(130) A Machine learning approach for estimating FEV₁% in Cystic Fibrosis using CT-derived lung parenchymal biomarkers

Background: Computed Tomography (CT) attenuation patterns in the lung become more complex with disease progression in patients with Cystic Fibrosis (CF) due to chronic inflammation of lung parenchyma [1]. Although Pulmonary function tests (PFT) remain the gold standard for monitoring lung disease in CF, some studies have demonstrated that CT-based radiology scores could detect changes earlier than PFT, suggesting CT sensitivity in CF [2]. However, obtaining radiology scores is time-consuming and subjective. Lung parenchymal biomarkers derived from CT image processing are quick, reproducible, and were previously demonstrated to quantify parenchymal abnormalities [3]. Our central hypothesis is that imaging biomarkers have translational applications in CF. To test this hypothesis, we developed a Machine Learning (ML) model to estimate FEV1% in patients with CF using CT-derived lung parenchymal measures alone.

Methods: Lung parenchymal biomarkers describing different components of CT attenuation distribution were derived using an in-house developed automated program and pyradiomics [4]. Subjects with CTs and PFTs within 3 months from our CF center were retrospectively included. We trained ML regression models using Decision Trees, Random Forest, and Extreme Gradient Boosting (XGBoost) to estimate FEV1% measured by PFT using CT-derived parenchymal measures. The dataset was split in a 70:30 ratio between training and testing tasks of ML using stratified sampling. 10-fold cross-validation was used to find the best hyperparameters for each model. Root mean squared error (RMSE) is used to assess the ML model performances.

Results: We included 153 CF patients in this study (Age: 14.3 ± 3.6 years; 59 male/94 female; FEV1% predicted as measured by PFT: 82.8 ± 22.8). A total of 121 lung parenchymal measures were derived from patient CT scans. ML regression models were trained on 23/121 lung measures after feature selection. XGBoost regression model obtained the best performance in estimating FEV1% with an RMSE of 6.4%. The ML model error in predicted FEV1% is on the same order of repeatability errors in FEV1% in spirometry reported in the 2012 NIOSH Spirometry Quality Assurance Study.

Conclusions: This study demonstrates that Machine Learning models have translational applications in predicting FEV1% in CF. Future studies will further refine this ML model by including clinical data, demographics and eventually deploy the model to inform clinical practice.

Acknowledgements: Supported by Cystic Fibrosis Foundation RDP grant

References: 1. Kołodziej, M., de Veer, M. J., Cholewa, M., Egan, G. F., & Thompson, B. R. (2017). Lung function imaging methods in Cystic Fibrosis pulmonary disease. Respiratory Research, 18(1), 1-11.

2. de Jong, P. A., Nakano, Y., Lequin, M. H., Mayo, J. R., Woods, R., Pare, P. D., & Tiddens, H. A. W. M. (2004). Progressive damage on high resolution computed tomography despite stable lung function in cystic fibrosis. European Respiratory Journal, 23(1), 93-97.

3. Mascalchi, M., Camiciottoli, G., & Diciotti, S. (2017). Lung densitometry: why, how and when. Journal of thoracic disease, 9(9), 3319.

4. van Griethuysen, J. J.; Fedorov, A.; Parmar, C.; Hosny, A.; Aucoin, N.; Narayan, V.; Beets-Tan, R. G.; Fillion-Robin, J.-C.; Pieper, S. & Aerts, H. J. Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 2017, 77, e104-e107