What's New

ATS: Application of Machine Learning in Pulmonary Arterial Hypertension Risk Stratification.

M. Kanwar , J. Kraisangka , L. Lohmueller , J.V. Scott , G. Strange , J. Anderson , M. Hoeper , C. Zhao , C. Pausch , J. Antaki , M. Druzdzel , C. Garnett , R.L. Benza. ATS MAY 2020 B97. WHAT’S NEW IN CLINICAL RESEARCH IN PULMONARY HYPERTENSION: LESSONS FROM THE BEST ABSTRACTS American Journal of Respiratory and Critical Care Medicine 2020;201:A4238

Rationale: Accurate risk stratification in patients with PAH is essential to allow clinicians to determine patients’ prognoses, identify treatment goals and monitor disease progression. Various risk stratification tools, including the recently updated REVEAL 2.0, are derived from large patient registries but are associated with inherent limitations. We hypothesize that the current risk assessment tools are limited due to their assumption of independent, linear relationships between significant clinical variables and that these limitations can be overcome by utilizing modern machine learning algorithms. We sought to demonstrate the utility of machine learning in enhancing the predictive ability of the existing REVEAL 2.0 tool (AUC 0.76).

Methods: Our derivation cohort included patients in REVEAL registry who survived ≥1-year post-enrolment (n=2,456) to develop a Bayesian network (BN) risk model to predict 1-year survival titled Pulmonary Hypertension Outcomes Risk Assessment (PHORA). We used the same variables and discretization cut points as the REVEAL 2.0 risk score calculator. A Tree Augmented Naïve (TAN) Bayes algorithm was used for structure and parameter learning. We performed a 10-fold cross validation within the REVEAL registry and then externally validated the PHORA algorithm in the European (COMPERA), Australian / New Zealand (PHSANZ) registries as well as patients enrolled in the The Ambrisentan and Tadalafil in Patients with Pulmonary Arterial Hypertension (AMBITION) clinical trial.

Results: PHORA was derived from 2,456 PAH patients in the REVEAL registry using Bayesian analysis. The majority of the patients were previously diagnosed (73%), females (80%), NYHA FC II (41.3%) or FC III (45.9%), and had a mean age of 53.6 years. The derived PHORA model (Figure 1A) had an AUC of 0.80 for predicting one-year survival when validated internally. External validation in COMPERA and PHSANZ registries resulted in an AUC of 0.74 and 0.80 respectively, while AMBITION trial patients were stratified with an AUC of 0.71. (Figure 1B)

Conclusion: A BN model for PAH (PHORA) demonstrated an improvement in accuracy over the existing multi-variate model of the REVEAL 2.0 calculator that persisted when externally validated in two large PAH registries and a contemporary clinical trial. This improvement stems from the BN model’s ability to understand both the dynamic influences of each risk factor on each other, as well as with the outcome itself.

This abstract is funded by: NIH NHBLI