Program Information
Head and Neck Cancer Survival Outcome Prediction Based On NRG Oncology RTOG 0522 with Random Forests and Random Survival Forests
M Huang1*, C Cheng2 , H Geng3 , H Zhong4 , J Wang5 , A Lin6 , D Guttmann7 , J van Soest8 , A Dekker9 , W Bilker10 , Z Zhang11 , D Rosenthal12 , R Axelrod13 , J Galvin14 , S Frank15 , W Thorstad16 , B Huth17 , A Hsu18 , A Trotti19 , Q Zhang20 , Y Xiao21 , (1) University of Pennsylvania, Philadelphia, PA, (2) University of Pennsylvania, Philadelphia, PA, (3) University of Pennsylvania, Bryn Mawr, Pennsylvania, (4) University of Pennsylvania, Philadelphia, PA, (5) Fudan University Shanghai Cancer Center, Shanghai, ,(6) University of Pennsylvania, Philadelphia, PA, (7) University of Pennsylvania, Philadelphia, PA, (8) MAASTRO, Maastricht, NT, (9) MAASTRO, Maastricht, NT, (10) University of Pennsylvania, Philadelphia, PA, (11) Fudan University Shanghai Cancer Center, Shanghai, shanghai, (12) MD Anderson Cancer Center, Houston, TX, (13) Thomas Jefferson University Hospital, Philadelphia, PA, (14) Thomas Jefferson University Hospital, Philadelphia, PA, (15) UT MD Anderson Cancer Center, Houston, TX, (16) Washington University School of Medicine, St. Louis, MO, (17) University of Cincinati, Philadelphia, PA, (18) University OF California San Francisco, San Francisco, CA, (19) H. Lee Moffitt Cancer Center & Research Institute, Tampa, FL, (20) American College of Radiology, Philadelphia, PA, (21) University of Pennsylvania, Philadelphia, PA
Presentations
SU-F-FS1-5 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: Four Seasons 1
Purpose: To use the machine learning Random Forest (RF) and novel survival prediction with Random Survival Forest (RSF) models to predict overall survival (OS) and local regional recurrence (LR) based on concurrent accelerated radiation, cisplatin, and cetuximab stage III&IV head and neck squamous cell carcinoma patients (940) enrolled in RTOG 0522 study.
Methods: The demographics, prognosis, dosimetric, and RT dose information were collected as features for prediction model of RF and RSF. Features included: gender, age, specific tumor site, tumor size, treatment arm, tumor staging (cT/cN/cM), hemoglobin_level, EGFR_score, smoking, HPV(P16), RT_dose, treatment_durations, tumor_volume (TV) and organ_at_risk (OAR) dosimetric score and contour scores. Patients with above features were processed into the RF model for predicting overall survival (OS) and local recurrence (LR). Feature importance was automatically ranked. RSF incorporates RF further assesses the corresponding impact of these features for LR and OS outcomes. To pursue RSF for event-specific survival prediction, the initial RF model was five-fold cross-validated.
Results: RSF prediction models stabilized at 5000 randomized decision trees, with concordance index (C-index) resulting in 0.711 and 0.692 for OS and LR respectively. The ‘out of bag’ error was 0.301 corresponding to a training set C-index of 0.711 and test set C-index of 0.715. The important features (Gini_metric) for OS from high-to-low are: hemoglobin_level, tumor_ size, age, RT_duration, smoking, cN, EGFR_score, cT, HPV(P16). For LR, they are: tumor_size, hemoglobin_level, age, RT_duration, EGFR_score, smoking, cN, TV_DVA score, cT, TV_contour_score. The RSF analysis further gives survival prediction over time, with 88.9%, 80.0%, 72.5%, 67.4%, 65.6% survival rates from 1 to 5 years.
Conclusion: RF and RSF provide a novel machine learning statistical prediction for OS and LR of stage III and IV head and neck carcinoma patients, which features automatic data mining and stable performance, a potential tool for decision support of precision medicine.
Funding Support, Disclosures, and Conflict of Interest: This project was supported by grants U10CA180868 (NRG Oncology Operations), U10CA180822 (NRG Oncology SDMC), U24CA180803 (IROC), from the National Cancer Institute (NCI), Eli Lilly. This project is funded, in part, under a grant with the Pennsylvania Department of Health. The Department specifically disclaims responsibility for any analyses, interpretations or conclusions.
Contact Email: