- Email: [email protected]
Natural History Following Recurrence After Definitive Locoregional Treatment in 1000+ Cases of Locally Advanced Head and Neck Squamous Cell Carcinoma
Volume 96 Number 2S Supplement 2016 with patient immobilized in treatment position. Treatment delivery systems include Linac with CBCT (n Z 10), robotic SBRT (n Z 8), Co-60 SRS (n Z 2) and proton systems (n Z 2). Twelve institutions use daily pretreatment imaging and online correction. Most centers (n Z 10) use robotic tracking as part of pre-treatment imaging verification and intrafractional monitoring. The action level for correction of set-up errors is 1-3 mm for 10 institutions and 5 mm for 1 institution. Nine centers routinely repeat imaging after couch adjustment prior to treatment delivery, and 3 institutions repeat imaging only if the shift was greater than 3-5 mm. Nine institutions apply a clinical target volume expansion of 1-10 mm and 13 institutions use a planning target volume margin of 1-5 mm. Fractionation and dose varied between 15-22 Gy in 1 fraction to 30-50 Gy in 5 or 6 fractions prescribed to an isodose line. Three institutions deliver treatment on consecutive days and 12 institutions every other day. Conclusion: There is considerable heterogeneity in the techniques used by 15 high volume centers in head and neck cancer SBRT. Further study is needed to understand the impact of these variables on outcomes, which can provide an evidence base for a consensus statement. Author Disclosure: I. Karam: None. M. Yao: None. D.E. Heron: None. I. Poon: None. S. Koyfman: None. S.S. Yom: Research Grant; Genentech. F. Siddiqui: research funding from Varian Medical Systems, Inc. and Philips Medical; Henry Ford Hospital Department of Radiation Oncology. Track Chair; ASTRO Annual Meeting. Panel Member; American College of Radiology. ASTRO Representative; Medical Dosimetry Certification Board. Operations; Henry Ford Hospital. E. Lartigau: Consultant; Accuray. M. Cengiz: None. H. Yamazaki: None. W. Hara: None. J. Phan: None. J. Vargo: None. V.H. Lee: None. R.L. Foote: None. K.W. Harter: None. N. Lee: Advisory Board; Pfizer, Vertex, Merck. A. Sahgal: Research Grant; Elekta AB. Honoraria; Varian Medical Systems, Medtronic. Belongs to research groups supported by Elekta AB; Elekta AB. S.S. Lo: Honoraria; Accuray and Varian Medical Systems. Partial research support; Elekta AB.
Poster Viewing E383 log-rank test. Multivariate analysis was not performed due to small sample size. Results: A total of 58 patients were assessed. Median follow-up for surviving patients was 42.7 months (range 23-53 months). Two-year OS, LRC, DFS and DMFS, for the entire cohort were 62%, 78.3%, 55.2% and 67.2%, respectively. Median pretreatment SUVmax for the primary tumor and lymph nodes was 11.85 (range 3.1e25.8) and 5.4 (range 1.8e18), respectively. Univariate analysis showed that patients with KPS <80% vs 80% (P<0.001), TNM stage IV vs III (P Z 0.037) and patients undergoing radiation therapy vs surgery (P Z 0.042) were significantly associated with lower 2-year OS (0% vs 67.9%; 37.5% vs 87.5% and 48.6% vs 85.7%, respectively). Patients with KPS <80% vs 80% (P Z 0.003) or age 65 years vs < 65 years (P Z 0.007) presented lower 2year LRC (40% vs 81% and 59% vs 90.6%, respectively). However, other parameters including SUVmax of the primary tumor or in the lymph nodes were not associated with OS or LRC. The KPS was the only factor associated with decreased 2-year DFS (0% and 60.4% for patients with KPS <80% and KPS 80%, respectively) (P Z 0.001). On univariate analysis nodal SUVmax was the only factor associated with decreased DMFS. Patients with pretreatment tumor SUVmax that exceeded the median value of the cohort demonstrated inferior 2-year DMFS relative to patients with SUVmax £ the median value of the cohort, 53.1% vs. 80.9%, respectively, P Z 0.016. Conclusion: The pretreatment SUVmax of nodal disease in patients with locally advanced HNSCC is prognostic for DMFS. Author Disclosure: J. Cacicedo: None. I. Fernandez: None. O. del Hoyo: None. A. Gomez-Iturriaga: None. L. Martinez-Indart: None. A. Frias: None. J. Gomez: None. E. Roden˜o: None. O. Rodriguez: None. J. Mendiola: None. A. Sancho: None. R. Ortiz de Zarate: None. F. Perez: None. F. Casquero: None. P. Bilbao: None.
