Multimodal model for predicting exercise-induced pulmonary hypertension validated by invasive exercise hemodynamics: a prospective study

Abstract

Background

Exercise-induced pulmonary hypertension (EiPH) represents an early stage of pulmonary vascular disease that remains challenging to identify noninvasively, particularly in patients with borderline resting haemodynamics. We aimed to develop and validate a multimodal non-invasive model integrating clinical characteristics, exercise echocardiography, and cardiopulmonary exercise testing (CPET) to accurately predict invasively confirmed EiPH.

Methods

This prospective cohort study consecutively enrolled adult patients who presented with exercise limitation following chronic pulmonary artery thrombosis or who exhibited increased tricuspid regurgitation velocity on transthoracic echocardiography. All patients underwent comprehensive clinical evaluation, stress echocardiography, CPET, and invasive exercise right heart catheterization as the diagnostic gold standard. To minimize the impact of physiological variability on the results, all three tests were conducted at the same time. Feature selection was conducted sequentially using Spearman correlation analysis, statistical testing, and LASSO regression to identify core features associated with EiPH. Subsequently, a logistic regression model with elastic net regularization was used. Five-fold cross-validation and grid search were employed to optimize model parameters and evaluate the diagnostic performance of different models. In addition, the DeLong test was used to assess whether the AUC differed significantly between models. Finally, subgroup analyses were performed to validate the robustness of the model.

Results

The study included a total of 78 patients, comprising 34 cases of EiPH and 44 cases without EiPH. Patients in the EiPH group were significantly older (68.5 vs. 51.5 years, P < 0.001) and demonstrated reduced exercise capacity, impaired pulmonary function, abnormal right ventricular function, and increased pulmonary vascular resistance. After integrating multimodal data from clinical features, stress echocardiography, and cardiopulmonary exercise testing, the Clinical+Echo+CPET model achieved the best performance, with an AUC of 0.951 (95% CI: 0.902–0.987), an accuracy of 0.871, and a Brier score of 0.128, indicating strong discriminative ability and good calibration. The model maintained high stability in the internal validation cohort, with an AUC of 0.900, a sensitivity of 0.833, and a specificity of 0.700. The DeLong test showed that the multimodal model (Clinical+Echo+CPET) had superior discriminative performance compared with the unimodal models (Clinical, Echo, or CPET). Subgroup analysis demonstrated that the model maintained good diagnostic performance across different age and sex groups.

Conclusions

A non-invasive multimodal model integrating clinical indicators, exercise echocardiography, and cardiopulmonary metabolic parameters can reliably identify EiPH.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12931-026-03814-z.

Introduction

Pulmonary hypertension (PH) is a progressive and highly heterogeneous pulmonary vascular disorder characterized by a sustained increase in right ventricular (RV) afterload, ultimately leading to right heart failure and death [1]. Despite substantial advances in targeted therapies for PH in recent years, the long-term prognosis remains poor, largely because irreversible structural and functional remodeling of the right ventricle has often occurred by the time of clinical diagnosis [2, 3]. Accordingly, current research in PH has increasingly shifted toward the early stages of pulmonary vascular disease. Exercise-induced pulmonary hypertension (EiPH) represents an early disease phenotype in which resting hemodynamics remain within normal limits but abnormal responses become evident during physiologic stress such as exercise [4]. The 2022 European Society of Cardiology/European Respiratory Society (ESC/ERS) guidelines formally reintroduced the hemodynamic definition of EiPH based on an abnormal increase in the slope of mean pulmonary arterial pressure relative to cardiac output (mPAP/CO slope) during exercise [5, 6]. This revised framework emphasizes the dynamic nature of the pressure–flow relationship in the pulmonary circulation and highlights early exhaustion of pulmonary vascular reserve (PVR) together with stress-induced impairment of right ventricle–pulmonary artery (RV–PA) coupling [7].

Accurate identification of EiPH remains challenging in routine clinical practice. Invasive exercise right heart catheterization (eRHC) is considered the reference standard for diagnosing EiPH. However, its invasive nature, technical complexity, and high resource utilization limit its use as a screening tool [8, 9]. Conventional noninvasive methods also have important limitations when used alone. Resting transthoracic echocardiography often fails to detect subtle abnormalities in RV–PA coupling, and nonspecific symptoms such as exertional dyspnea may overlap with a wide range of cardiopulmonary disorders [10]. Exercise stress echocardiography (ESE) can directly assess RV mechanical reserve and dynamic RV–PA coupling, whereas cardiopulmonary exercise testing (CPET) quantifies ventilatory efficiency and global cardiopulmonary metabolic reserve [11, 12]. However, EiPH is a multidimensional pathophysiologic condition involving abnormal pulmonary hemodynamics, chronotropic incompetence, and impaired metabolic efficiency. A single noninvasive modality therefore provides only a partial view of this dynamic process and may result in underrecognition of subclinical pulmonary vascular dysfunction.

Given the complexity of EiPH pathophysiology, a multidimensional noninvasive assessment strategy is needed. Recently, Martín et al. demonstrated that a multiparametric noninvasive score based on CPET and exercise echocardiography showed promising performance for risk stratification in EiPH [10]. To further clarify the physiological relationship between noninvasive predictors and invasive reference standards, ESE, CPET, and eRHC were performed simultaneously at the same location during the same exercise protocol, thereby eliminating time intervals among the three examinations and minimizing bias related to nonsynchronous measurements. In addition, RV strain parameters and their dynamic changes during exercise were incorporated into the analysis. On this basis, we aimed to develop and internally validate a noninvasive multimodal prediction model for EiPH, with the goal of providing a reliable tool for the early detection and risk stratification of pulmonary vascular dysfunction.

Methods

Study design and participants

This study was a prospective observational study. We consecutively enrolled adult patients who presented with exercise limitation due to chronic pulmonary artery thrombosis or who had increased tricuspid regurgitation velocity (TRV) as measured by transthoracic echocardiography. All participants underwent a comprehensive evaluation in a clinically stable condition. CPET, ESE, and eRHC were performed during the same evaluation period to minimize the influence of physiologic variability. A total of 119 patients were enrolled between January 2024 and April 2026. The study was approved by the Ethics Committee of Beijing Chaoyang Hospital, Capital Medical University (approval number 2023-KE-264), and all participants provided written informed consent.

