Network Analysis to Risk Stratify Patients with Exercise Intolerance

Abstract

Rationale

Current methods assessing clinical risk due to exercise intolerance in cardiopulmonary disease patients rely on a small subset of traditional variables. Alternative strategies incorporating the spectrum of factors underlying prognosis in at-risk patients may be useful clinically, but are lacking.

Objective

Use unbiased analyses to identify variables that correspond to clinical risk in patients with exercise intolerance.

Methods and Results

Data from 738 consecutive patients referred for invasive cardiopulmonary exercise testing (iCPET) at a single center (2011–2015) were analyzed retrospectively (derivation cohort). A correlation network of iCPET parameters was assembled using |r|>0.5. From an exercise network of 39 variables (i.e., nodes) and 98 correlations (i.e., edges) corresponding to P<9.5e⁻⁴⁶ for each correlation, we focused on a subnetwork containing peak rate of oxygen consumption (pVO₂) and 9 linked nodes. K-mean clustering based on these ten variables identified 4 novel patient clusters characterized by significant differences in 44 of 45 exercise measurements (P<0.01). Compared to a probabilistic model including 23 independent predictors of pVO₂ and pVO₂ itself, the network model was less redundant and identified clusters that were more distinct. Cluster assignment from the network model was predictive of subsequent clinical events. For example, a 4.3-fold (P<0.0001; 95% CI, 2.2–8.1) and 2.8-fold (P=0.0018; 95% CI, 1.5–5.2) increase in hazard for age- and pVO₂-adjusted all-cause 3-year hospitalization, respectively, were observed between the highest vs. lowest risk clusters. Using these data, we developed the first risk-stratification calculator for patients with exercise intolerance. When applying the risk calculator to patients in 2 independent iCPET cohorts (Boston, USA and Graz, Austria), we observed a clinical risk profile that paralleled the derivation cohort.

Conclusions

Network analyses were used to identify novel exercise groups and develop a point-of-care risk calculator. These data expand the range of useful clinical variables beyond pVO₂ that predict hospitalization in patients with exercise intolerance.

INTRODUCTION

Exercise intolerance is highly prevalent across a wide range of diseases encountered commonly in routine clinical practice, and is a principal cause of morbidity and increased healthcare cost burden.^1,2 The pathophysiology of exercise intolerance is generally ascribed to a cardiovascular, pulmonary, or skeletal muscle abnormality that impairs oxygen (O₂) delivery to or extraction by peripheral tissue.³ However, abnormalities in multiple organ systems are frequently observed in patients referred for evaluation, complicating efforts to establish the parameters that delineate different forms of exercise intolerance as well as the development of patient-specific risk stratification metrics.⁴ This dilemma is due, in part, to conventional methods that utilize only a narrow subset of available clinical data for analyzing exercise performance.³ As a result, peak volume of oxygen consumption (pVO₂) is often used as the single exercise variable for determining prognosis in patients with cardiopulmonary diseases.⁴ Interpreting exercise data using a wider range of clinical variables may have important implications for understanding exercise subtypes and clarifying patient prognosis, but such methods are not currently available.

Invasive cardiopulmonary exercise testing (iCPET) involves the simultaneous measurement of pulmonary function, ventilatory response, O₂ transport and utilization, cardiac performance, and hemodynamic data at rest and throughout exercise. Thus, iCPET has emerged as the gold standard diagnostic tool to unmask various abnormalities during physical activity, including left atrial hypertension, pulmonary vascular dysfunction, impaired venous return, and mitochondrial disease.^5–7 However, these diagnoses are based on abnormal findings utilizing a limited number of parameters, generally focusing on a single iCPET element (e.g., elevated pulmonary artery wedge pressure with exercise) and do not consider groups of variables that may characterize exercise limitation.^3,8 Moreover, many patients meet criteria for multiple exercise disorders while other symptomatic individuals fail to meet criteria for any disorder,⁹ suggesting limited discriminatory ability of conventional methods that use binary classification or branching logic for iCPET interpretation. As a result of these limitations, the use of iCPET to stratify at-risk patients for hard clinical end points has not been reported.

We hypothesized that a methodological approach to analyzing exercise data that emphasizes novel (and unexpected) relationships between iCPET variables could clarify different forms of exercise intolerance and allow for a systematic approach to assessing prognosis of at-risk patients in clinical practice.

METHODS

An expanded Methods section is located in the Supplemental Material.

Data, methods used in the analysis, and materials used to conduct the research will be made available to a qualified researcher for the purposes of reproducing the results or replicating the methods. Because of the sensitive nature of the data collected for this study, requests to access the dataset from qualified researchers trained in human subject confidentiality protocols may be sent to the corresponding author of this manuscript at Brigham and Women’s Hospital.

Study population

We analyzed retrospectively clinical and demographic data from consecutive patients referred to the Brigham and Women’s Hospital (BWH) Dyspnea Center (Boston, MA) between March, 2011, and December, 2015, who underwent iCPET as a part of clinical assessment for unexplained dyspnea and/or exertional intolerance.⁸ The study protocol was approved by the Partners Human Research Committee (IRB# 2011P000272). The requirement for informed consent was waived for all patients. The initial study group included 832 exercise tests. Studies missing data on peak cardiac output (N=48) or ≥ 10 other variables (N=27) were excluded from analysis. There were 19 patients for whom data from two iCPET studies were available, and in these cases the first iCPET was included in the analysis. Overall, there were 738 iCPET studies from 738 patients used for analysis. Further details on the completeness of data and exercise testing methods are provided in the Supplemental Material.

Network analysis methods

The overall approach to analysis of the iCPET dataset is illustrated in Figure 1, and a list of abbreviations and definitions of exercise variables is given in Online Table I. An overarching objective of this analysis was to first capture unexpected and functionally important relationships between exercise variables, which would then be used to identify distinct patient groups. However, many iCPET variables are interrelated (e.g., pVO₂ in mL/kg/min is closely related to pVO₂ % predicted⁴) and could be expected to correlate highly with each other (i.e., be collinear). To address this issue, we grouped iCPET variables according to their functional similarity. Specifically, variables were categorized into one of the following iCPET functional groups: pulmonary function, exercise capacity, ventilatory response to exercise, O₂ transport and utilization, non-invasive cardiac performance, invasive cardiac performance, and systemic and cardiopulmonary hemodynamics (Online Table II). Next, we assembled a network based on Pearson correlation coefficients between pairwise interactions,¹⁰ but excluded correlations among variables within the same functional group to avoid including expected associations between variables. We selected a correlation coefficient threshold of |r|>0.5 based on methods outlined in the Supplemental Material. At this correlation threshold, the threshold for the P-value for correlations was P<9.5e⁻⁴⁶ after Bonferroni correction for the number of all possible edges among the 73 iCPET variables. The network was refined further as described in the Supplemental Material. Network visualization was performed using Cytoscape v3.5.1.¹¹

Figure 1. Overall study design and approach to developing the invasive cardiopulmonary exercise testing (iCPET) networks.

