Predicting peak cardiorespiratory fitness in patients with cardiovascular disease using machine learning

doi:10.21203/rs.3.rs-5244290/v1

Download PDF

Article

Predicting peak cardiorespiratory fitness in patients with cardiovascular disease using machine learning

https://doi.org/10.21203/rs.3.rs-5244290/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Objective

This study aimed to develop machine learning (ML) models to predict peak cardiorespiratory fitness (CRF) before and after cardiac rehabilitation (CR).

Methods and Results

Data from 162 patients with cardiovascular disease were analyzed. Two predictive tasks were employed: Task 1 estimated peak oxygen consumption (VO₂ peak) using baseline clinical and functional data and Task 2 predicted changes in VO₂ peak after CR by additionally considering inter-visit exercise quantities and pre-CR cardiopulmonary exercise test (CPET) results. Four linear regression models and six ML models were trained and validated through 5-fold cross-validation technique. Both tasks demonstrated that the CatBoost and XGBoost models exhibited the highest predictive performance, effectively forecasting VO₂ peak values before and after CR. Task 1 highlighted the importance of the six-minute walk distance (6MWD), Korean Activity Scale Index (KASI), and hand grip strength (HGS) in predicting the initial VO₂ peak. Task 2 suggested a ceiling effect in the recovery of VO₂ peak following CR and emphasized the importance of resistance exercise.

Conclusion

The application of ML models provides a powerful tool for predicting the peak CRF in patients with CVD undergoing CR, both at the initial assessment and after completing rehabilitation programs.

Health sciences/Cardiology

Health sciences/Biomarkers/Predictive markers

physical fitness

cardiac rehabilitation

machine learning

predictive modeling

cardiovascular disease

Cardiovascular diseases (CVD) are the leading cause of death and illness globally. In the United States, CVD accounts for over 650,000 cases annually, making it the most common diagnosis among patients over the age of 65. The economic burden of CVD is significant, costing over $200 billion each year¹. Additionally, the prevalence of CVD among Koreans was reported to be 12.2% based on data from 2005 and 2007; this figure shows a substantial increase of 27.9% in the data from 2010 to 2017 ². Cardiac rehabilitation (CR) has been proven to be a part of the standard treatment for CVD, reducing cardiovascular mortality by approximately 26% and hospital readmissions by approximately 18%, while also enhancing patients' functional status and quality of life³. Furthermore, cardiorespiratory fitness (CRF) has been identified as a stronger potential predictor of mortality and CVD risk than factors such as obesity, diabetes, or dyslipidemia^4–6. Consequently, the 2016 Scientific Statement from the American Heart Association recommends regular assessment of CRF, as represented by peak oxygen consumption (VO₂ peak)⁷.

However, many patients participating in CR programs are unable to perform cardiopulmonary exercise testing (CPET), leading to the development of various non-exercise or maximal exercise-based predictive equations for estimating VO₂ peak indirectly⁸. Unfortunately, most previous VO₂ peak predictive equations were developed based on healthy cohorts, limiting their ability to reflect the characteristics of the cohorts in patients with CVD. Recently, machine learning (ML) models that consider a broader range of parameters for predicting VO₂ peak have been reported; however, these are also based on data from young, healthy adults^8,9. In 2023, Peterman et al. reported that predictive equations derived from a healthy cohort overestimated the VO₂ peak in patients with CVD, and developed a predictive equation that included variables such as the presence of underlying heart failure (HF) and myocardial infarction (MI) and whether a procedure or surgery was performed¹⁰. However, this equation utilized only simple clinical information such as age, sex, height, and weight without addressing individual patient-specific data such as physical function.

Little is known about the factors influencing the improvement in CRF in patients participating in CR programs. In 2017, Bargehr et al. reported predictors for the recovery of exercise capacity after CR, including baseline exercise capacity, age, percutaneous coronary intervention history, sex, systolic blood pressure (BP) at rest, body mass index (BMI), lipid-lowering drugs, low-density lipoprotein cholesterol, and triglycerides¹¹. However, this model predicted six-minute walk distance (6MWD) instead of VO₂ peak and did not include data that could influence exercise capacity, such as actual physical function and exercise performance records. Moreover, while predictive equations utilizing the 6MWD for estimating VO₂ peak have been introduced^12–15, Chirico et al. in 2020 reported that these equations are not suitable for monitoring changes in exercise capacity after CR in patients with HF¹⁶.

To accurately predict VO₂ peak and understand the importance of each factor, it is necessary to analyze a variety of factors. However, to date, no predictive model of VO₂ peak that considers all factors such as underlying diseases, functional ability, and exercise amount exists. Many predictive equations and models have failed to predict changes after CR programs, resulting in no models being applied in clinical practice. Therefore, this study aimed to 1) accurately predict VO₂ peak in patients with CVD using various clinical data without CPET through AI analysis, 2) predict VO₂ peak after CR implementation using previous data and exercise amount, and 3) explore the main factors influencing the improvement in CRF after CR.

Study population

The dataset used in this study was obtained from the Department of Physical Medicine and Rehabilitation at OO University OO Hospital. The medical records of 333 visits by 162 patients with CVD who underwent CPET for CR between March 2020 and May 2022 were retrospectively analyzed. The analyzed medical records comprised an initial postoperative evaluation and up to five follow-up assessments during one year, with patients advised to visit every 2–3 months. The exclusion criteria were: (i) significant orthopedic conditions or pain that limited participation in the CPET, (ii) unstable cardiopulmonary conditions, and (iii) severe cognitive impairment. This study was approved by the Institutional Review Board of OO University Hospital (IRB no. 2022AN0365) and conducted according to the principles of the Declaration of Helsinki.