Inclusion criteria were age ≥ 18 years, presence of exertional dyspnoea, and the ability to complete exercise testing. Exclusion criteria included clinically diagnosed heart failure, including left ventricular ejection fraction (LVEF) < 50%, heart failure with preserved ejection fraction (HFpEF) defined as resting pulmonary capillary wedge pressure (PCWP) ≥ 15 mmHg, or exercise-induced HFpEF defined as peak exercise PCWP ≥ 25 mmHg; moderate-to-severe valvular heart disease; severe parenchymal lung disease; acute pulmonary embolism; severe arrhythmia; or contraindications to exercise testing.

Exercise haemodynamic assessment

All participants underwent eRHC via venous access using a Swan–Ganz catheter. A symptom-limited graded exercise test was performed in the supine position using a cycle ergometer. The exercise load is increased gradually, rising by 30 watts every 2 min while maintaining a cadence of 55–65 revolutions per minute. Hemodynamic data were recorded at each stage until symptom limitation or maximal tolerance was reached [13]. At rest and at each exercise stage, the following parameters were continuously measured and recorded simultaneously: central venous pressure, mPAP, PCWP, right atrial pressure (RAP), and right ventricular pressure. CO was measured using the thermodilution method, with at least three measurements obtained at each stage and averaged. When significant tricuspid regurgitation was present or thermodilution measurements were unreliable, the direct Fick method was used for correction. PVR and cardiac index (CI) were calculated at each workload. The mPAP/CO slope was calculated using linear regression based on multipoint exercise data. EiPH was defined as an mPAP/CO slope > 3 mmHg/L/min calculated from resting and peak data [5, 9].

Exercise stress echocardiography

All participants underwent ESE at the time of eRHC. Echocardiographic examinations were performed using a Philips EPIQ 7 C ultrasound system equipped with an X5-1 phased-array transducer, with a frequency range of 1–5 MHz. RV function was assessed using an apical four-chamber view focused on the RV. Images were acquired at rest, during peak exercise, and during recovery, and tricuspid regurgitation spectra were recorded using continuous-wave Doppler. During exercise stress, because of increased respiratory motion, accelerated heart rate, and a shortened image acquisition window, echocardiographic images were acquired as far as possible after each workload stage had reached a relatively steady state. For conventional two-dimensional and Doppler measurements, consecutive cardiac cycles with the best image quality, stable cardiac rhythm, and no obvious respiratory artefacts were preferentially selected for analysis. Multiple high-quality cardiac cycles available within the same workload stage were measured and averaged to improve the reliability and representativeness of measurements obtained during exercise.

Offline echocardiographic image analysis was performed using TomTec software, and analysts were blinded to invasive haemodynamic data. Right heart-related parameters included RV end-diastolic area, RV end-systolic area, fractional area change, tricuspid annular plane systolic excursion (TAPSE), septal mitral annular early diastolic velocity (e′), the ratio of septal transmitral early diastolic inflow velocity to septal mitral annular early diastolic velocity (E/e′), tissue Doppler systolic peak velocity (S′), right ventricular free-wall longitudinal strain (RVFWSL), right ventricular 4-chamber longitudinal strain (RV4CSL), right atrial conduit strain (RAScd), right atrial reservoir strain (RASr), and right atrial contraction strain (RASct). Strain analysis was performed using two-dimensional speckle-tracking echocardiography. All exercise stress images included in the strain analysis had a frame rate of 50 frames per second. The starting point of the cardiac cycle for strain analysis was determined according to the simultaneously recorded electrocardiogram. For RVFWSL analysis, the R wave was used as the reference point for the onset of the ventricular cardiac cycle, whereas for right atrial strain analysis, the P wave was used as the reference point for the onset of the atrial cardiac cycle. All strain curves were manually reviewed, and endocardial borders were adjusted manually when necessary to ensure tracking quality.

During exercise, peak TRV was measured by continuous-wave Doppler, and the tricuspid regurgitation pressure gradient was calculated using the simplified Bernoulli equation as 4×TRV². Pulmonary artery systolic pressure (PASP) was obtained by adding estimated RAP to the tricuspid regurgitation pressure gradient. RAP at rest and during exercise was estimated based on inferior vena cava diameter and inspiratory collapsibility: RAP was assigned as 3 mmHg when the inferior vena cava diameter was < 21 mm with inspiratory collapse > 50%; 15 mmHg when the inferior vena cava diameter was > 21 mm with inspiratory collapse < 50%; and 8 mmHg in all other cases. Accordingly, PASP was calculated as follows: PASP = 4×TRV² + RAP [14].

The TAPSE/PASP ratio was calculated at each exercise stage as a noninvasive index of RV–PA coupling. Dynamic reserve indices included peak exercise values and their changes (Δ). The change value was defined as the difference between the peak exercise value and the resting value, reflecting RV systolic functional reserve and the compensatory capacity of the RV in response to exercise-induced increases in pulmonary vascular load.

Cardiopulmonary exercise test

All subjects underwent CPET on the same step-test platform, with continuous breath-by-breath gas exchange analysis and heart rate monitoring. The recorded CPET variables included maximal oxygen uptake (VO₂max), VO₂max expressed as a percentage of the predicted value (% predicted VO₂max), peak heart rate, and the percentage of the achieved workload relative to the predicted workload. The anaerobic threshold was determined using the V-slope method in conjunction with ventilatory parameters.

The minute ventilation versus carbon dioxide output slope (VE/VCO₂ slope) and the minute ventilation versus oxygen uptake slope (VE/VO₂ slope) were automatically calculated by the CPET system using breath-by-breath data obtained from rest to peak exercise. Specifically, the VE/VCO₂ slope was derived from the linear regression relationship between minute ventilation (VE) and carbon dioxide output (VCO₂), whereas the VE/VO₂ slope was derived from the linear regression relationship between VE and oxygen uptake (VO₂). CPET data were analyzed by certified personnel who were blinded to the invasive hemodynamic and echocardiographic results.