(A) Flow diagram of the overall study design. Developing the iCPET network was performed to identify novel and functionally related associations between iCPET variables, which then could be used to characterize subpopulations of patients with exercise intolerance. (B) Approach to developing the iCPET network in the derivation cohort. CO, cardiac output, %p, percent predicted.

Clinical end-points

Mortality was determined by events recorded in the U.S. National Social Security Death Index. Hospitalization was determined by the identification of a patient-linked BWH hospital discharge recorded in the electronic medical record at BWH. Time-to-event data for the outcome measures of all-cause mortality or hospitalization were modeled using a Cox proportional hazards model in which outcome time 0 was the date of the iCPET. In the event that multiple iCPET studies were performed for a single patient (N=19), the first iCPET was used as the index study. Observations without a documented event were censored at the time of database review. Multivariable modeling was performed adjusting for age, sex, and percent predicted pVO₂. Patients for whom outcome data were unavailable were not included in any end-point analyses.

Generating the iCPET risk calculator

For a patient undergoing iCPET, the 10 variables from the subnetwork (V_1–10) (see results section for a detailed description of the network analysis) are normalized (N) using the mean and standard deviation calculated for each variable from our iCPET cohort (N=738). These normalized values, N₁–N₁₀, compose a vector describing a patient’s exercise performance. Next, the Euclidean distance (d) from the patient’s vector of normalized variables to the center of each of the patient clusters is calculated. The patient is assigned to the cluster with the shortest vector-cluster center distance (d₁). The clinical risk (3-yr hospitalization rate) for the patient is based on outcome data for his/her assigned cluster as determined by our outcome analysis (Online Figure I).

Statistical analyses

Normality was assessed using the Shapiro-Wilk test. For continuous variables, data are presented as mean ± SD if normally distributed and median [IQR] if non-normally distributed, unless otherwise specified. Comparisons of proportions between groups for categorical variables were performed using the Fisher’s exact test and Chi-Square test for smaller and larger sample sizes, respectively. Comparisons involving continuous variables between two groups were performed using the Student’s t-test for normally distributed variables and the Mann-Whitney U-test for non-normally distributed variables. One-way ANOVA was used when comparing differences between more than two groups. Two-way ANOVA was used to compare the frequency of hospitalization by patient cluster between the derivation and Graz, Austria validation cohorts.

Patient clusters were determined by K-means analysis as described in detail in the Supplemental Material. The area under the curve (AUC) for the receiver operating characteristic (ROC) analysis was calculated using the method of DeLong and colleagues, as reported previously.¹² A normalized mutual information (NMI) value was used to calculate the overlap between patients by cluster vs. pVO₂ quartile, as reported previously.¹⁰

Analyses were performed using SPSS software, version 19 (IBM Company, Armonk, NY), Origin 9.1 (Northampton, MA), SAS 9.3 (SAS Institute, Cary, NC), and GraphPad Prism v7.03 (GraphPad Software, La Jolla, CA). Two-dimensional and three-dimensional visualization of the PC analyses was performed using the R packages ‘factoextra’ and ‘pca3d, respectively.

RESULTS

Study population

We identified 738 patients (58 [46–69] yr, 36% female) undergoing 738 iCPET studies during the study period (derivation cohort). The primary clinical indication for iCPET was unexplained exertional intolerance, which encompasses dyspnea on exertion, fatigue, and/or lightheadedness. In these patients, standard clinical testing performed at rest produced results that were either (i) normal or (ii) abnormal but insufficient to explain exertional intolerance as determined by the referring clinician. The criteria for traditional exercise dysfunction diagnoses and baseline clinical and exercise data for the study population are provided in Online Table III and IV. Systemic hypertension, hyperlipidemia, and diabetes mellitus were present in 43.5%, 36.6%, and 13.0% of patients, respectively. The average left ventricular ejection fraction (61 [57–68] %), resting right atrial pressure (RAP) (3 [1–6] mmHg), mean pulmonary artery pressure (mPAP) (15 [12–19] mmHg), pulmonary vascular resistance (PVR) (1.7 [1.2–2.4] WU), and pulmonary artery wedge pressure (PAWP) (7 [4–10] mmHg) were within the normal range, and the most common post-iCPET diagnoses were presumed normal (18.8%), left heart disease without pulmonary vascular disease (LHD-noPVD) (18.6%), and peripheral O₂ extraction disorder (in our practice this is due most commonly to dysautonomia or a mitochondrial myopathy) (18.6%). An associated pulmonary mechanical limit to exercise was observed in 33.3% of patients.

The exercise network

The exercise network contained 39 variables (i.e., nodes) and 98 correlations (i.e., edges) (Figure 2A). The most highly connected nodes (≥7 edges) in the exercise network are reported in Online Table V, and included pVO₂. Owing to its centrality in the exercise network and relevance to clinical outcome across the spectrum of medical diseases associated with dyspnea,¹³ we next focused on correlations involving pVO₂.

Figure 2. Exercise network and subnetwork used to select variables for determining exercise pathophenotypes.

(A) Data from 738 consecutive patients referred for invasive cardiopulmonary exercise testing (iCPET) were analyzed. Pairwise correlations (|r| > 0.5; this |r| threshold corresponded to P<9.5e⁻⁴⁶ for each correlation) for 73 iCPET variables were used to identify unanticipated relationships between exercise measurements (i.e., the exercise network) after grouping variables by iCPET function (box). A complete list of variable abbreviations and iCPET functional groups is provided in Online Tables I and II, respectively. The exercise network contained 39 nodes and 98 edges, but was too dense for further interpretation. Therefore, a subnetwork was constructed based on relationships between variables and pVO₂, which was a highly connected node in the larger exercise network indicating that this was a potentially important variable. (B) This subnetwork was used for further analyses to determine if specific exercise pathophenotypes could be identified within the iCPET cohort. pVO₂, volume of oxygen consumption at peak exercise; pSV, stroke volume at peak exercise; FVC, forced vital capacity in liters; MVV, maximal voluntary ventilation; pCa-vO₂, arterial to mixed-venous O₂ content difference at peak exercise; pVE, minute ventilation at peak exercise; FEV-1, forced expiratory volume-1 second in liters; pLactate, arterial lactate level at peak exercise; ppH; arterial pH at peak exercise.

We identified an exercise subnetwork with pVO₂ as a central node that contained 9 other variables (Figure 2B). Overall, this 10-variable subnetwork contained measures from 5 different iCPET functional groups: pulmonary function, exercise capacity, O₂ transport, invasive cardiac performance, and ventilatory response to exercise. The 10 variables were: pVO₂, maximal voluntary ventilation (MVV), peak minute ventilation (VE), peak arterial to mixed-venous O₂ content difference (pCa-vO₂), forced vital capacity (FVC), forced expiratory volume in 1^st second (FEV-1), peak arterial pH, peak stroke volume, peak arterial lactate, and peak arterial oxygen content (CaO₂). This subnetwork was used for further study based on the diversity of iCPET functional groups represented in it and a size amenable to additional analyses.

Identifying distinct patient groups from the exercise subnetwork

Using variables from the subnetwork, K-means analysis (Supplemental Material) identified 4 distinct patient clusters (Figure 3A, Online Figure IIA-F, Online Tables VI and VII). Importantly, all variables in the subnetwork were used to determine cluster assignment with pVO₂, peak CaO₂, and peak arterial lactate contributing to assignment in 100% of cases, FVC in 98% of cases, and peak stroke volume in 58% of cases (Online Figure IIG).

Figure 3. Normalized value for each variable in the subnetwork stratified by cluster.

Data from 738 patients referred for invasive cardiopulmonary exercise testing (iCPET) were used to classify patients into clusters based on their performance on the 10 variables from the exercise subnetwork. (A) The distribution of patients in each cluster is plotted according to variance for iCPET data from all patients in the study cohort (principal component [PC] 1 and PC 2). (B) The normalized value of each variable in the exercise subnetwork is presented as a function of cluster assignment. The line between variables is used to visually distinguish data belonging to each particular cluster. (C) Heat map illustrating the individual patient level differences by cluster for each subnetwork variable. Scale indicates the range of normalized fold-change in performance. (D) The prevalence of patients meeting standard exercise diagnoses in each iCPET cluster. pVO₂, volume of oxygen consumption at peak exercise; pSV, stroke volume at peak exercise; FVC, forced vital capacity in liters; MVV, maximal voluntary ventilation; pCa-vO₂, arterial to mixed-venous O₂ content difference at peak exercise; pVE, minute ventilation at peak exercise; FEV-1, forced expiratory volume-1 second in liters; pLactate, arterial lactate level at peak exercise; ppH; arterial pH at peak exercise; PVD, pulmonary vascular disease; LHD-noPVD, left heart disease without PVD; LHD+PVD, left heart disease with PVD.

Each subnetwork variable for all patients was normalized to have a mean = 0 and variance = 1, with the cluster centers shown in Figure 3B. This analysis revealed a gradient across clusters for many variables. Specifically, the lowest normalized values for 9 of the 10 subnetwork variables were observed in cluster 4, with incremental increases in the normalized value for these 9 variables observed in cluster 3, cluster 2, and then cluster 1. For arterial pH at peak exercise, the trend was directionally opposite and the magnitude in difference across the clusters was less compared to the other 9 variables. A heat map of normalized values for variables in the subnetwork shows homogeneity within clusters, but strong differences between clusters (Figure 3C).

Comparing clusters from the subnetwork and multiple linear regression models

We performed probabilistic modeling as a comparison to our network-based approach for patient clustering. From a total of 72 iCPET variables, multiple linear regression analysis identified 24 variables as independent predictors of pVO₂ (Online Table VIII). This approach was also performed after adjusting for collinear variables (Online Table IX) (see Supplemental Material). Compared to the unadjusted model, the adjusted model (N=23 variables) did not include age or resting venous content of oxygen (CvO₂), but added peak Ca-vO₂. Otherwise, no differences were observed between the unadjusted and adjusted multiple linear regression results.

As with the network model, K-means analysis identified 4 distinct patient clusters using either the unadjusted or adjusted multiple linear regression analysis. The PC1 and PC2 as well as the heat map from variables in both multiple linear regression analyses are provided in Online Figure III. There were 5 variables in both the network and adjusted multiple linear regression models (peak VE, peak stroke volume, peak arterial lactate, pCa-vO₂, and pVO₂). Redundancy was noted among independent predictors in the multiple linear regression model, as 10 variables were derived from the same measurement (e.g., pVO₂ adjusted for weight [pVO₂_Cal] and pVO₂ percent predicted [pVO₂%]), and numerous variables are functionally similar and known to be strongly correlated (e.g., pVO₂ percent predicted and exercise watts).

The Davies-Boudin index, a function of the ratio of within cluster scatter to between cluster separation where lower values indicate better clustering, was 1.93, 2.02, and 2.16 for the subnetwork, unadjusted multiple linear regression, and adjusted multiple linear regression approaches, respectively. This finding indicates that greatest separation (and ‘tightness’) of the patient clusters, and, thus, greater discrimination, was achieved by the subnetwork approach as compared to the multiple linear regression approaches.

Clinical profile and relevance of the subnetwork-based clusters

The clinical profile and values for each of the subnetwork variables is presented in Table 1, and is expanded in Online Table X to include all iCPET exercise variables. The mean values of most subnetwork variables improved across clusters: 4è3è2è1, which included pVO₂, MVV, FEV-1, FVC, peak Ca-vO₂, peak stroke volume, and peak arterial lactate (P<0.0001 for all comparisons) (Online Table X). A similar trend was also observed for iCPET measurements not included in the subnetwork, such as peak CO, peak RAP, peak PVR, and peak PAWP (P<0.0001 for all comparisons). Among the 45 exercise and clinical variables we analyzed for each patient, there were significant differences in 44 (97%) across the 4 patient clusters (Online Table X). The greatest difference between most clinical, hemodynamic, and pulmonary function variables was observed when comparing cluster 4 to cluster 1, which was also the case for age (70 [61–76] vs. 50 [35–60] yr) and hemoglobin (12.5 [11.5–13.4] vs. 15.5 [14.6–16.3] g/dL).

