Background: Cardiopulmonary exercise testing (CPET) is a cornerstone of prognostication for patients with heart failure. CPET involves maximal exertion and measured general exercise capacity and respiratory function, for each patient breath during the test, however, in clinical practice this data is discarded and prognostication is based on a few summary indicators. We hypothesize that prognostication from exercise testing could be improved by considering the sum total of the data generated during CPET as opposed to using summary indicators alone.

Method: Complete data (including staged exercise measures and breath-by-breath data) from 1123 patients was extracted from the CPET station and used to predict progression of heart failure (defined as need for mechanical circulatory support, heart transplant, or death) within one year of the CPET study Patient-specific fourth-order polynomial equations were used to include breath-by-breath data as features in the predictive models. We compared the predictive ability using the area under the receiver operative curve (AUC) of individual parameters using logistic regression models and three levels of neural nets including: 1) indicators used in classic clinical predictive models, 2) all summary and staged data other than breath-by-breath data and 3) all generated data including breath-by-breath data. Model performance was assessed using a 100-iteration cross-validation. A subset of 598 patients had a second CPET study at least 300 days after the first test and in this subgroup, we assessed whether using data from both studies would improve algorithm performance over using the most recent study alone.

Results: A total of 105/1123 (9.3%) patients in the primary cohort and 45/598 (7.5%) in the secondary cohort experienced clinical deterioration (primary outcome). Statistical differences were observed between patients with clinical deterioration and stable patients for almost all indicators. In terms of prognostication, individual CPET indices showed AUC values below 0.80 including: peak VO2 (AUC: 0.772), VE/VCO2 slope (AUC: 0.772) and oxygen uptake efficiency slope (0.796). At the same time, the 3 neural nets developed for this study achieved higher AUCs: all summary indices (AUC: 0.807), summary indices and staged data (AUC: 0.817) and all data including breath-by-breath data (AUC: 0.813). Interestingly, in the secondary cohort with two CPET studies; none of the models with two years’ worth of data demonstrated significant improvements over the models created with only the most recent data.

Conclusions: This study suggests that the full value of CPET studies in terms of prognostication is not fully exploited by the current clinical practice of considering individual indices in isolation. We have also shown that using neural nets to make full use of the data generated through CPET studies for prognostication improves overall performance. Future studies will need to investigate alternate ways of modelling the breath-by-breath data and to determine the relative value of CPET studies compared to other clinical characteristics in predicting clinical deterioration for patients with heart failure.




Author: Brigitte Mueller

Coauthor(s): Brigitte Mueller, Heather J. Ross, Jason Hearn, Chun Po Fan, Edgar Crowdy, Joe Duhamel, M. Walker, A. Carolina Alba, Cedric Manlhiot

Status: Completed Work