Feature selection

To reduce the impact of multicollinearity on model stability, Spearman correlation analysis was first performed on candidate features. For variable pairs with an absolute correlation coefficient > 0.7, the variable more strongly correlated with the outcome was preferentially retained, and only features with statistical significance were included in subsequent analyses. On this basis, feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression with L1 regularization. The regularization parameter λ was optimized within a predefined parameter space using ten-fold cross-validation (λ = 1/C), and the optimal model was determined using the minimum error or 1-SE criterion to balance predictive performance and model parsimony. To improve robustness, LASSO analysis was repeatedly implemented in the training cohort, and feature selection stability was evaluated by calculating the selection frequency of variables. All the above procedures were conducted only in the training cohort to avoid information leakage.

Model construction and evaluation

After feature selection, predictive models were constructed based on the selected features. All continuous variables were standardized (Z-score normalization) before modeling to eliminate the influence of scale differences on parameter estimation. This study employs a logistic regression model with elastic net regularization. Model parameters were optimized using five-fold stratified cross-validation combined with grid search, with tuning parameters including the regularization strength parameter C and the L1 ratio parameter. Class weights were introduced during model training to correct for bias caused by class imbalance. Finally, three levels of models were sequentially constructed to evaluate incremental diagnostic value: unimodal models (clinical, Echo, CPET), bimodal models (clinical + CPET), and a comprehensive trimodal model (clinical + CPET + Echo). In the comprehensive trimodal model, only the top two features with the highest absolute coefficients from each category were included.

After building the model, we conducted subgroup analyses to assess the model’s robustness across predefined clinical strata (gender; age < 60 years vs. ≥60 years) and calculated interaction P-values to rule out significant heterogeneity in predictive performance.

Statistical analysis

Continuous variables were expressed as the mean ± standard deviation or median (interquartile range), according to their distribution, and categorical variables were expressed as frequencies and percentages. Normality was assessed using the Shapiro–Wilk test. Between-group comparisons were performed using the independent-samples t-test or the Mann–Whitney U test, while categorical variables were analyzed using the χ² test or Fisher’s exact test. Correlations between variables were assessed using Spearman’s rank correlation analysis. In the training cohort, receiver operating characteristic (ROC) curves were constructed based on predicted probabilities, and the area under the curve (AUC) was calculated. The optimal cutoff value was determined using the Youden index, and the 95% confidence interval of the AUC was estimated using bootstrap resampling with 1,000 iterations. This cutoff value was then fixed in the internal validation cohort to avoid bias from threshold re-optimization. In the internal validation cohort, the AUC, accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1-score, and Brier score were further calculated, and classification performance was evaluated using a confusion matrix. The DeLong test was used to assess whether differences in AUC between models were statistically significant. Subgroup consistency was examined by introducing interaction terms into multivariable models. All statistical analyses were performed using SPSS 25.0 (IBM) and Python 3.12, and a two-sided P < 0.05 was considered statistically significant.

Results

Study population characteristics

Figure 1 shows the patient enrollment flowchart for this study. A total of 78 patients were enrolled in the study and divided into two groups: EiPH (n = 34) and no EiPH (n = 44). All enrolled patients had resting PCWP < 15 mmHg and peak exercise PCWP < 25 mmHg. The two groups were generally similar in demographic characteristics, with no statistically significant differences in sex distribution or disease composition. However, the EiPH group was significantly older (68.50 vs. 51.50 years, P < 0.001), and NT-proBNP levels were significantly elevated, suggesting potential impairment of cardiac function. In terms of pulmonary function, there were no significant differences in FVC and FEV₁ between the two groups, but FEV₁/FVC and diffusing capacity (DLCO and DLCO/VA) were significantly reduced in the EiPH group, indicating decreased ventilation and gas exchange efficiency (Table 1).

Fig. 1.

Flowchart of study population selection and grouping according to EiPH status. EiPH: exercise-induced pulmonary hypertension; mPAP: mean pulmonary artery pressure

Table 1.

Comparison of baseline demographics, pulmonary function, and resting/exercise hemodynamic characteristics between patients with and without exercise-induced pulmonary hypertension