Study design

Two tasks were designed to predict VO₂ peak using the frameworks illustrated in Fig. 1. Task 1 estimated VO₂ peak at the same visit point using only clinical information and functional assessments, excluding CPET data. Task 2 predicted VO₂ peak for the next visit point (post-CR VO₂ peak) and change in VO₂ peak between two visits (ΔVO₂ peak, Positive: recovery, Negative: deterioration) based on the amount of exercise performed between visits and pre-CR data, including CPET results. For Task 1, the data utilized included demographic information and physical measurements collected at the initial visit, along with medical history. Additionally, questionnaire data and functional assessments involving strength and endurance tests conducted at each visit were used in Task 1. In Task 2, in addition to the top eight variables shown to have high importance in Task 1, the CPET results and CR information collected through exercise logs were analyzed.

The detailed items utilized for each task analysis are listed in Table 1. The clinical information was categorized into demographic and disease-related information. Variables such as sex, underlying CVD, history of cardiac procedures and surgeries, and use of cardiovascular medication, were encoded as binary data. Ejection fraction was assessed via echocardiography during hospitalization, and categorized as follows: normal (≥ 50%), mildly reduced (35–49%), and reduced (< 35%).

Table 1

Summary of factors used to predict VO₂ peak
Demographic information	Age Sex Height Weight Body Mass Index (BMI)
Disease-related information	Underlying cardiovascular disease Cardiac surgery history Cardiac procedure history Cardiovascular medication Ejection fraction (EF) on echocardiography
Self-reported measures	Korean Activity Scale Index (KASI) EuroQol-5 dimension (EQ-5D) Drinking history Smoking history Previous exercise history
Performance-based measures	Six-minute walk distance (6MWD) Hand grip strength (HGS)
CPET information*	Peak oxygen consumption (VO₂ peak) Peak ventilatory threshold (VT peak) Peak heart rate (HR peak) Ventilatory equivalent for carbon dioxide (VE/VCO₂) Peak O₂ Pulse Peak systolic BP (SBP peak) Peak diastolic BP (DBP peak) Peak Respiratory Exchange Ratio (RER peak) Peak Rate Pressure Product (RPP peak) Peak Rate Perceived Exertion (RPE peak) Total exercise duration
CR exercise information*	Exercise type (resistance exercise, aerobic exercise (walking, cycling, others)) Duration (minutes per day) Frequency (days per week) Total exercise time (minutes per week)
* indicates that it was used only in the Task 2 study.
CPET, cardiopulmonary exercise test; CR, Cardiac rehabilitation.

At each visit, patients underwent functional assessments consisting of self-reported measures, performance-based measures, and CPET. Performance-based measures included hand grip strength (HGS), which was measured using a JAMAR PLUS hand dynamometer. Measurements were taken alternately from the left and right hands twice each, with the highest values recorded¹⁷. The 6MWD was assessed by instructing the patients to walk as far as possible within six minutes, maintaining an intensity level between 3 (moderate) and 4 (somewhat strong) on the Borg CR 10 scale¹⁸.

The self-reported Korean Activity Scale Index (KASI) was used to evaluate the feasibility of 15 daily activities by assigning a weighted score to each item¹⁹. The EuroQol-5 Dimension (EQ-5D) was used to assess the quality of life and general health status across five dimensions: mobility, self-care, usual activities, pain/discomfort, and anxiety/depression²⁰. From the second visit onwards, the patients completed questionnaires regarding their CR exercises, the details of which are specified in Table 1.

Patients performed symptom-limited progressive treadmill exercises as part of the standardized CPET. The exercise testing protocol was terminated at the patient's request or upon signs of gait instability or cardiovascular decompensation, following the guidelines of the American College of Sports Medicine²¹. The parameters derived from each assessment are listed in Table 1.

Following the American Association of Cardiovascular and Pulmonary Rehabilitation guidelines, patients were classified into low-, moderate-, and high-risk groups and prescribed target metabolic equivalents (METs) and heart rates (HR)²². During CR, patients maintained exercise diaries that included the duration of aerobic and resistance exercises, resting and peak HR, Rating of Perceived Exertion (RPE), and Respiratory Disturbance Index (RDI).

ML modeling

To ensure accurate validation and comparison of the dataset, regression and ML models were employed. Specifically, for Linear Regression models, basic linear regression is used along with various regularizations, such as LASSO²³, RIDGE²⁴, and SGD²⁵. Additionally, six ML models were employed for comparison with the linear regression models. Support Vector Regression (SVR)²⁶ was used because of its robustness against outliers and generalizations. Ensemble methods have been utilized, including GradientBoost²⁷, RandomForest²⁸, CatBoost²⁹, XGBoost³⁰, and LightGBM³¹. The optimal model was selected and parameter tuning was conducted accordingly.

For rigorous validation of each model's performance, five rounds of ML analyses were conducted, with each analysis randomly splitting the data into training and validation datasets. Given the small size of the dataset, the K-Fold Validation technique was employed to ensure that all data were used for both training and validation. This approach allows the performance of the model to be validated without data loss, offering a more reliable assessment than using a single validation set.

Among the metrics used to evaluate the regression model, the Sum of Squared Error (SSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE) were selected to prevent outliers and overestimation. These indicators enable an objective comparison of performance, especially when the values are either positive or negative.