Table 1. The clinical and exercise subnetwork profile of patients by cluster.

Data from 738 patients referred for invasive cardiopulmonary exercise testing (iCPET) were used to classify patients into groups (clusters) based on the 10 exercise variables in the exercise subnetwork. An ‘r’ prior to a variable indicates a measurement at rest and a ‘p’ prior to a variable indicates a measurement at peak exercise. Data for continuous variables that are normally distributed are presented as mean ± SD and non-normally distributed variables as median [IQR]. Data for categorical variables are presented as N (%). The P-values were calculated as described in the Methods. BMI, body mass index; LVEF, left ventricular ejection fraction; CHD, coronary heart disease; ACE, angiotensin converting enzyme; pVO₂, volume of oxygen consumption at peak exercise; pVE, minute ventilation at peak exercise; FEV1, forced expiratory volume in 1 second in liters; FVC, forced vital capacity in liters; MVV, maximal voluntary ventilation; ppH; arterial pH at peak exercise; pLactate, arterial lactate level at peak exercise; pCaO₂, arterial O₂ content at peak exercise; pCa-vO₂, arterial to mixed-venous O₂ content difference at peak exercise; pSV, stroke volume at peak exercise; *LVEF data were available for N=432 patients.

Clinical Characteristic	Cluster 4(N=205)	Cluster 3(N=260)	Cluster 2(N=173)	Cluster 1(N=100)	P Value
Age (yr)	70 [61–76]	58 [49–68]	50 [36–62]	48 [35–60]	<0.0001
Female (n, %)	159 (78)	195 (75)	105 (61)	14 (14)	<0.0001
Weight (kg)	80 [68–95]	73 [64–91]	77 [64–95]	92 [78–109]	<0.0001
BMI (kg/m2)	29.3 [24.3–34.4]	26.8 [23.1–31.6]	26.0 [23.0–31.5]	27.4 [24.4–31.1]	0.001
Age (yr)	70 [61–76]	58 [49–68]	50 [36–62]	48 [35–60]	<0.0001
LVEF* (%)	61.2 [56.2–66.3]	62.8 [56.6–68.8]	62.9 [56.3–68.9]	61.2 [57.5–65.6]	0.39
Co-morbidities
Systemic hypertension	132 (64)	100 (38)	54 (31)	28 (28)	<0.0001
Hyperlipidemia	103 (50)	90 (35)	45 (26)	24 (24)	<0.0001
Diabetes mellitus	52 (25)	25 (10)	13 (8)	4 (4)	<0.0001
≥1 CHD risk factor	36 (18)	47 (18)	25 (14)	15 (15)	0.7429
Valvular disease	27 (13)	16 (6)	5 (3)	3 (3)	<0.0001
History of tobacco use	5 (2)	6 (2)	5 (3)	3 (3)	0.8485
Coronary artery disease	32 (16)	23 (9)	16 (9)	3 (3)	0.0039
Medication Use
Digoxin	10 (5)	5 (2)	1 (1)	0	0.0135
β-adrenergic receptor antagonist	80 (39)	69 (27)	45 (26)	10 (10)	<0.0001
Calcium channel antagonist	54 (26)	37 (14)	11 (6)	6 (6)	<0.0001
ACE Inhibitor	39 (19)	42 (16)	20 (12)	13 (13)	0.2131
Diuretic	90 (44)	60 (23)	28 (16)	13 (13)	<0.0001
Aspirin	82 (40)	73 (28)	39 (23)	19 (19)	<0.0001
Insulin	17 (8)	8 (3)	4 (2)	3 (3)	0.0229
Oral hypoglycemic	35 (17)	14 (5)	7 (4)	4 (4)	<0.0001
Exercise subnetwork variables
pVO2 (mL/kg/min)	10.5 [8.8–12.5]	15.1 [12.5–17.9]	19.7 [16.6–24.1]	24.8 [19.4–31.4]	<0.0001
pVE (L)	34 [28–41]	45 [39–54]	61 [55–70]	87 [75–99]	<0.0001
FEV-1 (% predicted)	68 ± 21	85 ± 19	92 ± 18	97 ± 16	<0.0001
FVC (% predicted)	67 ± 19	87 ± 17	93 ± 16	98 ± 16	<0.0001
MVV (L)	57 [44–69]	85 [71–96]	104 [92–117]	132 [117–156]	<0.0001
ppH	7.40 [7.38–7.45]	7.39 [7.36–7.43]	7.37 [7.34–7.39]	7.36 [7.32–7.39]	<0.0001
pLactate (mg/dL)	3.4 [2.4–4.5]	4.9 [3.9–5.9]	6.6 [5.3–7.8]	7.0 [5.3–9.2]	<0.0001
pCaO2 (mL/dL)	16.2 ± 1.9	18.5 ± 1.8	19.2 ± 1.7	21.1 ± 2.0	<0.0001
pCa-vO2 (mL/dL)	10.1 [8.8–11.0]	11.5 [10.2–13.0]	12.0 [11.0–13.5]	14.2 [12.2–15.7]	<0.0001
pSV (mL)	76.7 [64.1–94.2]	76.5 [64.0–88.9]	86.2[75.1–107.2]	110.5[92.7–129.3]	<0.0001

We next determined if cluster assignment was associated with differences in clinical measures of exercise tolerance. We observed a step-wise increase in maximum watts achieved during iCPET across clusters 4è3è2è1 (55 [34–65] vs. 87 [71–102] vs. 126 [107–141] vs. 176 [149–210] watts, P<0.0001), which was consistent with findings that patients in cluster 4 and cluster 1 generally had the most and least abnormal clinical, pulmonary function, metabolic, and cardiopulmonary hemodynamic profiles, respectively (Online Table X). Owing to the strong correlation between watts achieved and pVO₂,^1–3 which was a variable in the subnetwork, we next assessed if pVO₂ was a critical determinant of the observed differences in functional status by cluster. Approximately half of the patients in cluster 4 (52%), the cluster with the lowest average pVO₂, were not, however, also classified in the lowest pVO₂ % predicted quartile. Overall agreement for patient cluster assignment and pVO₂ quartile assignment was low (NMI [0,1]=0.0725) (Online Figure IV and Online Table XI). Further, the accuracy rates of pVO₂ for predicting cluster assignment for a patient were 66.4% and 66.3% by linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA), respectively, whereas the 10 variables in the subnetwork predicted the cluster assignment for a patient with accuracy rates of 94.1% and 92.6% for LDA and QDA, respectively. Taken together, these findings suggest that cluster assignment was not predominately a consequence of pVO₂.