Variables	ALL (N = 78)	No EiPH (N = 44)	EiPH (N = 34)	P
Demographics
Sex, male	30 (38.5%)	19 (43.2%)	11 (32.4%)	0.459
Diagnosis				0.117
CTEPD	21 (26.9%)	14 (31.8%)	7 (20.6%)
CTEPH	42 (53.8%)	20 (45.5%)	22 (64.7%)
CTD-PAH	8 (10.3%)	7 (15.9%)	1 (2.9%)
Resting TRV ≥ 2.8 m/s	7 (9.0%)	3 (6.8%)	4 (11.8%)
WHO, FC				0.121
Class I	34 (43.6%)	23 (52.3%)	11 (32.4%)
Class II	43 (55.1%)	20 (45.5%)	23 (67.6%)
Class III	1 (1.3%)	1 (2.3%)	0 (0.0%)
Age, years	61.00 (49.25, 69.75)	51.50 (45.75, 61.75)	68.50 (63.00, 73.00)	< 0.001
BMI, kg/m2	25.50 ± 3.90	26.30 ± 3.97	24.46 ± 3.62	0.038
6MWD, m	462.60 (422.00, 504.42)	463.39 (435.00, 500.02)	460.22 (415.50, 503.25)	0.354
NT-proBNP, pmol/L	65.45 (39.90, 95.72)	46.60 (34.17, 76.32)	78.55 (62.88, 151.70)	0.001
Peak HR, bpm	130.72 ± 21.81	139.82 ± 21.15	118.94 ± 16.55	< 0.001
Resting HR, bpm	75.00 (66.00, 84.00)	78.00 (70.00, 85.25)	70.50 (65.00, 79.75)	0.188
dBP, mmHg	78.50 (72.25, 89.00)	83.00 (74.50, 89.25)	74.50 (69.50, 87.25)	0.089
sBP, mmHg	91.36 ± 11.86	93.05 ± 11.62	89.18 ± 11.97	0.154
SpO2, %	96.00 (95.00, 98.00)	96.00 (95.00, 97.00)	96.00 (95.00, 98.00)	0.819
Pulmonary Function Test
FVC, % pred	109.26 ± 15.78	106.34 ± 14.28	113.04 ± 17.00	0.063
FEV1, % pred	98.24 ± 15.05	97.26 ± 14.00	99.51 ± 16.44	0.515
FEV1%FVC, %	89.30 (85.12, 96.55)	91.70 (86.80, 97.33)	87.10 (83.40, 90.57)	0.004
DLCOc SB, mL/min/mmHg	83.35 (77.25, 89.30)	84.81 (78.69, 93.88)	79.95 (76.53, 84.15)	0.020
DLCOc/VA, mL/min/mmHg/L	86.35 (78.98, 93.28)	90.40 (82.17, 95.95)	82.19 (78.80, 91.44)	0.039
TLC, % pred	99.14 ± 12.59	98.24 ± 12.19	100.31 ± 13.19	0.477
Rest and Exercise Right Heart Catheterization
SvO2, %	72.97 ± 5.52	74.63 ± 5.43	70.81 ± 4.92	0.002
TPR, Wood Units	2.37 (1.80, 2.90)	2.15 (1.61, 2.55)	2.79 (2.20, 3.30)	< 0.001
Resting mPAP, mmHg	16.00 (15.00, 18.75)	16.00 (14.00, 18.00)	17.00 (16.00, 19.00)	0.190
CVP, mmHg	3.00 (2.00, 4.00)	3.00 (2.00, 4.00)	3.00 (2.00, 3.00)	0.042
RAP, mmHg	3.00 (2.00, 4.00)	3.00 (2.00, 4.00)	2.50 (2.00, 3.00)	0.051
RVP, mmHg	9.86 ± 2.21	9.93 ± 2.30	9.76 ± 2.12	0.743
PCWP, mmHg	7.00 (5.00, 8.00)	7.00 (5.75, 8.25)	7.00 (5.00, 8.00)	0.859
Peak PCWP, mmHg	13.00 (10.00, 16.00)	11.50 (9.00, 15.25)	14.00 (12.00, 17.75)	0.013
CO, L/min	5.58 ± 1.23	6.12 ± 1.16	4.87 ± 0.91	< 0.001
CI, L/min/m²	3.19 ± 0.65	3.41 ± 0.63	2.91 ± 0.56	< 0.001
Peak mPAP, mmHg	38.27 ± 9.83	35.50 ± 9.35	41.85 ± 9.39	0.004
ΔmPAP/ΔCO, mmHg/L/min	2.91 (2.11, 3.47)	2.15 (1.38, 2.55)	3.58 (3.23, 4.00)	< 0.001

6MWD 6 min walk distance, BMI Body mass index, BSA Body surface area, CI Cardiac index, CO Vardiac output; CVP Central venous pressure, CTD-PAH Connective Tissue Disease-associated Pulmonary Arterial Hypertension, CTEPD Chronic Thromboembolic Pulmonary Disease, CTEPH Chronic Thromboembolic Pulmonary Hypertension, dBP Diastolic blood pressure, DLCOc SB Diffusing capacity of the lung for carbon monoxide corrected for hemoglobin, single breath, DLCOc/VA transfer coefficient of the lung for carbon monoxide corrected for hemoglobin, EiPH Exercise-induced pulmonary hypertension, FEV₁: Forced expiratory volume in 1 s, FVC Forced vital capacity, HR Heart rate, mPAP mean pulmonary artery pressure, mPAP/CO slope slope of the mean pulmonary artery pressure–cardiac output relationship, NT-proBNP N-terminal pro-B-type natriuretic peptide, PCWP Ppulmonary capillary wedge pressure, RAP Right atrial pressure, RVP Right ventricular pressure, sBP systolic blood pressure, SpO₂ peripheral oxygen saturation, SvO₂ Mixed Venous Oxygen Saturation, TLC Total lung capacity, TPR Total pulmonary resistance, WHO-FCx World Health Organization functional class

Δ: change from rest to peak exercise

Resting echocardiographic parameters showed no substantial between-group differences, except for a reduced septal e’ in the EiPH group. In contrast, stress echocardiographic indices demonstrated more pronounced differentiation: peak TRV was significantly increased in the EiPH group, while right heart systolic function indices (such as Peak TAPSE, S’, and TAPSE/PASP) were significantly decreased, indicating impaired RV–PA coupling function. In addition, reserve indices such as ΔTAPSE and ΔS’ were also significantly reduced (Table 2). CPET results showed that exercise capacity was significantly decreased in the EiPH group, with a significant reduction in VO₂max, while ventilatory efficiency indices VE/VCO₂ slope and VE/VO₂ slope were increased. In terms of hemodynamics, the EiPH group exhibited significantly decreased CO and cardiac index, while pulmonary vascular resistance was significantly increased. These results indicate that EiPH patients exhibit more pronounced impairment of PVR and right heart functional decompensation under exercise conditions.

Table 2.