Feature importance refers to the techniques that determine the extent to which each feature in a dataset contributes to the predictive power of the model. The SHapley Additive explanation (SHAP) algorithm³² is a technique in Explainable AI that calculates the SHAP value, which is the Shapley value for the conditional expectation function of an ML model, to analyze how much each feature contributes to individual prediction outcomes. It assumes that if the predictive performance changes significantly with the removal of a specific variable, then that variable is highly important. This approach consistently provides coherent interpretations by identifying variable importance and their positive or negative relationship with the target metric, taking into account correlations between variables.

Participant characteristics

The demographics, disease-related characteristics, baseline functional measurements, and CPET results of the patients are listed in Tables 2 and 3. The mean patient age was 60.0 ± 13.4 years. Of the patients, 75.9% were male and 24.1% were female. The average BMI of patients was 24.8 ± 4.0 kg/m². Angina pectoris was the most common underlying CVD, followed by congestive HF (CHF). The results of CR, collected through the self-reported exercise questionnaire, are shown in Fig. 2.

Table 2

Demographic and clinical characteristics of the patients (N = 162)
Variables	Values
Demographic information Age (years)	60.0 ± 13.4
Sex, male/female (number)	123 (75.9%) / 39 (24.1%)
Height (cm)	165.1 ± 8.3
Weight (kg)	67.9 ± 14.1
BMI (kg/m²)	24.8 ± 4.0
Disease-related information
Cardiovascular diagnosis (number)
Angina	69 (42.6%)
Congestive heart failure	40 (24.7%)
Valvular heart disease	30 (18.5%)
STEMI	21 (13.0%)
Arrhythmia	20 (12.4%)
NSTEMI	12 (7.4%)
Comorbidities (number)
Hypertension	105 (64.8%)
Diabetes mellitus	53 (32.7%)
Dyslipidemia	66 (40.7%)
Chronic liver disease	1 (0.6%)
Chronic lung disease	3 (1.9%)
Malignancy	9 (5.6%)
Ejection fraction (number) Normal (≥ 50%) Mildly reduced (35–49%) Reduced (< 35%)	97 (61%) 36 (22.6%) 29 (16.4%)
Values represent mean ± standard deviation or number (%)
BMI, Body mass index; STEMI, ST-elevation myocardial infarction; NSTEMI, Non-ST-elevation myocardial infarction.

Table 3

Initial assessment results of patients (N = 162)
Variables	Values
Self-reported measures KASI EQ- 5D	38.8 ± 18.2 0.85 ± 0.11
Smoking history (number) Never smoker Ex-smoker Current smoker	64 (39.8%) 84 (52.2%) 13 (8.1%)
Drinking history (number) Never drinker Ex-drinker Current drinker	77 (47.8%) 50 (31.1%) 34 (21.1%)
Performance-based measures 6MWD (m) HGS (kg)	457.3 ± 89.1 32.6 ± 9.9
CPET results VO2 peak (mL kg^− 1 min^− 1) VT peak (L) HR peak (bpm) VE/VCO2 O2 pulse peak (mL min^− 1) SBP peak (mmHg) DBP peak (mmHg) RPE peak RPP peak (bpm mmHg) CPET duration (sec) RER peak	18.9 ± 5.9 14.6 ± 5.3 131.7 ± 25.1 39.9 ± 18.6 10.6 ± 7.3 171.7 ± 29.5 75.2 ± 13.7 14.8 ± 1.6 21244.8 ± 6503.5 626.3 ± 217.0 1.0 ± 0.2
Risk classification (number) Low risk Moderate risk High risk	81 (50%) 58 (35.8%) 23 (14.2%)
Values represent mean ± standard deviation or number (%)
KASI, Korean Activity Status Index; EQ5D, EuroQol-5 dimension; 6MWD, Six-minute walk distance; HGS, Hand grip strength; CPET, Cardiopulmonary exercise test; VO2 peak, Peak oxygen consumption; VT peak, Peak Ventilatory threshold; HR peak, Peak heart rate; VE/VCO2, Ventilatory equivalent for carbon dioxide; SBP peak, peak systolic blood pressure; DBP peak, Peak diastolic blood pressure; RPE peak, Peak rate perceived exertion; RPP peak, Peak rate pressure product; RER peak, Peak respiratory exchange ratio.

Task 1: prediction of VO₂ peak using clinical data without CPET

Model performance

To predict the VO₂ peak at the same visit point using only clinical information and functional assessments, excluding CPET data, a total of ten models were utilized, and training was conducted using the 5-fold cross-validation technique. Table 4 presents the results of the study. The best-performing model for predicting the VO₂ peak was CatBoost, which showed an SSE of 811.29 and an RMSE of 3.70. Other models in the boosting family based on CatBoost, such as GradientBoost, XGBoost, and LightGBM, have also demonstrated high performance. However, the XGBoost model exhibited a slightly better performance based on the MAE metric.

Table 4

Performance of ten models for task 1
		SSE	MAE	RMSE
Linear Regression	Linear	1065.93 ± 262.73	3.10 ± 0.23	4.23 ± 0.51
	Lasso	2174.76 ± 169.31	4.97 ± 0.30	6.08 ± 0.25
	Ridge	921.48 ± 96.30	3.05 ± 0.14	3.95 ± 0.20
	SGD	1007.40 ± 150.47	3.20 ± 0.16	4.13 ± 0.30
Machine Learning	SVR	1358.07 ± 54.48	3.79 ± 0.08	4.81 ± 0.10
	Random forest	854.94 ± 182.16	2.90 ± 0.23	3.79 ± 0.42
	Gradient Boost	877.53 ± 115.96	2.97 ± 0.21	3.86 ± 0.25
	CatBoost	811.29 ± 143.65	2.81 ± 0.23	3.70 ± 0.33
	XGBoost	818.05 ± 176.23	2.77 ± 0.28	3.71 ± 0.43
	LightGBM	931.07 ± 103.40	3.03 ± 0.17	3.97 ± 0.22
Values represent mean ± standard deviation.
Bolded text indicates highest performance.
SGD, Stochastic Gradient Descent; SVR, Support Vector Regression; CatBoost, Category Boosting; XGBoost, eXtreme Gradient Boosting; LightGBM, Light Gradient Boosting Machine; SSE, Sum of Squared Errors; MSE, Mean Squared Error; RMSE, Root Mean Squared Erro.

To compare the prediction performance, a pre-existing VO₂ peak predictive equation derived from a CVD cohort¹⁰ was applied to our dataset, resulting in an MAE of 4.442 and an RMSE of 5.55. Excluding LASSO regression, the remaining nine models demonstrated superior performance.

Feature importance analysis (SHAP)

The SHAP algorithm was used to identify the factors influencing the VO₂ peak in Task 1. The graph on the left in Fig. 3, representing the mean absolute values of SHAP, indicates the importance of each variable. The results revealed that 6MWD, KASI, EQ-5D, and HGS were significant factors, along with age. Notably, the importance of 6MWD was more than double that of the other variables. The graph on the right in Fig. 3 shows the distribution of the SHAP values for each data point, indicating the direction of the impact of each variable on the VO₂ peak. For the 6MWD, there is a cluster of red points (indicating higher values) to the right of the SHAP value of zero, suggesting a positive relationship, where a higher 6MWD corresponds to a higher VO₂ peak. Regarding age, the red points were more widely dispersed than the blue points, suggesting that age plays a more critical role in older patients than in younger patients, affecting the VO₂ peak variably across different age groups.

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information

Model performance

After the CR program, the VO₂ peak was predicted using information from the previous visit point and CR data. Table 5 lists the predictive performance of each model. Similar to Task 1, ML performed better than the linear regression models. Overall, the best-performing model was XGBoost, which achieved an SSE of 457.18, an MAE of 1.86, and an RMSE of 2.55.Table 5 additionally details the results of the ML analysis set to predict ΔVO₂ peak. CatBoost exhibits the best performance across all indicators (SSE: 477.75, MAE: 1.92, RMSE: 2.62). Across all models, Task 2 showed significantly improved prediction accuracy compared with Task 1, suggesting that having CR and CPET data enables more accurate predictions.

Table 5

Performance of ten models for task 2
Post-CR VO₂ peak prediction
		SSE	MAE	RMSE
Linear Regression	Linear	718.78 ± 82.02	2.48 ± 0.16	3.23 ± 0.18
	Lasso	1961.69 ± 176.32	4.34 ± 0.22	5.34 ± 0.25
	Ridge	655.13 ± 80.16	2.41 ± 0.17	3.08 ± 0.19
	SGD	759.98 ± 115.46	2.54 ± 0.18	3.31 ± 0.27
Machine Learning	SVR	742.86 ± 122.08	2.53 ± 0.20	3.28 ± 0.27
	Random forest	500.02 ± 143.61	1.96 ± 0.19	2.67 ± 0.36
	Gradient Boost	469.65 ± 143.81	1.85 ± 0.17	2.59 ± 0.38
	CatBoost	459.05 ± 112.05	1.86 ± 0.11	2.57 ± 0.30
	XGBoost	457.18 ± 135.69	1.86 ± 0.17	2.55 ± 0.36
	LightGBM	589.65 ± 78.45	2.30 ± 0.12	2.92 ± 0.19
ΔVO₂ peak prediction
		SSE	MAE	RMSE
Linear Regression	Linear	690.49 ± 87.65	2.46 ± 0.15	3.16 ± 0.19
	Lasso	1229.27 ± 97.82	3.18 ± 0.16	4.22 ± 0.16
	Ridge	669.79 ± 69.14	2.42 ± 0.13	3.12 ± 0.16
	SGD	789.55 ± 111.49	2.62 ± 0.23	2.91 ± 0.36
Machine Learning	SVR	890.88 ± 88.98	2.72 ± 0.09	3.59 ± 0.17
	Random forest	592.86 ± 153.62	2.16 ± 0.23	2.91 ± 0.36
	Gradient Boost	559.98 ± 189.06	2.04 ± 0.28	2.81 ± 0.47
	CatBoost	477.75 ± 120.99	1.92 ± 0.14	2.62 ± 0.31
	XGBoost	549.00 ± 128.03	2.06 ± 0.16	2.81 ± 0.32
	LightGBM	680.64 ± 103.46	2.419 ± 0.17	3.14 ± 0.24
Values represent mean ± standard deviation.
Bolded text indicates highest performance.
SGD, Stochastic Gradient Descent; SVR, Support Vector Regression; CatBoost, Category Boosting; XGBoost, eXtreme Gradient Boosting; LightGBM, Light Gradient Boosting Machine; SSE, Sum of Squared Errors; MSE, Mean Squared Error; RMSE, Root Mean Squared Error.

Feature importance analysis (SHAP)

In the same way as Task 1, the SHAP algorithm was used to analyze feature importance. Figure 4(a) illustrates the importance of features for predicting the post-CR VO₂ peak, with the sequence of importance being pre-CR VO₂ peak, age, pre-CR VE/VCO₂, and 6MWD. Overall, the CPET results demonstrated a significant influence. Figure 4(b) represents the importance of variables for predicting ΔVO₂ peak. Consistently, the pre-CR VO₂ peak was ranked at the top, but the mean SHAP value was more than twice that of the other variables, indicating a relatively higher importance. The CPET results, including pre-CR VO₂ peak, along with HGS and KASI, exhibited a positive correlation with post-CR VO₂ peak but a negative correlation with ΔVO₂ peak. Notably, the 6MWD consistently showed a positive correlation in both analyses, though its mean SHAP value in ΔVO₂ peak prediction was relatively low at 0.17.

