QP 11 - DHI 2: Innovative Uses of AI in Cancer Treatment and Patient Care
1052 - Predicting Pulmonary Toxicity in Lung Cancer Patients with Interstitial Lung Disease Receiving SABR Using Machine Learning: A Secondary Analysis of ASPIRE-ILD
M. Gil1,2, S. Harrow2, W. Nailon2,3, S. Marshall1, R. Doucet4, H. Bahig5, J. P. Bissonnette6,7, A. Hope8, B. J. Debenham9, S. Gaede10, A. Warner10,11, C. Ryerson12, and D. A. Palma10; 1University of Strathclyde, Glasgow, United Kingdom, 2Edinburgh Cancer Centre, Edinburgh, United Kingdom, 3University of Edinburgh, Edinburgh, United Kingdom, 4Centre Hospitalier de lUniversité de Montréal, Montreal, QC, Canada, 5Centre Hospitalier de lUniversité de Montréal, Montréal, QC, Canada, 6Department of Medical Physics, University Health Network, Toronto, ON, Canada, 7Department of Radiation Oncology, University of Toronto, Toronto, ON, Canada, 8Department of Radiation Oncology, University Health Network, Toronto, ON, Canada, 9University of Alberta, Edmonton, AB, Canada, 10London Health Sciences Centre, London, ON, Canada, 11Department of Radiation Oncology, Western University, London Health Sciences Centre, London, ON, Canada, 12Department of Medicine, University of British Columbia, Vancouver, BC, Canada
Purpose/Objective(s): Stereotactic ablative body radiotherapy (SABR) is an emerging treatment option for lung cancer patients with interstitial lung disease (ILD). Identifying patients who are at greater risk of toxicity after SABR would help personalize their treatment. The aim of this study was to develop and validate methods to predict pulmonary toxicity in patients with ILD receiving SABR using radiomic and dosimetric-based machine learning methods.Materials/
Methods: This retrospective study used data from the ASPIRE-ILD clinical trial, currently the largest multi-centre trial investigating the use of SABR for the treatment of lung cancer patients with ILD. Dose-volume histogram features were calculated from the treatment plans and included the Vx, as both a percentage and volume, in steps of 5 Gy from 5 to 60 Gy, the mean and max lung dose, the D2cm, and the R50. Radiomic features were calculated from the lung volume of each patient’s diagnostic pre-treatment CT scan using PyRadiomics. Additionally, deep learning-based features were extracted from each CT scan using a UNet convolutional neural network (CNN). The AdaBoost algorithm for boosted decision trees was used to train a prediction model for the binary prediction of grade = 2 pulmonary toxicity, splitting the patient cohort into low and high-risk groups. Different combinations of feature sources were tested to determine the source of any predictive power and a leave-one-out cross-validation approach was applied to test over the full dataset. Results: Baseline planning scans and diagnostic imaging was available from all 39 patients, of whom 11 developed grade = 2 CTCAE v5 pulmonary toxicity with dyspnea (n=7) and pneumonitis (n=3) being the predominant toxicities. The best-performing model included dose, CT radiomic and CT deep learning features to achieve an AUC of 0.841 with a sensitivity of 81.8% and a specificity of 78.6%. This model correctly classified 9/15 patients as high-risk and 22/24 patients as low-risk giving a precision and negative predictive value of 60% and 92% respectively. Table 1 displays all model results. Conclusion: A radiomic- and dosimetric-based machine learning method to predict the occurrence of pulmonary toxicity in lung cancer patients with ILD receiving SABR has been developed and validated. To our knowledge, this is the first study showing the benefit of radiomic and deep learning CT features, or to use of machine learning for stratifying these patients into low- and high-risk groups. Validation in a larger and more diverse dataset is needed.
Test results for the AdaBoost model trained to predict grade = 2 toxicity using different feature sets. AUC is the area under the ROC curve.