Comparison of transthoracic echocardiographic, stress echocardiographic, and cardiopulmonary exercise testing parameters between patients with and without exercise-induced pulmonary hypertension

Variables	ALL (N = 78)	No EiPH (N = 44)	EiPH (N = 34)	P
Transthoracic Echocardiography
EDV, mL	93.00 (83.00, 103.75)	97.00 (85.75, 108.00)	90.00 (83.00, 101.75)	0.141
ESV, mL	30.71 ± 6.14	31.25 ± 5.77	30.00 ± 6.61	0.376
LVEF, %	67.50 (64.00, 71.00)	68.00 (64.75, 71.00)	66.50 (64.00, 72.75)	0.746
Septal e’, cm/s	7.00 (6.00, 8.00)	7.00 (6.50, 8.00)	6.40 (5.22, 7.07)	0.005
Septal E/e’	9.95 ± 2.26	9.83 ± 2.08	10.11 ± 2.49	0.582
RV basal diameter, mm	33.00 (31.00, 35.00)	33.00 (31.00, 35.00)	32.50 (31.00, 34.00)	0.435
MPA, mm	27.00 (24.00, 29.75)	27.00 (23.75, 28.25)	26.00 (24.00, 30.00)	0.746
RVEDA, cm²	22.40 ± 4.86	23.29 ± 4.67	21.24 ± 4.92	0.063
RVESA, cm²	13.98 ± 3.82	14.29 ± 3.59	13.58 ± 4.13	0.425
FAC, %	38 ± 11	39 ± 11	36 ± 11	0.346
MPI	0.48 (0.41, 0.57)	0.47 (0.41, 0.55)	0.52 (0.41, 0.58)	0.292
Stress Echocardiography
Resting TAPSE, mm	21.00 (19.00, 22.97)	20.16 (18.61, 23.00)	21.62 (20.00, 22.79)	0.276
Resting S’, cm/s	12.90 (11.05, 15.38)	13.00 (11.97, 16.00)	12.60 (11.00, 14.30)	0.463
Resting TRV, cm/s	223.31 ± 54.76	208.90 ± 53.67	241.95 ± 51.08	0.007
Resting RVFWSL, %	-21.94 ± 4.05	-21.86 ± 3.88	-22.04 ± 4.30	0.848
Resting RV4CSL, %	-18.00 (-20.58, -15.60)	-17.25 (-20.02, -15.50)	-19.10 (-21.30, -16.80)	0.126
Resting RASr, %	21.75 (18.15, 27.98)	21.30 (18.30, 26.85)	23.15 (18.00, 31.10)	0.604
Resting RAScd, %	-12.55 (-18.60, -9.00)	-11.55 (-17.45, -8.93)	-13.70 (-18.60, -9.62)	0.569
Resting RASct, %	-10.08 ± 4.54	-9.73 ± 4.22	-10.54 ± 4.96	0.441
Resting TAPSE/PASP, mm/mmHg	1.05 (0.77, 1.58)	1.33 (0.83, 1.73)	0.87 (0.70, 1.25)	0.017
Peak TAPSE, mm	23.94 ± 3.28	24.75 ± 2.70	22.90 ± 3.69	0.013
Peak S’, cm/s	19.33 ± 4.16	20.32 ± 4.24	18.04 ± 3.72	0.015
Peak TRV, cm/s	332.50 (252.75, 372.00)	282.20 (234.25, 347.68)	361.50 (331.75, 383.20)	< 0.001
Peak RVFWSL, %	-24.34 ± 4.91	-25.62 ± 4.14	-22.69 ± 5.38	0.008
Peak RV4CSL, %	-20.49 (-23.80, -16.92)	-21.60 (-23.80, -18.70)	-19.45 (-24.02, -15.10)	0.171
Peak RASr, %	25.85 ± 6.63	26.76 ± 6.75	24.67 ± 6.37	0.169
Peak RAScd, %	-14.30 (-16.15, -10.53)	-14.60 (-16.35, -11.65)	-13.30 (-15.85, -10.07)	0.185
Peak RASct, %	-12.93 ± 4.32	-13.57 ± 4.31	-12.11 ± 4.26	0.142
Peak TAPSE/PASP, mm/mmHg	0.52 (0.40, 0.96)	0.80 (0.50, 1.12)	0.42 (0.38, 0.52)	< 0.001
ΔTAPSE, mm	3.00 (0.19, 5.92)	4.00 (2.12, 6.79)	1.52 (-1.00, 4.00)	0.007
ΔS’, cm/s	5.81 ± 4.43	5.82 ± 4.44	5.76 ± 4.54	0.956
ΔTRV, cm/s	93.97 ± 53.08	80.69 ± 44.95	111.15 ± 58.32	0.011
ΔRASct, %	-2.85 ± 5.78	-3.84 ± 5.86	-1.58 ± 5.50	0.087
ΔRAScd, %	0.04 ± 8.90	-0.78 ± 8.37	1.11 ± 9.56	0.356
ΔRASr, %	3.00 (-3.85, 7.50)	5.60 (-3.75, 8.72)	1.90 (-3.93, 6.38)	0.143
ΔRV4CSL, %	-2.00 (-6.40, 2.10)	-3.90 (-7.25, 0.98)	0.70 (-4.17, 2.27)	0.059
ΔRVFWSL, %	-2.40 ± 5.90	-3.76 ± 5.71	-0.65 ± 5.75	0.020
ΔTAPSE/PASP, mm/mmHg	-0.38 (-0.79, -0.22)	-0.38 (-0.84, -0.17)	-0.37 (-0.72, -0.27)	0.736
Cardiopulmonary Exercise Testing
PRP, %	0.96 (0.77, 1.07)	0.89 (0.74, 1.04)	1.02 (0.85, 1.10)	0.073
HRR 3 min	28.00 (21.00, 41.25)	33.00 (26.50, 45.25)	24.00 (16.50, 28.75)	< 0.001
VO2max, mL/min	1059.00 (881.75, 1290.00)	1234.50 (1011.25, 1457.25)	950.50 (810.75, 1081.00)	< 0.001
VO2 peak %pred, %	71.02 ± 13.22	71.79 ± 12.41	70.03 ± 14.33	0.564
VO2(AT) %pred, %	43.95 ± 11.47	42.79 ± 11.37	45.46 ± 11.59	0.311
Peak HR, bpm	130.72 ± 21.81	139.82 ± 21.15	118.94 ± 16.55	< 0.001
VE/VCO2 slope	38.84 ± 4.94	37.47 ± 4.37	40.62 ± 5.12	0.005
VE/VO2 slope	53.25 ± 9.80	51.20 ± 8.80	55.90 ± 10.52	0.035