Prevalence of standard exercise diagnoses by subnetwork-based cluster

The prevalence of patients according to standard exercise diagnoses differed across clusters for PVD (P=0.0003), LHD with PVD (LHD+PVD) (P<0.0001), peripheral O₂ extraction disorder (P<0.0001), low ventricular filling syndrome (i.e., failure to augment venous return as the primary identifiable cause of impaired cardiac output)⁵ (P=0.0252), and presumed normal status (P<0.0001), but not LHD-noPVD (P=0.16) (Table 2 and Figure 3D). The criteria for a single specific exercise diagnosis were not met in 136 patients (17.9%), and the prevalence of this patient subgroup (i.e., alternative diagnosis) across all clusters was also significantly different (P=0.0018) with the greatest difference identified between clusters 3 and 4 (23% vs. 10%, P<0.0001). The highest proportion of patients with PVD (16.2%), LHD+PVD (19.5%), LHD-noPVD (22.4%), and peripheral O₂ extraction disorder (27.2%) were in cluster 4. By contrast, the prevalence of patients with presumed normal cardiopulmonary hemodynamic responses to exercise was lowest in cluster 4 and, overall, the distribution of this subgroup correlated inversely with clinical event rates by cluster ([cluster 4] 5.2 vs. [cluster 3] 16.4 vs. [cluster 2] 29.2 vs. [cluster 1] 31.1 % of patients).

Table 2. The prevalence of patients according to standard exercise diagnoses for the derivation (BWH) and validation (Graz, Austria) cohorts.

Patients may have met criteria for more than a single diagnosis. BWH, Brigham and Women’s Hospital; PVD, pulmonary vascular disease; LHD, left heart disease; Periph, peripheral; Pulm, pulmonary. The hemodynamic criteria for each definition are provided in the Supplement. Data are expressed as N (%).

Cohort	PVD	LHD+PVD	LHD-noPVD	Periph. O2 Extraction Disorder	Impaired Ventricular Filling Syndrome	Presumed Normal	Pulm Mechanical Limit	Alternative Diagnosis
BWH	66 (8.9)	63 (8.5)	137 (18.6)	137 (18.6)	84 (11.4)	139 (18.8)	226 (30.6)	136 (18.4)
Graz, Au	13 (11.1)	12 (10.3)	51 (43.6)	25 (21.4)	3 (2.6)	5 (4.3)	38 (32.5)	21 (17.9)

Patient clusters and clinical outcome

Outcome data were available for 700 patients (95%). During the study period, death was reported in 27 patients (6.8%) and the all-cause mortality for clusters 4, 3, 2, and 1 was 8.6%, 1.9%, 2.2%, and 0%, respectively (P<10⁻⁴). Inpatient hospitalization occurred in 194 patients (25.6%) at a median of 556 [214–926] days following iCPET. The median follow-up for non-hospitalized subjects for clusters 4, 3, 2, and 1 were 619, 632, 756, and 862 days, respectively. Hospitalized patients were older at the time of iCPET (63 [50–71] vs. 57 [44–68] yr, P<0.001), more likely to be female (45% vs. 35%, P<0.05), and had more severe abnormalities in variables in the subnetwork compared to non-hospitalized counterparts (Online Table XII). The Kaplan-Meier estimates for 3-year all-cause hospitalization rates for clusters 4, 3, 2, and 1 was 47%, 33%, 22%, and 16%, respectively (P<0.0001). A Kaplan-Meier plot for all-cause hospitalization by cluster is presented in Figure 4A. Compared to cluster 1, which served as the referent group, cluster 4 had the highest hazard ratio for all-cause hospitalization (HR=4.54; 95% CI, 2.47–8.33, P<0.001), followed by cluster 3 (HR=2.56; 95% CI, 1.39–4.76, P=0.0028) and cluster 2 (HR=1.71; 95% CI, 0.88–3.33, P=0.012). Directionally similar findings were observed after adjusting for age, pRAP, pmPAP, pPVR, or pVO₂ (Table 3).

Figure 4. Clinical outcome by subnetwork cluster and exercise risk calculator.

Data from 738 patients referred for invasive cardiopulmonary exercise testing (iCPET) were used to classify patients into four clusters based on their performance on the 10 variables from the exercise subnetwork, including peak volume of oxygen consumption (pVO₂). (A) The study cohort was stratified by patient cluster and Kaplan-Meier analysis of the probability of all-cause hospitalization was performed. (B) The iCPET subnetwork, cluster, and outcome data were used to generate an on-line risk stratification calculator. This clinical tool automates patient cluster assignment from performance results on each of the 10 variables in the subnetwork, and reports prognosis based on the corresponding cluster-derived 3-year all-cause hospitalization rate (95% confidence interval) (accessible at: https://icpet.partners.org/). (C) The risk calculator was applied to a validation cohort of 113 iCPET patients referred for exercise intolerance to Medical University in Graz, Austria. (D) These results were similar to findings from the derivation cohort studying patients with exercise intolerance referred to Brigham and Women’s Hospital. The difference in hospitalization rate by cluster between cohorts was not statistically significantly different by 2-way ANOVA (P=0.28).

Table 3. All-cause hospitalization rate for patients according to iCPET cluster.

The unadjusted and adjusted hazard ratio for all-cause hospitalization among patients assigned to clusters 3, 2, or 1 compared to cluster 4. pVO₂%, peak volume of oxygen consumption expressed as a percent of predicted; pRAP, peak right atrial pressure; pmPAP, peak mean pulmonary artery pressure; pPVR, peak pulmonary vascular resistance.