The importance of exercise-related factors for predicting ΔVO₂ peak is detailed in Fig. 5. Regarding CR, the total exercise time was the most critical factor, followed by the duration of resistance exercise. Except for the frequency of cycling, most exercise-related factors were positively correlated with ΔVO₂ peak.

Task 1: prediction of VO₂ peak using clinical data without CPET

In our study, the newly developed ML model successfully predicted the VO₂ peak in the CVD cohort based solely on simple functional evaluations and clinical characteristics without CPET data.

Numerous models and equations have been developed to estimate VO₂ peak indirectly. However, most studies have been based on data from healthy cohorts, and to our knowledge, the only regression equation developed based on a CVD cohort is by Peterman et al.¹⁰ Our study has significant clinical implications, as it presents the first model that uses ML to analyze a diverse range of characteristics integratively in a CVD cohort.

Moreover, our model demonstrated excellent predictive performance for the VO₂ peak. Compared with the previously reported prediction equation by Peterman et al. (RMSE: 5.55), our overall ML models showed superior performance. The best-performing model, CatBoost, achieved an RMSE of 3.70. This performance was better than that of the two ML models developed based on the 20-m shuttle run test in healthy young adults (RMSEs: 4.78 & 4.07), and was not significantly inferior to the performance of a model built by Abut et al. in 2019, which was based on maximal treadmill test data from healthy young adults (RMSE: 2.91)^33,34.

The comprehensive consideration of various parameters highlighted that among the functional evaluations, 6MWD, KASI, and HGS showed high importance in predicting the VO₂ peak. Among the clinical parameters, age, resting diastolic BP (DBP), and BMI were significant predictors. Notably, the SHAP value for the 6MWD, which demonstrated the highest importance, showed a positive correlation with a wide dispersion in negative direction. This suggests that a lower 6MWD is associated with a lower VO₂ peak and vice versa, indicating a concordant relationship. The 6MWD is a relatively inexpensive and rapidly administered tool for assessing functional aerobic capacity. Unlike CPET, it can be performed safely in patients with moderate to severe risk. Considerable literature has been published on the relationship between 6MWD and VO₂ peak. In 2010, Ross et al. utilized data from 11 study groups to develop a prediction equation for the VO₂ peak based on 6MWD, but they reported that the SEE was approximately 27% of the mean VO₂ peak, which limits its clinical application for individual patients¹³. Our findings highlight that while the 6MWD is a useful metric for building VO₂ peak prediction models, it should be considered alongside additional indicators.

In our study, the KASI ranked second in importance among the functional evaluation data used, displaying a distribution similar to that of the 6MWD, with a positive correlation. The KASI, a self-reported activity status questionnaire, is a safe and cost-effective assessment tool. Previous research involving patients with acute MI (AMI) reported a significant linear correlation between KASI and VO₂ peak, a trend that was confirmed in our model³⁵. The high variable importance of the KASI suggests that this correlation extends beyond AMI to encompass broader CVD categories, indicating its utility as an effective tool for evaluating exercise capacity across diverse CVD cohorts.

HGS has a strong correlation with maximum upper and lower body strength and overall muscle strength³⁶. Additionally, low HGS has been reported as an independent and strong risk factor for all-cause mortality and CVD, aligning with trends observed in VO₂ peak³⁷. In 2018, Chang et al. demonstrated a strong association between HGS and VO₂ peak in paraplegic men³⁸, whereas in 2021, Zhou et al. reported a significant relationship between HGS and VO₂ peak in healthy young adults, particularly noting higher correlations in males³⁹. Based on our study, HGS was an important predictor of VO₂ peak in the CVD cohort.

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information

By incorporating the high-importance indicators identified in Task 1 along with CPET and CR data, we have successfully predicted post-CR VO₂ peak and ΔVO₂ peak. Similar to Task 1, CatBoost model displayed the highest performance in predicting ΔVO₂ peak. Moreover, all models showed improved performance over Task 1, highlighting the significance of the CPET and exercise-related data.

Although many studies have shown that CR significantly influences functional aerobic capacity, few have quantitatively predicted changes in functional aerobic capacity due to CR and investigated the indicators that affect the outcomes of CR. Bargehr and Rubén et al. reported factors influencing improvements in the 6MWD post-CR, but no studies have previously examined the VO₂ peak^10,40. Our research developed the first model to successfully predict VO₂ peak after CR in patients with CVD. Additionally, by exploring additional variables that influence the recovery of the VO₂ peak, our study provides insights that could be utilized in a clinical decision support system for tailoring personalized CR prescriptions based on individual patient characteristics.

The highest-ranked indicator for VO₂ peak recovery was the pre-CR VO₂ peak, which was more than twice as important compared to other indicators in terms of ΔVO₂ peak. Although it showed a positive correlation with post-CR VO₂ peak, it demonstrated a negative correlation with ΔVO₂ peak. This suggests that although a higher initial VO₂ peak indicates a better overall prognosis, the recuperative effect of CR is expected to be greater in patients with an initially lower CRF. Specifically, the SHAP value distribution for ΔVO₂ peak was broad in negative direction, supporting the need for CR even in patients with severe CVD who had very low initial CRF. Conversely, a narrow distribution of SHAP values in the positive direction suggests a ceiling effect in the CR. Similarly, the negative correlation with total visit period supports the existence of this ceiling effect. Therefore, transitioning from active Phase II CR to maintenance Phase III CR at an appropriate time is crucial. Utilizing ML predictions can facilitate the identification of these transition points, enabling the design of cost-effective CR programs.