%pred percentage of predicted, AT Anaerobic threshold, E Early diastolic peak transmitral flow velocity, e’ early diastolic mitral annular velocity, EDV End-diastolic volume, ESV End-systolic volume, FAC Fractional area change, HR Heart rate, HRR 3 min 3-minute Heart Rate Recovery, LVEF Left ventricular ejection fraction, MPA Main pulmonary artery, MPI Myocardial performance index, PASP Pulmonary artery systolic pressure, PRP Predicted Work Rate Percentage, RAScd Right atrial conduit strain, RASct Right atrial contractile strain, RASr Right atrial reservoir strain, RV Right ventricular, RV4CSL Right ventricular 4-chamber longitudinal strain, RVEDA Right ventricular end-diastolic area, RVESA Right ventricular end-systolic area, RVFWSL Right ventricular free wall longitudinal strain, S’ peak systolic tricuspid annular velocity, TAPSE Tricuspid annular plane systolic excursion, TRV Tricuspid regurgitation velocity, VE/VCO₂ slope Minute ventilation versus carbon dioxide output slope, VE/VO₂ slope Minute ventilation versus oxygen uptake slope, VO₂max maximal oxygen consumption

Δ: change from rest to peak exercise

This study also compared baseline data between the training set and the validation set (e-Table 1). Overall, the training set (n = 62) and the validation set (n = 16) were comparable in terms of demographic characteristics and most clinical features. There were no statistically significant differences in gender distribution or WHO functional classification. The age distributions of the two groups were similar, but the validation group had a higher BMI, and there were also differences in blood pressure levels. In exercise echocardiography, most parameters showed no significant differences, with only Peak RASct and ΔRASr exhibiting statistical differences. CPET revealed that the validation group had significantly lower VO₂ peak %pred and VO₂(AT) %pred, while VE/VO₂ slope was elevated, suggesting reduced exercise tolerance and impaired ventilatory efficiency. No significant differences were observed in resting right heart catheterization parameters between the two groups. Overall, the validation set exhibited more pronounced abnormalities in certain metabolic and functional parameters; however, the baseline characteristics of the two groups remained generally comparable.

Feature selection

During the feature selection process, after removing variables through Spearman correlation analysis and statistical testing, a total of 17 candidate features were retained for subsequent LASSO analysis. After LASSO analysis, 9 variables remained. Ranked in descending order by the absolute value of coefficients, they were: FEV₁/FVC, VO₂max, VE/VO₂ slope, Peak TAPSE/PASP, HRR 3min, Age, ΔTRV, Peak S’, and PRP. Among them, FEV₁/FVC showed the strongest negative correlation, while VE/VO₂ slope was the most important positively correlated predictor (e-Figs. 1, 2 and 3). Overall, the selected variables covered clinical characteristics, echocardiographic structural and functional indices, and cardiopulmonary exercise metabolic parameters, suggesting that structural remodeling, RV–PA coupling abnormalities, and decreased exercise cardiopulmonary compensatory capacity jointly contribute to the prediction of EiPH.

Model performance evaluation

In performance evaluation, all models showed a consistent trend of performance improvement in both the training and internal validation cohorts (Fig. 2). In the training cohort, unimodal models already demonstrated good discriminative ability, among which the CPET model performed best (AUC = 0.872, 95% CI: 0.782–0.948), outperforming the clinical model (AUC = 0.854) and the echocardiographic model (AUC = 0.814). With the integration of multimodal information, model performance was further improved: the AUC of the Clinical+Echo model increased to 0.945, and the Clinical+Echo+CPET model performed best, with an AUC of 0.951 (95% CI: 0.902–0.987), an accuracy of 0.871, and a relatively low Brier score, indicating advantages in both discrimination and calibration. In the internal validation cohort, the echocardiographic model performed relatively well (AUC = 0.833), while the performance of the CPET model slightly declined (AUC = 0.717). The multimodal model still demonstrated more stable overall performance, among which the Clinical+Echo+CPET model had an AUC of 0.900 (95% CI: 0.683–1.000), sensitivity and specificity of 0.833 and 0.700, respectively, an accuracy of 0.750, and a Brier score of 0.172 (Table 3).

Fig. 2.

Diagnostic performance for predicting exercise-induced pulmonary hypertension and confusion matrices across the training and internal validation cohorts A-B Receiver operating characteristic (ROC) curves of the five models in the training (A) and internal validation (B) cohorts C-L Confusion matrices corresponding to the five models in the training and internal validation sets. CPET: cardiopulmonary exercise testing; Echo: echocardiography; EiPH: exercise-induced pulmonary hypertension; AUC: area under the curve; CI: confidence interval

Table 3.

Diagnostic performance of various models in the training and internal validation cohorts

Dataset	Model	Cutoff	AUC (95% CI)	Sen	Spe	Accuracy	PPV	NPV	Brier Score
Training cohort	Clinical	0.731	0.854 (0.755, 0.936)	0.607	0.971	0.806	0.944	0.750	0.159
	Echo	0.490	0.814 (0.699, 0.917)	0.821	0.706	0.758	0.697	0.828	0.178
	CPET	0.630	0.872 (0.782, 0.948)	0.750	0.912	0.839	0.875	0.816	0.145
	Clinical+Echo	0.349	0.945 (0.893, 0.985)	1.000	0.735	0.855	0.757	1.000	0.097
	Clin+Echo+CPET	0.482	0.951 (0.902, 0.987)	1.000	0.765	0.871	0.778	1.000	0.128
Internal validation cohort	Clinical	0.731	0.800 (0.491, 1.000)	0.500	0.900	0.750	0.750	0.750	0.207
	Echo	0.490	0.833 (0.603, 1.000)	0.833	0.600	0.688	0.556	0.857	0.180
	CPET	0.630	0.717 (0.344, 1.000)	0.667	0.800	0.750	0.667	0.800	0.221
	Clinical+Echo	0.349	0.817 (0.484, 1.000)	0.833	0.400	0.563	0.455	0.800	0.230
	Clin+Echo+CPET	0.482	0.900 (0.683, 1.000)	0.833	0.700	0.750	0.625	0.875	0.172