Cluster			All-Cause Hospitalization
	Unadjusted	P	Adjustment
			Age	P	pRAP	P	pmPAP	P	pPVR	P	pVO2%	P
1	1.0(referent)	-	1.0(referent)		1.0(referent)		1.0(referent)		1.0(referent)		1.0(referent)	-
2	1.71(0.88–3.33)	0.116	1.70(0.87–3.31)	0.1192	1.65(0.85–3.21)	0.143	1.77(0.91–3.5)	0.0914	1.68	0.129	1.56(0.80–3.04)	0.1932
3	2.56(1.39–4.76)	0.0028	2.48(1.33–4.64)	0.0044	2.50(1.35–4.62)	0.0037	2.65(1.4–4.9)	0.0019	2.44	0.005	1.90(1.01–3.6)	0.0437
4	4.54(2.47–8.33)	<0.001	4.25(2.23–8.07)	<0.0001	4.10(2.19–7.48)	<0.0001	4.00(2.17–7.38)	<0.001	3.95	<0.0001	2.76(1.47–5.21)	0.0018

We used ROC analysis to compare the predictive value of the 10-variable subnetwork model with individual variables reported previously to increase clinical risk in referral populations. The AUC for pVO₂ quartile alone was increased significantly by adding the 10-variable subnetwork model when predicting 3-year mortality (0.774 vs. 0.816, P=0.026) and 3-year hospitalization (0.636 vs. 0.667, p=0.024). By contrast, peak exercise PAWP was a weak predictor of both 3-year mortality (AUC=0.608, P=0.04) and 3-year hospitalization (AUC=0.6036, P=0.06). There was, however, an incremental improvement in predicting 3-year hospitalization when adding the 10-variable model to PAWP (AUC +0.036, p=0.005).

Although iCPET provides the most comprehensive assessment of exercise physiology, its availability in clinical practice remains limited compared to conventional (non-invasive) CPET. We performed additional ROC analyses to determine if iCPET provided an incremental difference in predictive power compared to CPET. From the 10 variables in the exercise subnetwork, a model including just the 5 non-invasive variables (FVC, FEV-1, VE, MVV, and pVO₂) was predictive of 3-year hospitalization, but adding the additional 5 invasively obtained variables improved the AUC significantly (0.6833 vs. 0.7093, P=0.026). This analysis was not performed for mortality given the small absolute number of events relative to predictors.

iCPET clinical risk calculator and validation cohorts

The iCPET subnetwork, cluster, and outcome data were used to generate an on-line risk calculator. This clinical tool automates patient cluster assignment from performance results on each of the 10 variables in the subnetwork, and reports prognosis based on the corresponding cluster-derived Kaplan-Meier estimate for 3-year all-cause hospitalization. The calculator expresses the estimated risk (with 95% confidence interval) corresponding to each cluster (Figure 4B; accessible at: https://icpet.partners.org/).

Second cohort at BWH, Boston, MA, USA

To investigate the reproducibility of our network findings, we applied the risk calculator to an independent iCPET cohort from BWH. We analyzed cardiopulmonary hospitalization events in 78 iCPET patients at BWH not included in the primary study cohort and in whom at least 1-year of follow-up data were available (median follow-up was 765 [422–1316] days) (Online Table XIII). Inpatient hospitalization due to cardiopulmonary disease was observed in 10.2% of patients. The 1-year cardiopulmonary disease-specific hospitalization rate was 24%, 5%, 5%, and 0% for Clusters 4, 3, 2, and 1 (P<0.031 for cluster 4 vs. 1) (Online Figure V).

Validation cohort at Graz, Austria

To focus on the validity and generalizability of our risk calculator, we next studied a patient population referred for iCPET at the Medical University in Graz, Austria (see Supplemental Material for details on this study population). From a cohort of 117 consecutive patients referred for iCPET (53.8 ± 12.9 yr, 79% female), data were available for analysis on 113 patients. The prevalence of patients according to standard exercise diagnoses is provided in Table 2. The risk calculator was used to assign patients to cluster 1 (N=10), cluster 2 (N=48), cluster 3 (N=35), and cluster 4 (N=20). Patients in cluster 1 were significantly younger compared to patients in the other clusters; no significant difference in age was observed between clusters 2, 3, and 4 (Online Figure VI). This finding contrasted with the age profile by cluster observed in the derivation cohort, and suggested that differences in exercise performance in this validation cohort were unlikely to be due solely to age. Data for each network variable and BMI by cluster are provided in Table 4.

Table 4. Clinical, pulmonary function, metabolic, and cardiopulmonary hemodynamic characteristics of patients from the Graz, Austria validation cohort by iCPET cluster.

The risk calculator was used to assign 113 patients referred for invasive cardiopulmonary exercise testing (iCPET) at Medical University in Graz, Austria into clusters derived from our 10-variables exercise subnetwork. A ‘p’ prior to a variable indicates a measurement at peak exercise. FVC, forced vital capacity; FEV-1, forced expiratory volume in 1 second; MVV, maximum voluntary ventilation; VE, minute ventilation; VO₂, volume of oxygen consumption; CaO₂, content of arterial oxygen; Ca-vO₂, difference in content of arterial – venous oxygen content; SV, stroke volume; CO, cardiac output. Data for continuous and normally distributed variables are expressed as mean ± SD. Data for continuous variables that are not normally distributed are expressed as median [IQR]. Data for categorical variables are presented as N (%). The P-values were calculated as described in the Methods.

