External validation of clustering models for cardiovascular risk stratification based on cardiopulmonary exercise testing
E J Santana, E Ntalianis, J Christle, N Cauwenberghs, J Claes, M T Wheeler, R Brouwers, G Fahed, E A Ashley, J Myers, V Cornelissen, F Haddad, T KuznetsovaAbstract
Introduction
Cardiopulmonary exercise testing (CPET) enables the assessment of cardiorespiratory fitness. Previously, sex-specific clustering models were trained on time series from cycle ergometry in a Belgian cohort. Those models provided a promising artificial intelligence-based integrative tool for cardiovascular (CV) risk stratification, warranting external validation before its implementation in clinical practice.
Purpose
To evaluate the performance of the CPET time series clustering models in an external cohort from the United States tested on a treadmill.
Methods
The external cohort consisted of individuals who achieved a respiratory exchange ratio (RER)≥1.0 during CPET at a university medical center between 2014 and 2020. We included individuals without a diagnosis of heart failure or congenital heart disease that were referred to CPET mainly due to symptoms or for cardiopulmonary disease evaluation. The previously developed models, derived using dynamic time warping with k-medoids (unsupervised machine learning), were applied to time series of oxygen uptake, heart rate, RER, and end-tidal partial pressure of oxygen and carbon dioxide of the external cohort. The resulting clusters were evaluated regarding clinical characteristics and health outcomes. Events included stroke, atrial fibrillation, cardiac arrest, acute coronary syndrome, acute heart failure, heart or lung transplants, cardiac and valve surgeries, cardiogenic shock, shockable rhythm or arrythmia surgery. The median follow-up time was 2.5 years with maximum time set to 4 years.
Results
Among the 751 patients, 277 (36.9%) were females; the mean age was 47±16 years, body mass index was 27±5 kg/m2, and 447 (59.5%) had a history of CV disease. Cluster 1 represented the most favorable CV profile whereas cluster 5 represented the worst profile, consistent with the observations in the training data. Compared to the other clusters, cluster 5 had significantly higher values for age and hypertension, and lower peak VO2 values (Table 1). In 1841 person-years (py) of follow-up, 183 individuals experienced at least one CV event (99.4/1000 py) and the event rates increased gradually across the clusters: 52.5/1000 py in cluster 1 (35 events), 67.5/1000 py in cluster 2 (31 events), 113.3/1000 py in cluster 3 (45 events), 207.6/1000 py in cluster 4 (42 events) and 259.4/1000 py in cluster 5 (30 events). In line, the Kaplan-Meier curves showed differences in cumulative incidence of events between clusters (Figure 1A). After adjustments for traditional risk factors, clusters 4 and 5 presented significantly higher CV risks compared to cluster 1 (Figure 1B).
Conclusion
The previously developed CPET-based clustering models were able to stratify CV risk and event rates in an external cohort, tested in a different country and CPET modality. These findings increase the confidence of using those models as a decision support tool for CV-risk assessment based on CPET time series.