OBJECTIVES: Cystic Fibrosis (CF) is a progressive genetic disease affecting the lungs. Referral for lung transplant (LTx) is recommended when patients face a high likelihood of short-term mortality. Current models for predicting mortality in CF have poor performance, making LTx referral decision-making difficult and inconsistent in practice. Our objective is to apply machine learning (ML) methods to develop a high-performing risk prediction model for short-term mortality in CF.

METHODS: We used data from the Cystic Fibrosis Foundation Patient Registry to develop a mortality risk prediction model for adults with CF. We used the lasso method to develop a preliminary ML model on a limited initial set of clinical and demographic variables. We evaluated the time-varying performance of baseline predictions using the area under the receiver operating curve (AUC) over time and summarized 2-year performance using the survival concordance index (c-index). Performance was compared to the existing model, forced expiratory volume in 1 second (FEV₁) alone.

RESULTS: Our lasso model identified 24 predictors of mortality from the initial limited dataset and had a higher AUC at all time points compared to the existing FEV₁ only model (c-index 0.89(0.86 - 0.92) vs 0.85(0.81 - 0.88)).

CONCLUSIONS: The lasso predicted mortality better than FEV₁ alone for adults with CF in the US using a preliminary dataset. We are now training models on an expanded dataset using additional ML approaches, including ridge, elastic net, support vector machines, random forests, and boosting. Instead of choosing only one ML model, we will create an optimally weighted combination of these different models using ensemble learning methods. We hypothesize that even greater gains in performance will be achieved.