	Cluster 1 (N=10)	Cluster 2 (N=48)	Cluster 3 (N=35)	Cluster 4 (N=20)	P Value
BMI (kg/m2)	27.4 ± 2.3	26.4 ± 4.5	24.2 [20.8–26.9]	26.4 ± 5.0	0.09026
FVC (L)	5.1 ± 0.7	3.7 ± 0.6	3.0 ± 0.4	2.6 ± 0.5	<0.0001
FEV-1 (L)	4.0 ± 0.5	3.0 ± 0.5	2.4 ± 0.3	2.0 ± 0.4	<0.0001
pVO2 (mL/min)	2254 ± 358	1613 ± 236	1297 ± 203	1095 ± 213	<0.0001
pVE (L/min)	81 ± 17	64 ± 14	54 ± 10	45 ± 11	<0.0001
MVV (L/min)	123 ± 15	91 ± 15	73 ± 8.9	62 ± 14	<0.0001
ppH	7.34 [7.34–7.37]	7.36 ± 0.04	7.38 ± 0.04	7.39 ± 0.05	<0.005
pSV (ml/min)	129 ± 17.0	106 [96–120]	89 [81–105]	96 [80–111]	<0.0001
pCaO2 (ml/dL)	19.3 ± 0.4	16.9 ± 1.9	17.0 ± 1.2	15.5 ± 2.3	<0.0001
pCa-vO2 (mL/dL)	12.1 ± 2.2	10.7 ± 1.8	10.6 ± 1.7	9.3 ± 1.4	<0.0001
Lactate (mmol/L)	7.5 ± 2.3	6.6 ± 1.9	5.9 ± 1.4	4.3 ± 1.6	<0.0001

The median follow-up for all-cause hospitalization was 3125 [2580–3560] vs. 3329 [2530–3692] vs. 2870 [2145–3628] vs. 3161 [1793–3602] days (P=0.65) for clusters 4, 3, 2, and 1, respectively. The 3-year all-cause hospitalization rate in this validation cohort was 75%, 37%, 33%, 10%, for clusters 4, 3, 2, and 1, respectively (P<0.001) (Figure 4C), which paralleled findings in the derivation cohort. Overall, a significant difference was not observed for the 3-year all-cause hospitalization rate by cluster between the derivation and Graz, Austria, cohorts (P=0.28) (Figure 4D). Compared to pVO₂ alone, the AUC for pVO₂ was increased by adding the 10-variable subnetwork model (0.59 vs. 0.69), although the analysis did not achieve statistical significance owing to limited sample size (P=0.195).

DISCUSSION

In this study, network analyses were used to describe a novel association between 10 iCPET measurements. Based on these findings, 4 specific groups were identified from a large cohort of patients referred for evaluation of exercise intolerance. Patient groups were characterized by a unique exercise profile across pulmonary function, exercise hemodynamic, and metabolic data independent of post-test diagnosis, and corresponded to differences in rates of hard clinical end-points. These observations provided the basis for developing the first clinical risk calculator derived from iCPET data collected systematically, which was then verified in two additional patient cohorts.

A number of recent reports have expanded the clinical characteristics of diseases that influence exercise performance, such as heart failure with preserved ejection fraction,¹⁴ chronic obstructive pulmonary disease,¹⁵ and sarcopenia¹⁶, among others.¹⁷ However, comprehensive studies utilizing network-based methods to analyze exercise performance directly are lacking. Moreover, prior reports investigating the clinical relevance of analyzing combinations of exercise measurements have aimed to alter the diagnostic yield for a specific disease sub-type, such as adjusting the pulmonary artery pressure threshold at peak exercise to inform the prevalence of ‘exercise-induced pulmonary hypertension’.^7,18,19 By contrast, findings from our study demonstrate that unique and clinically relevant exercise profiles emerge when considering the gamut of available exercise data.

We observed that a wide range of factors underlie exercise intolerance and influence prognosis. For example, the patient group at highest risk (i.e., cluster 4) demonstrated diminished exercise capacity in association with distinct differences in nearly all of the analyzed exercise variables. These patients were characterized by abnormalities across a heterogeneous array of exercise measures, including baseline FVC and FEV-1, as well as low Ca-vO₂ and stroke volume, and increased PVR, right atrial pressure, and mPAP at peak exercise. Moreover, a continuum toward normalization of values for each of these measures (and many others) was observed across the other three clusters and corresponded to an attendant decrement in clinical risk. Considering that differences in outcome by cluster were largely maintained after adjusting for age and pVO₂, these collective data emphasize the advantage of utilizing all exercise variables when assessing exercise pathophysiology and prognosis in clinical practice.

Findings from this study have several implications for clinical practice. Our observations are consistent with the growing body of evidence in cardiopulmonary medicine indicating that comprehensive methods are useful for prognosticating patients optimally.^20–22 Specifically, our data expand the range of clinical variables beyond pVO₂ (or other single variables) alone that are useful for assessing prognosis.⁴ We show that exercise intolerance is a continuum based on multiple exercise elements, suggesting the potential benefit of multi-dimensional therapeutic strategies to address functional decline.^23,24 It may also be the case that effective exercise pathophenotype-specific treatments are lacking currently due to under-recognition of alternative factors (some measured non-invasively) that contribute to exercise impairment in these patients.

For example, our study cohort included a sizeable population of patients with normal left ventricular ejection fraction and elevated PAWP at rest or with exercise suggesting a diagnosis of heart failure with reduced ejection fraction (HFpEF). Yet, PAWP did not emerge as a variable in the network and outcome differences across the patient clusters were maintained after adjusting for exercise PAWP level. Thus, our data imply that targeting PAWP by enhancing diuresis alone may exclude other treatable contributors to exercise intolerance and adverse outcome in HFpEF. This could include escalating pharmacotherapies to improve pulmonary function, targeting physical conditioning to improve peripheral O₂ extraction, and perhaps adjusting medical therapies that have been shown to affect right ventricular stroke volume, which was a variable in the exercise subnetwork, such as β-adrenergic receptor antagonists.²⁶

Similarly, our network analysis suggests that treating patients with pulmonary vascular disease provoked by exercise^7,18,19 may require targeting as a collective: aerobic conditioning to improve peripheral O₂ extraction, pulmonary function to improve pulmonary mechanics/ventilation, and RV stroke volume to improve cardiac pump function and O₂ delivery. This combined goal may be achieved, for example, through prescription exercise, optimization of pulmonary bronchodilator therapy, and pulmonary vasodilator therapy in appropriate patients.^27,28 Further studies assessing the efficacy of such multidimensional approaches to the management of exercise intolerant patients are warranted based on these study findings, with particular emphasis on diseases in which decision-making is based on a narrow subset of exercise clinical variables, such as cardiac transplantation for hypertrophic cardiomyopathy (e.g., pVO₂)²⁵ or exercise-pulmonary arterial hypertension (e.g., PVR).³