AUC Area under the receiver operating characteristic curve, CI Confidence interval, Clin Clinical, CPET Cardiopulmonary exercise testing, Echo Echocardiography, NPV Negative Predictive Value, PPV Positive Predictive Value, Sen Sensitivity, Spe Specificity

The DeLong test results showed that the multimodal model (Clinical+Echo+CPET) had superior discriminative performance compared with unimodal models (Clinical, Echo, or CPET), with statistically significant differences observed in comparisons with the Clinical model (P = 0.013), Echo model (P = 0.008), and CPET model (P = 0.043) (e-Table 2). e-Table 3 summarizes the application of all models. As shown in Fig. 3, the multimodal model generated individualized probability scores based on noninvasive parameters, enabling accurate risk stratification and correctly distinguishing true positive EiPH cases (P = 0.566) and true negative cases (P = 0.324).

Fig. 3.

Clinical Translation and Representative Case Illustrations of the Multimodal EiPH Predictive Model A The stepwise diagnostic inference workflow. The model calculates the log-odds (Logit(P)) using six non-invasive predictors (VO₂max, FEV₁/FVC, Peak TAPSE/PASP, ΔTRV, Age, and VE/VO₂ slope), converts it to a predicted probability (P), and classifies the patient using an optimal threshold of 0.482 B A true-positive case (Patient 1). The multimodal inputs generated a high predicted probability (P = 0.566), correctly predicting EiPH, which was consistent with the invasive gold standard C A true-negative case (Patient 2). The multimodal inputs generated a low predicted probability (P = 0.324), correctly excluding EiPH, which was consistent with the invasive gold standard. ΔTRV: Change in tricuspid regurgitation velocity from rest to peak exercise; cm/s: Centimeters per second; EiPH: Exercise-induced pulmonary hypertension; Exp: Exponential function; FEV₁: forced expiratory volume in 1 s; FVC: forced vital capacity; Logit(P): Natural logarithm of the odds; mmHg: Millimeters of mercury; P: Predicted probability; PASP: Pulmonary artery systolic pressure; TAPSE: Tricuspid annular plane systolic excursion; VE/VO₂ slope: Minute ventilation versus oxygen uptake slope; VO₂max: maximal oxygen consumption

To evaluate the generalizability and robustness of the multimodal model under different clinical contexts, subgroup analyses were performed based on age (< 60 and ≥ 60 years) and sex (Table 4). The results showed that the model maintained excellent diagnostic performance across all subgroups, with AUCs in the internal validation cohort ranging from 0.830 to 1.000. Notably, no statistically significant interaction was observed between model predictive performance and age or sex (all interaction P values > 0.05 in both the training and internal validation cohorts). In particular, the model still demonstrated robust and excellent discriminative ability in elderly patients (≥ 60 years; internal validation cohort AUC: 0.936) and in females (internal validation cohort AUC: 0.889), confirming its broad clinical applicability.

Table 4.

Subgroup analysis of the multimodal model’s performance for predicting exercise-induced pulmonary hypertension

Dataset	Subgroup	N	AUC (95% CI)	P
Training cohort	< 60 years	27	0.956 (0.852, 1.000)	0.809
	≥ 60 years	35	0.952 (0.867, 1.000)
	Female	21	1.000 (1.000, 1.000)	1.000
	Male	41	0.920 (0.737, 0.979)
Internal validation cohort	< 60 years	8	1.000 (0.737, 1.000)	0.944
	≥ 60 years	8	0.936 (0.737, 1.000)
	Female	9	0.889 (0.650, 1.000)	0.862
	Male	7	0.830 (0.705, 01.000)

AUC Area under the receiver operating characteristic curve, CI Confidence interval

Discussion

This study demonstrates that a noninvasive multimodal model integrating clinical features, ESE, and CPET can effectively identify EiPH. Through feature selection, nine key predictors were ultimately identified, including FEV₁/FVC, VO₂max, VE/VO₂ slope, peak TAPSE/PASP, HRR at 3 min, age, ΔTRV, peak S’, and PRP. These variables encompass pulmonary ventilatory function, exercise metabolic capacity, RV systolic reserve, RV–PA coupling, and age-related factors, suggesting that EiPH is not a single hemodynamic abnormality but rather a comprehensive manifestation of multidimensional impairment in cardiopulmonary reserve. Model performance analysis further showed that although unimodal models already possessed a certain degree of discriminative ability, diagnostic performance improved significantly with the progressive integration of clinical, echocardiographic, and CPET information. The final Clinical+Echo+CPET model maintained a high AUC in both the training and internal validation cohorts and exhibited relatively low Brier scores, indicating that multimodal integration helps improve both the discrimination and calibration of the model.

Previous studies have also explored the noninvasive prediction of EiPH. Among them, Kenya et al. reported an AUC of 0.74 for a fusion model based on clinical features combined with deep learning for the diagnosis of EiPH [15]. In the study by Irene et al., the constructed noninvasive risk score model achieved a maximum sensitivity and specificity of 0.86 and 0.71, respectively; however, this evaluation was performed within the same dataset [10]. In contrast, the integrated model in the present study, which incorporated clinical, ESE, and CPET data, maintained a relatively high AUC in the internal validation cohort and demonstrated relatively stable sensitivity and specificity, suggesting that multidimensional information fusion contributes to improved model generalizability. It is noteworthy that although the multimodal model overall outperformed the unimodal models, the differences between some bimodal and trimodal models did not reach statistical significance, which may be related to the limited sample size and overlapping information across different modalities.