Among the top-ranked indicators for predicting ΔVO₂ peak, pre-CR VT peak, age, total visit period, KASI, and DBP peak exhibited negative correlations, while resting HR, visit interval and 6MWD showed relatively positive correlations. Similar to the predictions made using the pre-CR VO₂ peak, groups with a lower VT peak, higher resting HR, and lower KASI, indicative of lower baseline CRF and physical function, demonstrated greater recovery shortly after CR. Conversely, a higher baseline function, as indicated by the 6MWD, was correlated with greater recovery.

Interestingly, although the 6MWD was identified as more than twice as important as the other indicators in Task 1, its significance decreased considerably in Task 2. Moreover, the importance of 6MWD and HGS in predicting ΔVO₂ peak was significantly lower compared to their role in predicting post-CR VO₂ peak. This suggests that, while the 6MWD and HGS are critical predictors of the VO₂ peak at the same time point, their relevance diminishes when predicting subsequent recovery. This characteristic may explain why prediction equations using the 6MWD fail to forecast changes following CR¹⁶. On the other hand, the relative importance of KASI in predicting ΔVO₂ peak suggests its potential clinical utility in predicting changes in CRF. These results suggest that the drivers of VO₂ peak recovery post-CR may differ from those at baseline and require unique modeling approaches.

In CR, exercise strategies are one of the primary areas in which healthcare professionals can intervene. Most exercise-related indicators showed a positive correlation with ΔVO₂ peak, revalidating the positive effects of CR. Notably, the duration and total exercise time of the resistance exercises were highly significant indicators. Resistance exercise improves functional performance and prognosis in patients with coronary artery disease (CAD) and HF, and recent guidelines recommend integrating resistance exercise into CR programs⁴¹. However, compared to aerobic exercise, resistance exercise strategies have been less explored⁴². Our study supports the importance of resistance exercise in a comprehensive CVD cohort. Additionally, future AI analyses could facilitate the integration of resistance and aerobic exercises, allowing the prescription of customized CR strategies that are anticipated to maximize CRF improvements.

Comprehensively, our research highlights the importance of using individualized patient data to enhance prediction models. In Task 1, baseline metrics such as the 6MWD, KASI, and HGS emerged as critical predictors of VO₂ peak, affirming their value in functional evaluations. Task 2 revealed a ceiling effect along with the differentiated impact of resistance exercise on patient outcomes. Interestingly, Task 2 showed that the predictors of VO₂ peak recovery during CR may differ from baseline values.

Limitations

Our study had several limitations. First, it was conducted at a single center, and the model requires external validation in a multicenter setting with a broader participant background to confirm its generalizability and effectiveness. Second, the study was retrospective and the interval between patient visits was not strictly controlled. Although patients visiting the outpatient clinic for CR were advised to come every 2–3 months, some did not adhere to these intervals. These visit intervals could be associated with the patients' CR compliance or health status, representing a potential limitation. Third, patients who were unable to undergo CPET were excluded. Those contraindicated for CPET owing to high cardiovascular risk or those with very low initial physical function were excluded. Thus, the applicability of this model to patients with these characteristics requires further validation.

Our study demonstrates the significant utility of ML models in predicting the peak CRF in patients with CVD who underwent CR. By incorporating a comprehensive range of clinical, functional, and exercise-related data, our models have proven superior to traditional VO₂ peak predictive equations in terms of both accuracy and relevance to the specific needs of patients with CVD.

Furthermore, the predictive capability of our model enables a more personalized approach for CR. By accurately forecasting the trajectory of VO₂ peak recovery, clinicians can tailor interventions to maximize individual patient outcomes and enhance the efficiency and effectiveness of rehabilitation programs.

In conclusion, the application of advanced ML techniques marks a transformative advancement in CVD management. Future research should aim to expand these models to include a broader range of patient demographics and longer follow-up periods to improve their generalizability and effectiveness across various clinical settings.

Acknowledgements

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIP) (No. RS-2024-00336696).

Author contributions statement

All authors contributed to the study design and approval of the final version of the manuscript.

All authors contributed to the study design and approval of the final version of the manuscript.

JWS, HBK: Writing - Original Draft, Formal Analysis, Investigation, Validation, Resources

BRK, HKL: Conceptualization, Methodology, Formal Analysis, Investigation, Writing - Review & Editing, Funding Acquisition

JSJ, HJK, HSS: Investigation, Resources, Writing - Review & Editing

JHK, CYP: Formal Analysis, Investigation

Consent statement

Due to the retrospective nature of the study, (IRB of Korea University Medicine) waived the need of obtaining informed consent

Additional information

Competing interests

No potential conflict of interest relevant to this article was reported.

Data availability

All data generated or analyzed during this study are included in this article.

Arnett, D. K. & ACC/AHA Guideline on the Primary Prevention of Cardiovascular Disease. : Executive Summary: A Report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines. J. Am. Coll. Cardiol. 74, 1376–1414. (2019). https://doi.org/10.1016/j.jacc.2019.03.009 (2019).
Oh, M. S. & Jeong, M. H. Sex differences in cardiovascular disease risk factors among Korean adults. Korean J. Med. 95, 266–275. https://doi.org/10.3904/kjm.2020.95.4.266 (2020).
Anderson, L. et al. Exercise-based cardiac rehabilitation for coronary heart disease. Cochrane Database Syst. Rev. CD001800. (2016). https://doi.org/10.1002/14651858.CD001800.pub3 (2016).
Blair, S. N. Physical inactivity: the biggest public health problem of the 21st century. Br. J. Sports Med. 43, 1–2 (2009).
Harber, M. P. et al. Impact of Cardiorespiratory Fitness on All-Cause and Disease-Specific Mortality: Advances Since 2009. Prog Cardiovasc. Dis. 60, 11–20. https://doi.org/10.1016/j.pcad.2017.03.001 (2017).
Myers, J. et al. Physical activity and cardiorespiratory fitness as major markers of cardiovascular risk: their independent and interwoven importance to health status. Prog Cardiovasc. Dis. 57, 306–314. https://doi.org/10.1016/j.pcad.2014.09.011 (2015).
Ross, R. et al. Importance of Assessing Cardiorespiratory Fitness in Clinical Practice: A Case for Fitness as a Clinical Vital Sign: A Scientific Statement From the American Heart Association. Circulation. 134, e653–e699. https://doi.org/10.1161/cir.0000000000000461 (2016).
Buttar, K. K., Saboo, N. & Kacker, S. A review: Maximal oxygen uptake (VO2 max) and its estimation methods. Int. J. Phys. Educ. Sports Health. 6, 24–32 (2019).
Ashfaq, A., Cronin, N. & Müller, P. Recent advances in machine learning for maximal oxygen uptake (VO2 max) prediction: A review. Inf. Med. Unlocked. 28, 100863. https://doi.org/10.1016/j.imu.2022.100863 (2022).
Peterman, J. E. et al. A Nonexercise Prediction of Peak Oxygen Uptake for Patients With Cardiovascular Disease: data from the fitness registry and the importance of exercise international database(FRIEND). J. Cardiopulm. Rehabil Prev. 43, 115–121. https://doi.org/10.1097/hcr.0000000000000722 (2023).
Bargehr, J. et al. Predictors of Suboptimal Gain in Exercise Capacity After Cardiac Rehabilitation. Am. J. Cardiol. 119, 687–691. https://doi.org/10.1016/j.amjcard.2016.08.005 (2017).
Adedoyin, R. A. et al. Prediction of functional capacity during six-minute walk among patients with chronic heart failure. Niger J. Clin. Pract. 13, 379–381 (2010).
Cahalin, L. P., Mathier, M. A., Semigran, M. J., Dec, G. W. & DiSalvo, T. G. The six-minute walk test predicts peak oxygen uptake and survival in patients with advanced heart failure. Chest. 110, 325–332. https://doi.org/10.1378/chest.110.2.325 (1996).
Mandic, S. et al. Estimating exercise capacity from walking tests in elderly individuals with stable coronary artery disease. Disabil. Rehabil. 35, 1853–1858. https://doi.org/10.3109/09638288.2012.759629 (2013).
Ross, R. M., Murthy, J. N., Wollak, I. D. & Jackson, A. S. The six minute walk test accurately estimates mean peak oxygen uptake. BMC Pulm Med. 10, 31. https://doi.org/10.1186/1471-2466-10-31 (2010).
Chirico, D. et al. Using the 6-min Walk Test to Monitor Peak Oxygen Uptake Response to Cardiac Rehabilitation in Patients With Heart Failure. J. Cardiopulm. Rehabil Prev. 40, 378–382. https://doi.org/10.1097/hcr.0000000000000517 (2020).
Fu, L. et al. Perioperative poor grip strength recovery is associated with 30-day complication rate after cardiac surgery discharge in middle-aged and older adults - a prospective observational study. BMC Cardiovasc. Disord. 19, 266. https://doi.org/10.1186/s12872-019-1241-x (2019).
Enright, P. L. The six-minute walk test. Respir Care. 48, 783–785 (2003).
Sung, J. et al. Development of Korean activity scale/index (KASI). Korean Circ. J. 30, 1004–1009 (2000).
Kim, M. H., Cho, Y. S., Uhm, W. S., Kim, S. & Bae, S. C. Cross-cultural adaptation and validation of the Korean version of the EQ-5D in patients with rheumatic diseases. Qual. Life Res. 14, 1401–1406. https://doi.org/10.1007/s11136-004-5681-z (2005).
American College of Sports Medicine. ACSM's guidelines for exercise testing and prescription (Lippincott williams & wilkins, 2013).
American Association of Cardiovascular Pulmonary Rehabilitation. Guidelines for Cardia Rehabilitation and Secondary Prevention Programs-(with Web Resource) (Human Kinetics, 2013).
Kim, S. J., Koh, K., Lustig, M., Boyd, S. & Gorinevsky, D. An Interior-Point Method for Large-Scale ℓ1-Regularized Least Squares. IEEE J. Sel. Top. Signal. Process. 1, 606–617. https://doi.org/10.1109/jstsp.2007.910971 (2007).
Rifkin, R. M. & Lippert, R. A. Notes on regularized least squares in Computer Science and Artificial Intelligence Laboratory Technical Report (2007).
Zhang, T. Solving large scale linear prediction problems using stochastic gradient descent algorithms. in Proceedings of the twenty-first international conference on Machine learning. 116.
Smola, A. J. & Schölkopf, B. A tutorial on support vector regression. Stat. Comput. 14, 199–222. https://doi.org/10.1023/b:Stco.0000035301.49549.88 (2004).
Friedman, J. H. Stochastic gradient boosting. Comput. Stat. Data Anal. 38, 367–378. https://doi.org/10.1016/S0167-9473(01)00065-2 (2002).
Breiman, L. Random forests. Mach. Learn. 45, 5–32. https://doi.org/Doi 10.1023/A:1010933404324 (2001).
Prokhorenkova, L., Gusev, G., Vorobev, A., Dorogush, A. V. & Gulin, A. CatBoost: unbiased boosting with categorical features in 32nd Conference on Neural Information Processing Systems (NIPS) Vol. 31, 6639–6649 (2018).
Chen, T. Q., Guestrin, C., Assoc Comp, M. & XGBoost: A Scalable Tree Boosting System. in 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD). 785–794Assoc Computing Machinery, (2016).
Ke, G. L. et al. LightGBM: A Highly Efficient Gradient Boosting Decision Tree. in 31st Annual Conference on Neural Information Processing Systems (NIPS). (Neural Information Processing Systems (Nips), (2017).
Lundberg, S. M. & Lee, S. I. A Unified Approach to Interpreting Model Predictions. in 31st Annual Conference on Neural Information Processing Systems (NIPS). (Neural Information Processing Systems (Nips), (2017).
Abut, F., Akay, M. F. & George, J. A robust ensemble feature selector based on rank aggregation for developing new VO < sub > 2 max prediction models using support vector machines. Turkish J. Electr. Eng. Comput. Sci. 27, 3648–3664. https://doi.org/10.3906/elk-1808-138 (2019).
Przednowek, K. et al. Predictive Modeling of VO₂max Based on 20 m Shuttle Run Test for Young Healthy People. Appl. Sci. (Basel). 8, 14. https://doi.org/10.3390/app8112213 (2018).
Kim, Y. J. et al. Comparison of the Korean Activity Status Index with cardiopulmonary exercise test in patients with acute myocardial infarction. J. Med. Life Sci. 19, 57–65. https://doi.org/10.22730/jmls.2022.19.2.57 (2022).
Roberts, H. C. et al. A review of the measurement of grip strength in clinical and epidemiological studies: towards a standardised approach. Age Ageing. 40, 423–429. https://doi.org/10.1093/ageing/afr051 (2011).
Wu, Y., Wang, W., Liu, T. & Zhang, D. Association of Grip Strength With Risk of All-Cause Mortality, Cardiovascular Diseases, and Cancer in Community-Dwelling Populations: A Meta-analysis of Prospective Cohort Studies. J. Am. Med. Dir. Assoc. 18 551.e517-551.e535 (2017).
Chang, J. S., Lee, Y. H. & Kong, I. D. Predictive factors of peak aerobic capacity using simple measurements of anthropometry and musculoskeletal fitness in paraplegic men. J. Sports Med. Phys. Fit. 59, 925–933. https://doi.org/10.23736/s0022-4707.18.08531-6 (2019).
Zhou, M. et al. Handgrip Strength-Related Factors Affecting Health Outcomes in Young Adults: Association with Cardiorespiratory Fitness. Biomed. Res. Int. 6645252. (2021). https://doi.org/10.1155/2021/6645252 (2021).
Fuentes Artiles, R. et al. Predictors of gain in exercise capacity through cardiac rehabilitation: Sex and age matter. Heart Lung. 62, 200–206. https://doi.org/10.1016/j.hrtlng.2023.08.003 (2023).
Verdicchio, C. et al. A Clinical Guide for Assessment and Prescription of Exercise and Physical Activity in Cardiac Rehabilitation. A CSANZ Position Statement. Heart Lung Circ. 32, 1035–1048. https://doi.org/10.1016/j.hlc.2023.06.854 (2023).
Kirkman, D. L., Lee, D. C. & Carbone, S. Resistance exercise for cardiac rehabilitation. Prog Cardiovasc. Dis. 70, 66–72. https://doi.org/10.1016/j.pcad.2022.01.004 (2022).

No competing interests reported.

Download PDF

Editor invited by journal
15 Oct, 2024
Submission checks completed at journal
14 Oct, 2024
First submitted to journal
11 Oct, 2024

You are reading this latest preprint version

Predicting peak cardiorespiratory fitness in patients with cardiovascular disease using machine learning

Status:

Version 1

Abstract

Objective

Methods and Results

Conclusion

Figures

Introduction

Methods

Study population

Study design

ML modeling

Results

Participant characteristics

Task 1: prediction of VO₂ peak using clinical data without CPET

Model performance

Feature importance analysis (SHAP)

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information

Model performance

Feature importance analysis (SHAP)

Discussion

Task 1: prediction of VO₂ peak using clinical data without CPET

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information

Limitations

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1

Predicting peak cardiorespiratory fitness in patients with cardiovascular disease using machine learning

Status:

Version 1

Abstract

Objective

Methods and Results

Conclusion

Figures

Introduction

Methods

Study population

Study design

ML modeling

Results

Participant characteristics

Task 1: prediction of VO2 peak using clinical data without CPET

Model performance

Feature importance analysis (SHAP)

Task 2: Prediction of VO2 peak using clinical data with CPET and CR information

Model performance

Feature importance analysis (SHAP)

Discussion

Task 1: prediction of VO2 peak using clinical data without CPET

Task 2: Prediction of VO2 peak using clinical data with CPET and CR information

Limitations

Conclusion

Declarations

References

Additional Declarations

Status:

Version 1

Task 1: prediction of VO₂ peak using clinical data without CPET

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information

Task 1: prediction of VO₂ peak using clinical data without CPET

Task 2: Prediction of VO₂ peak using clinical data with CPET and CR information