicting fatigue state.

Detection of the Anaerobic Threshold in Endurance Sports: Validation of a New Method Using Correlation Properties of Heart Rate Variability

Past attempts to define an anaerobic threshold (AnT) have relied upon gas exchange kinetics, lactate testing and field-based evaluations. DFA a1, an index of heart rate (HR) variability (HRV) fractal correlation properties, has been shown to decrease with exercise intensity. The intent of this study is to investigate whether the AnT derived from gas exchange is associated with the transition from a correlated to uncorrelated random HRV pattern signified by a DFA a1 value of 0.5. HRV and gas exchange data were obtained from 15 participants during an incremental treadmill run. Comparison of the HR reached at the second ventilatory threshold (VT2) was made to the HR reached at a DFA a1 value of 0.5 (HRVT2). Based on Bland-Altman analysis and linear regression, there was strong agreement between VT2 and HRVT2 measured by HR (r = 0.78, p < 0.001). Mean VT2 was reached at a HR of 174 (±12) bpm compared to mean HRVT2 at a HR of 171 (±16) bpm. In summary, the HR associated with a DFA a1 value of 0.5 on an incremental treadmill ramp was closely related to that of the HR at the VT2 derived from gas exchange analysis. A distinct numerical value of DFA a1 representing an uncorrelated, random interbeat pattern appears to be associated with the VT2 and shows potential as a noninvasive marker for training intensity distribution and performance status.

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

The problem of the automatic determination of the first and second ventilatory thresholds (VT1 and VT2) from cardiopulmonary exercise test (CPET) still leads to controversy. The reliability of the gold standard methodology (i.e. expert visual inspection) feeds into the debate and several authors call for more objective automatic methods to be used in the clinical practice. In this study, we present a framework based on a collaborative approach, where a web-application was used to crowd-source a large number (1245) of CPET data of individuals with different aerobic fitness. The resulting database was used to train and test an artificial intelligence (i.e. a convolutional neural network) algorithm. This automatic classifier is currently implemented in another web-application and was used to detect the ventilatory thresholds in the available CPET. A total of 206 CPET were used to evaluate the accuracy of the estimations against the expert opinions. The neural network was able to detect the ventilatory thresholds with an average mean absolute error of 178 (198) mlO₂/min (11.1%, r = 0.97) and 144 (149) mlO₂/min (6.1%, r = 0.99), for VT1 and VT2 respectively. The performance of the neural network in detecting VT1 deteriorated in case of individuals with poor aerobic fitness. Our results suggest the potential for a collective intelligence system to outperform isolated experts in ventilatory thresholds detection. However, the inclusion of a larger number of VT1 examples certified by a community of experts will be likely needed before the abilities of this collective intelligence can be translated into the clinical use of CPET.

Estimating an individual's oxygen uptake during cycling exercise with a recurrent neural network trained from easy-to-obtain inputs: A pilot study

Measurement of oxygen uptake during exercise ([Formula: see text]) is currently non-accessible to most individuals without expensive and invasive equipment. The goal of this pilot study was to estimate cycling [Formula: see text] from easy-to-obtain inputs, such as heart rate, mechanical power output, cadence and respiratory frequency. To this end, a recurrent neural network was trained from laboratory cycling data to predict [Formula: see text] values. Data were collected on 7 amateur cyclists during a graded exercise test, two arbitrary protocols (Prot-1 and -2) and an "all-out" Wingate test. In Trial-1, a neural network was trained with data from a graded exercise test, Prot-1 and Wingate, before being tested against Prot-2. In Trial-2, a neural network was trained using data from the graded exercise test, Prot-1 and 2, before being tested against the Wingate test. Two analytical models (Models 1 and 2) were used to compare the predictive performance of the neural network. Predictive performance of the neural network was high during both Trial-1 (MAE = 229(35) mlO2min-1, r = 0.94) and Trial-2 (MAE = 304(150) mlO2min-1, r = 0.89). As expected, the predictive ability of Models 1 and 2 deteriorated from Trial-1 to Trial-2. Results suggest that recurrent neural networks have the potential to predict the individual [Formula: see text] response from easy-to-obtain inputs across a wide range of cycling intensities.