From a pathophysiological perspective, the features selected in this study effectively reflect the core abnormalities of patients with EiPH under exercise stress. Compared with patients without EiPH, those with EiPH exhibited increased peak TRV, whereas peak TAPSE, peak S’, and peak TAPSE/PASP were decreased, indicating that during exercise, pulmonary circulatory afterload is abnormally increased while RV systolic reserve and RV–PA coupling capacity are impaired [16–18]. Compared with peak mPAP alone, the mPAP/CO slope can more comprehensively characterize the ability of the pulmonary vasculature to adapt to increased blood flow across different levels of exercise load; its elevation indicates reduced pulmonary vascular compliance and an abnormal afterload response [19, 20]. In this study, increased ΔTRV and decreased peak TAPSE/PASP noninvasively corresponded to an enhanced pressure response and impaired RV–PA coupling, respectively, which are directionally consistent with the pathological processes reflected by an increased mPAP/CO slope [21, 22].

In addition to right heart mechanical dysfunction, the roles of pulmonary function and exercise metabolic indices in the model are also noteworthy. FEV₁/FVC showed the strongest negative predictive effect, suggesting that airflow limitation or decreased ventilatory function may be involved in abnormal pulmonary circulatory responses during exercise [23, 24]. Patients with EiPH exhibited reduced DLCO and DLCO/VA, and, together with decreased VO₂max and increased VE/VO₂, these findings indicate not only impaired PVR but also reduced gas exchange efficiency and overall exercise oxygen utilization capacity [25, 26]. As one of the most important positive predictors, VE/VO₂ slope suggests an increased ventilatory demand relative to oxygen uptake, which may reflect increased ineffective ventilation during exercise, impaired ventilation/perfusion matching, or reduced efficiency of the pulmonary circulation–gas exchange unit [12, 27]. In addition, reduced HRR at 3 min indicates impaired autonomic regulation and heart rate recovery capacity, while a decrease in PRP, reflecting the achieved workload relative to the predicted workload, reveals limited exercise tolerance and insufficient integrated cardiopulmonary reserve from an overall functional perspective [28]. Therefore, the contribution of CPET indices to the model does not simply reflect reduced exercise capacity but also complements, at a functional level, the cardiopulmonary coupling abnormalities that cannot be fully captured by echocardiographic and clinical variables. This study also suggests that age-related factors are important in the identification of EiPH. Patients with EiPH were older, and reduced exercise tolerance, impaired right heart reserve, and decreased pulmonary vascular compliance may jointly influence the development and identification of EiPH [29, 30]. Notably, subgroup analysis showed that the model maintained good diagnostic performance in patients aged ≥ 60 years.

This study has several limitations. First, this was a single-center study with a relatively limited sample size and without an independent external validation cohort; therefore, the generalizability of the model needs to be further validated in multicenter populations. Second, this study used a synchronized testing protocol integrating ESE, CPET, and eRHC. Although this design reduced bias related to nonsynchronous measurements, it requires specialized equipment, operator expertise, and multidisciplinary coordination, which may limit its implementation in some centers. Third, echocardiographic parameters obtained during exercise may be affected by image quality, respiratory motion, increased heart rate, and doppler alignment, and therefore some measurement variability may remain. Finally, this was a cross-sectional analysis without long-term follow-up data; thus, the association between model-predicted EiPH and future progression to resting PH or adverse clinical outcomes could not be assessed. Future studies with larger, multicenter prospective cohorts and long-term prognostic data are needed to further validate the clinical value of this model.

Conclusion

This study showed that a noninvasive multimodal model integrating clinical variables, ESE, and CPET demonstrated good discriminative performance for EiPH. Comprehensive noninvasive assessment may help improve the early recognition of exercise-induced pulmonary vascular abnormalities, but further validation in larger multicenter cohorts is needed.

Abbreviations

AUC

Area under the curve

CI

Cardiac index

CO

Cardiac output

CPET

Cardiopulmonary exercise testing

CTEPH

Chronic thromboembolic pulmonary hypertension

Echo

Echocardiography

EiPH

Exercise-induced pulmonary hypertension

E

Early diastolic peak transmitral flow velocity

ESE

Exercise stress echocardiography

e’

Early diastolic mitral annular velocity

FAC

Fractional area change

HR

Heart rate

HRR 3min

3-minute heart rate recovery

LASSO

Least absolute shrinkage and selection operator

mPAP

Mean pulmonary arterial pressure

PASP

Pulmonary arterial systolic pressure

PH

Pulmonary hypertension

PVR

Pulmonary vascular resistance

RAP

Right atrial pressure

RAScd

Right atrial conduit strain

RASct

Right atrial contraction strain

RASr

Right atrial reservoir strain

eRHC

exercise right heart catheterization

RV

Right ventricle

RVEDA

Right ventricular end-diastolic area

RVESA

Right ventricular end-systolic area

RVFWS

Right ventricular free-wall strain

RV–PA

Right ventricle–pulmonary artery

SD

Standard deviation

TAPSE

Tricuspid annular plane systolic excursion

TPR

Total pulmonary resistance

TRV

Tricuspid regurgitation velocity

VE/VCO₂ slope

Minute ventilation versus carbon dioxide output slope

VE/VO₂ slope

Minute ventilation versus oxygen uptake slope

VO₂max

Peak oxygen uptake

WHO

World Health Organization

Authors’ contributions

(I) Conception and design: XPD, YDL. (II) Administrative support: YDL, XPD. (III) Provision of study materials or patients: JNG, SQY, YHY, QMX. (IV) Collection and assembly of data: XPD, RF, JWZ. (V) Data analysis and interpretation: XPD, YDL, JYH, RF, DCG, RF. (VI) Manuscript writing: All authors. (VII) Final approval of manuscript: All authors.

Funding

Supported by Beijing Research Ward Excellence Program (No.BRWEP2024W112030103);the National Natural Science Foundation of China (No.82572244)༛Beijing Municipal Administration of Hospitals Incubating Program (No. Z221100007422028).

Data availability

All data used and analyzed during this study are included in the manuscript.

Declarations

Ethics approval and consent to participate

This prospective study was approved by the Institutional Ethics Committee of Beijing Chaoyang Hospital, and all patients signed informed consent forms (approval number 2023-KE-264). The study was conducted in strict accordance with the ethical standards of the institutional research committee, the 1964 Declaration of Helsinki and its later amendments, or comparable ethical standards.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.