The exercise risk calculator is another important result of this study that bears on clinical practice by providing physicians without formal training in exercise physiology an opportunity to interpret the clinical relevance of iCPET study data at point-of-care relative to 3-year hospitalization rate. This may be particularly relevant to patients who are classified as normal using classical exercise diagnosis criteria. Nearly one-fifth of the study population met this classification, but this subset of patients populated all 4 clusters. Thus, the point-of-care risk calculator is an effective strategy to quantify prognosis in at-risk patients and its use may identify specific exercise parameters that are abnormal and require clinical attention, but would have otherwise been unrecognized by the referring physician. Importantly, the calculator uses a novel approach that permits integrating values from multiple exercise variables in combination, and matching the aggregate of this analysis to patients with a similar exercise profile. In this way, the model determines cluster assignment for a patient based on the interrelatedness of 10 functionally related variables, irrespective of clinical comorbidities. Therefore, the network-based analytical strategy used in this study may function as a complementary approach to traditional methods utilizing linear regression, which emphasize variable weighting and single positive or negative predictors of outcome.^29–30 However, analyses in this study were not designed to show superiority of network-based methods for risk stratification compared with any other approach, and the results should not be interpreted in this regard.

Another important finding from this study pertains to the prevalence of standard exercise diagnoses across iCPET clusters. In cluster 4, which had the lowest average pVO₂ and highest clinical event rate, the prevalence of three diagnoses (e.g., PVD, LHD+PVD, and peripheral O₂ extraction disorder) accounted collectively for ~60% of patients. Similarly, presumed normal patients were well represented in clusters 3, 2, and 1, while the rate of pulmonary mechanical limit to exercise or patients not meeting any traditional exercise diagnosis was evenly distributed across all clusters. This finding exposes our limited knowledge of overlapping pathophenotypes,^31–34 and suggests that isolating specific variables to define exercise subtypes using conventional diagnostic approaches may limit recognition of the diverse contributors to exercise intolerance clinically.³⁵

Several limitations of this study merit consideration when interpreting our conclusions. Adjusting the criteria used to group iCPET variables based on their functional similarities would likely influence network topography and scale, and, therefore, alter our findings. Additionally, a contribution by phenotype extremes to the observed differences in event rate by cluster cannot be excluded. Analyzing the change in cardiopulmonary hemodynamic indices between rest and peak exercise was not performed in this analysis, but could provide an important systems-based dimension by which to hone further interpretation of iCPET datasets.³⁶ Although the combination of invasive and non-invasive parameters from the subnetwork provided a modest incremental advantage for predicting outcome compared to non-invasive variables alone, these findings do not negate the importance of non-invasive or minimally invasive testing for diagnosing and risk-stratifying common causes of exercise intolerance, such as a pulmonary mechanical limitation to exercise and heart failure. Similarly, these findings should not be used to justify iCPET rather than standard CPET clinical practice, which is a decision requiring point-of-care and individualized consideration.

Risk estimates may have been subject to bias within and between centers participating in this study due to variability in patient referral pattern, indications for iCPET, clinical volume, threshold for hospitalization, as well as minor differences in iCPET methodologies. Indeed, event rates for each cluster varied across the primary and validation cohorts, suggesting that these factors, in addition to differences in the study population size and follow-up period, could influence prognosis estimates. In addition, by normalizing values to calculate the k-means cluster, it is possible that patients with extreme exercise profiles or outlier values on individual iCPET variables skewed risk estimates. These confounders were not analyzed specifically in this study. Therefore, additional prospective research in other populations is needed to validate the generalizability of the network strategy and risk calculator, particularly in populations characterized by common forms of exercise intolerance that may have been underrepresented (or underdiagnosed) in this study, such as patients with heart failure with reduced left ventricular ejection fraction or valvular heart disease.

In conclusion, this study provides for the first-time results from a clinically relevant systems-based analysis of patients with exercise intolerance. These methods permitted an unbiased assessment of factors that are important for classifying patients following iCPET, and resulted in the assembly of a novel 10-variable model from pulmonary function, cardiopulmonary hemodynamic, and metabolic elements that defined 4 exercise groups. These groups were associated with significantly different exercise profiles and clinical outcomes, which were not contingent solely upon pVO₂, thereby illustrating the importance of expanding the range of variables under consideration when interpreting exercise intolerance in clinical practice. Further prospective studies are warranted to validate these data and the derivative exercise clinical risk calculator to characterize appropriately the implications of our findings for patient prognosis, treatment candidacy, and therapy selection for the sizeable population of patients with exercise intolerance.

Novelty and Significance.

What Is Known?

• Unexplained exercise intolerance is common and associated with increased clinical morbidity and mortality.

• Exercise is a complex process that requires coordinated activity between multiple physiological systems.

• Current methods for diagnosing patients focus on a very narrow subset of exercise variables.

• Data integrating the totality of available exercise variables to classify and risk stratify patients are lacking.

What New Information Does This Article Contribute?

• Network methods were applied to a large cohort of patients with unexplained exercise intolerance referred for invasive cardiopulmonary exercise testing (iCPET).

• A novel collection of 10 functionally related iCPET variables emerged from the network, and these variables were used to optimally phenotype and risk-stratify exercise intolerant patients.

• These data might be useful for prognosticating exercise intolerant patients using a network approach that integrates information from multiple physiological systems.

Exercise intolerance is an independent risk factor for adverse clinical events, and may be a consequence of cardiovascular, pulmonary, musculoskeletal, hematological, or metabolic dysfunction. However, current methods to diagnose patients with exercise intolerance rely on a limited number of exercise variables. Network medicine permits analysis of multiple interactions between variables simultaneously. In this study, network methods identified a novel collection of 10 exercise variables that were functionally related to one another. These 10 variables were used to optimally phenotype exercise intolerance patients and develop a risk calculator tool for prognosis. This approach differs from classical methods, such as logistical regression, by assessing risk based on interrelated variables, rather than single variables that are independent predictors of outcome. Findings inform exercise intolerance phenotypes and allow risk assessment. Methods from this study may also bear on diagnosing and prognosticating patients with other complex diseases that are characterized currently by overemphasizing a small subset of available variables.

Nonstandard Abbreviations and Acronyms

iCPET

invasive cardiopulmonary exercise testing

NMI

normalized mutual information

LDA

linear discriminant analysis

QDA

quadratic discriminant analysis

PVD

pulmonary vascular disease

LHD+PVD

left heart disease with pulmonary vascular disease

LHD-noPVD

left heart disease without pulmonary vascular disease