Integrative Interpretation of Cardiopulmonary Exercise Tests for Cardiovascular Outcome Prediction: A Machine Learning Approach

Figures

• 

  Figure 1. Workflow of the unsupervised machine learning. The blue and orange rhomboids represent the input data and the output of the analysis, respectively. The rectangles describe analytical steps. The dashed line represents data supporting the phenogrouping. CPET, cardiopulmonary exercise testing.

• 

  Figure 2. Correlations between the spirometric and the CPET features. Strong correlations, those with an absolute Pearson correlation coefficient > 0.8, were framed in red in the correlation matrix (A) and presented below (B). The spirometric measurements were eventually excluded from the phenogrouping, due to the extensive amount of spirometric data missing (23.0%). In addition, five CPET metrics were excluded from the phenogrouping due to high collinearity (see B), leaving ten features for the phenomapping (in bold in (A)). Abbreviations as in Table 1.

• 

  Figure 3. Radar charts of the CPET features used for the phenogrouping illustrate the superposition of these features in each of the four phenogroups identified for each sex. The plot lines compare the clusters’ standardized values expressed as z-scores relative to the global average (=0) across the ten dimensions used as inputs to the clustering process. Abbreviations as in Table 1.

• 

  Figure 4. The incidence of cardiovascular events by the integrative CPET profiles. * and ** respectively indicate a p value below 0.05 and below 0.001 for a pairwise logrank test.

• 

  Figure 5. The risk for cardiovascular events by the integrative CPET profiles. The adjusted hazard ratios (95% CI) accounted for differences in age, height, weight, heart rate, systolic and diastolic blood pressure, antihypertensive drug intake and a history of diabetes mellitus, chronic kidney disease, and cardiovascular intervention at baseline. Table 2 details the hazard ratios and p values from the unadjusted and adjusted models, including the models accounting for differences in VO2/kgpeak or % predicted VO2,peak.