2945 2944 Prognostic Value of Maximum Standardized Uptake Value Measured by Pretreatment 18F-FDG PET/CT in Locally Advanced Head and Neck Squamous Cell Carcinoma: Adding Value to Clinical Staging J. Cacicedo,1 I. Fernandez,2 O. del Hoyo,3 A. Gomez-Iturriaga,4 L. Martinez-Indart,1 A. Frias,1 J. Gomez,1 E. Roden˜o,2 O. Rodriguez,5 J. Mendiola,1 A. Sancho,1 R. Ortiz de Zarate,1 F. Perez,1 F. Casquero,2 and P. Bilbao2; 1Cruces University Hospital, Bilbao, Spain, 2Hospital Universitario Cruces, Barakaldo, Spain, 3Cruces University hospital, Bilbao, Spain, 4Cruces University Hospital/ Biocruces Health Research Institute, Barakaldo, Spain, 5Cruces university hospital, Bilbao, Spain Purpose/Objective(s): To evaluate the prognostic significance of maximum standardized uptake value (SUVmax) in locally advanced headand-neck squamous cell carcinoma (HNSCC) in patients undergoing pretreatment [F-18] fluoro-D-glucose-positron emission tomography/ computed tomography (FDG-PET/CT) imaging. Materials/Methods: This study is a subanalysis of a previous prospective study performed in our institution to determine the incremental staging information provided by PET/CT and its impact on management plans in patients with untreated stage IIIe IV HNSCC. Patients with HNSCC who underwent PET/CT before a radical treatment with definitive radiation therapy (concomitant chemotherapy) or surgery + postoperative (chemo) radiation were analyzed. Patients undergoing treatment with palliative intent or presenting with distant metastasis at diagnosis where excluded. FDG uptake by both the primary lesion and the neck node was measured using the SUVmax value. The effects of clinicopathological factors (age, gender, tumor location, tumor stage, Karnofsky Performance Status (KPS), and treatment strategy) including primary tumor SUVmax and nodal SUVmax on overall survival (OS), disease-free survival (DFS), locoregional control (LRC), and distant metastasis-free survival (DMFS) were evaluated. Kaplan-Meier survival curves were generated and compared via the
Natural History Following Recurrence After Definitive Locoregional Treatment in 1000+ Cases of Locally Advanced Head and Neck Squamous Cell Carcinoma J.E. Leeman,1 J.G. Li,2 P. Venigalla,1 P.B. Romesser,1 Z.S. Zumsteg,3 S.M. McBride,1 C.J. Tsai,1 D.S. Higginson,1 N. Katabi,1 J.O. Boyle,4 B.R. Roman,1 E.J. Sherman,1 N. Lee,1 and N. Riaz1; 1Memorial Sloan Kettering Cancer Center, New York, NY, 2Jiangxi Cancer Hospital, Nanchang, China, 3Cedars-Sinai Medical Center, Los Angeles, CA, 4 MSKCC, New York, NY, United States Purpose/Objective(s): Squamous cell carcinomas of the head and neck (HNSCC) arising from different subsites have unique biologic features and differing clinical outcomes. Herein, we report on the natural history following recurrence in a large cohort of locally advanced HNSCC patients treated with definitive loco-regional treatment with modern radiation techniques (IMRT). Materials/Methods: We reviewed outcomes from 1062 consecutive locally advanced HNSCC patients treated with definite chemo-RT or surgery with post-operative RT +/- chemotherapy from 2001-2013 at a large cancer center, including 734 cases of oropharyngeal carcinoma (OPC), 155 oral cavity carcinoma (OCC), 126 laryngeal carcinoma (LRX) and 47 hypopharyngeal carcinoma (HPX). We analyzed outcomes across sub-sites with a focus on survival following locoregional failure (LRF) or distant metastasis (DM). Median follow-up was 57.7 months in surviving patients. Results: Crude rates of LRF and DM, respectively, by sub-site were: OPC 9.4% and 12%, OCC 25.8% and 12.9%, LRX 19.8% and 18.2%, HPX 31.9%, 29.8%. Time from LRF to death was not significantly different between sub-sites. Across all sub-sites, patients who underwent salvage surgery or salvage re-irradiation following LRF experienced longer survival (surgery vs no surgery, 21.7 vs 6.0 mo, P < 0.01, re-RT vs no re-RT, 21.9 vs. 10.3 mo, P Z 0.04). Time from DM to death was significantly shorter in OCC compared to other sub-sites (OCC median 3.9 months, OPC 13.4 mo, LRX 17.5 mo, HPX 7.5 mo, P < 0.01). Metastatic patients
E384
International Journal of Radiation Oncology Biology Physics
treated after 2006 experienced longer survival across all sub-sites (13.2 vs 6 mo, interquartile ranges [6.2-28.7] vs [3.2-13.4], P < 0.01). Conclusion: Salvage surgery and re-RT following LRF are associated with improved survival. DM from OCC portends inferior outcomes compared to other sub-sites. Metastatic patients treated after 2006, when cetuximab was FDA approved for HNSCC, experienced longer survival. Understanding the natural history of recurrent/metastatic HNSCC after definitive locoregional treatment suggests several unique features which influence outcomes and may affect interpretation of clinical trials in these patients. Author Disclosure: J.E. Leeman: None. J. Li: None. P. Venigalla: None. P.B. Romesser: None. Z.S. Zumsteg: None. S.M. McBride: None. C. Tsai: None. D.S. Higginson: None. N. Katabi: None. J.O. Boyle: None. B.R. Roman: None. E.J. Sherman: None. N. Lee: Advisory Board; Merck, Pfizer, Vertex. N. Riaz: None.
methods provided higher accuracies in keeping with its known strengths in handling large and incomplete datasets. Ensemble models should be studied in greater detail to further understand the meaning of the data. Ensemble methods obtained better predictive performance than other single algorithms by allowing for a more flexible structure in the models. Diverting from the traditional CART models and other single models may be a more effective way of extracting knowledge, especially with sizeable and often incomplete datasets. Author Disclosure: A.M. Hernandez: None. Z. Cheng: None. X. Hui: None. A.P. Kiess: None. S.P. Robertson: None. J. Moore: None. M.R. Bowers: None. A. Choflet: None. J.W. Wong: None. T.R. McNutt: None. H. Quon: None. L. Burns: None. A. Thompson: None.
2946 The Role of Ensemble Machine Learning Algorithms to Predict Weight Loss Following Head and Neck Radiation Therapy A.M. Hernandez,1 Z. Cheng,2 X. Hui,3 A.P. Kiess,3 S.P. Robertson,4 J. Moore,2 M.R. Bowers,5 A. Choflet,3 J.W. Wong,3 T.R. McNutt,4 H. Quon,3 L. Burns,6 and A. Thompson6; 1George Washington University (Graduate Student), Washington, DC, 2Johns Hopkins University School of Medicine, Baltimore, MD, 3Johns Hopkins University, Department of Radiation Oncology and Molecular Radiation Sciences, Baltimore, MD, 4 Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 5Johns Hopkins University, Baltimore, MD, 6Johns Hopkins Hospital, BSN, Baltimore, MD Purpose/Objective(s): Head and neck cancer (HNC) patients often experience weight loss due to various radiation-related toxicities. Predicting weight loss can assist with interventions. The aim of this study was to evaluate the performance of various Machine Learning (ML) classification algorithms and determine dominant predictors for weight loss building on our prior weight loss prediction model using the Classification and Regression Trees (CART) algorithm. Materials/Methods: HNC patients receiving intensity modulated radiation therapy (RT) were queried from Oncospace. Oncospace aggregates prospective structured outcome data captured during routine clinical care with RT planning data also generated during the routine clinical workflow. Multiple ML algorithms available in the computer algorithm’s Classification Learner (CL): Decision Trees, Discriminant Analysis, Support Vector Machines, Nearest Neighbor Classifier, were used to develop prediction models for the primary end point of weight loss (5kg) at 3-month post RT, and compared to Ensemble Classifiers. The predicted response from a trained ensemble is the average of predictions from individual trees. Ten-fold cross-validation was used to protect data against over fitting. Results: From 2007-14, 326 patients with 729 variables collected during treatment and in follow up were analyzed. The primary outcome was a weight loss of (5kg), despite the use of a PEG tube. The predictors utilized with the CL were determined based on the best performance of various ‘weak learning’ ML classification algorithms (decision trees, Knearest neighbors, etc.), then compared to those with a high-quality ensemble model. The patient reported outcomes (PRO) such as “able to eat foods I like” were used with further expanding trees through Ensemble Bagged Trees and anatomic tumor location by ICD-9 resulting in a learning accuracy of 0.844. A comparison of weight loss of (5kg) with radiation dose and chemotherapy predictors resulted in 0.89 and 0.927 accuracies. Additionally, significant accuracies were also obtained for mucositis (0.862), xerostomia (0.865), T-stage (0.869), pain intensity (0.896), PEG used (0.890), and PRO “content with the quality of life” (0.902). Conclusion: ML algorithms were successfully applied modeling weight loss in our head and neck Oncospace informatics platform. Ensemble
2947 Comorbidity With Age as a Predictor of Survival for Patients With Nasopharyngeal Cancer Following Radiation Treatment: A Nationwide Population-Based Study C.C. Yang,1 P.C. Chen,2 L.C. Lin,1 S.L. Chang,3 and C.C. Lee4; 1 Department of Radiation Oncology, Chi-Mei Medical Center, Tainan, Taiwan, 2Department of Radiation Oncology, Pingtung Christian Hospital, Pingtung, Taiwan, 3Department of Otolaryngology, Chi-Mei Medical Center, Tainan, Taiwan, 4Department of Otolaryngology, Head and Neck Surgery, Kaohsiung Veterans General Hospital, Kaohsiung, Taiwan Purpose/Objective(s): To characterize the impact of comorbidity and age on survival outcomes for patients with nasopharyngeal carcinoma (NPC) post radiation therapy (RT). Materials/Methods: A total of 4095 patients with newly diagnosed NPC treated by RT or RT plus chemotherapy (CT) in the period from 2007 to 2011 were included through Taiwan’s National Health Insurance Research Database. Comorbidities present prior to the NPC diagnosis was obtained and adapted to the Charlson Comorbidity Index (CCI) and Age-Adjusted Charlson Comorbidity Index (ACCI). Overall survival was estimated using the KaplandMeier method and the difference between groups was analyzed using log-rank test. The Cox proportional hazards regression model was used for multivariate analysis. Differences between variables with categorical data were examined using the chi-square test. Receiver Operating Characteristic (ROC) curves was generated to assess the accuracy and the predictive ability of each index for survival. Results: Most of the patients (75%) were male (age 5113 years) and 2470 of them (60%) had at least one comorbid condition. The most common comorbid condition was diabetes mellitus. According to these two different comorbidity index (CCI and ACCI), higher scores were associated with worse overall survival (P< 0.001). The Receiver Operating Characteristic (ROC) curve was used to assess the discriminating ability of CCI and AACI and it demonstrated the predictive ability for mortality with the ACCI (0.693, 95% CI 0.670-0.715) was superior to that of the CCI (0.619, 95% CI 0.593-0.644). Conclusion: Comorbidities with age greatly influenced the clinical presentations, therapeutic interventions, and outcomes of patients with NPC post RT. Higher comorbidity index scores accurately was associated with worse survival. The ACCI seems to be a more appropriate prognostic indicator and should be considered in further clinical studies. Author Disclosure: C. Yang: None. P. Chen: None. L. Lin: None. S. Chang: None. C. Lee: None.
2948 Adjuvant Radiation Therapy Improved Local Control for Primary Mucosal Melanoma of the Head and Neck M. Abugideiri,1 K. Patel,2 J.M. Switchenko,3 K. Magliocca,4 Z.S. Buchwald,1 J. Delgaudio,5 D.H. Lawson,6 H. Danish,2 J.J. Beitler,2,5 and M.K. Khan2; 1Department of Radiation Oncology at Emory University, Atlanta, GA, 2Department of Radiation Oncology, Winship Cancer Institute at Emory University, Atlanta, GA, 3Department of Biostatistics and Bioinformatics, Winship Cancer Institute at Emory
Poster Viewing E383 log-rank test. Multivariate analysis was not performed due to small sample size. Results: A total of 58 patients were assessed. Median follow-up for surviving patients was 42.7 months (range 23-53 months). Two-year OS, LRC, DFS and DMFS, for the entire cohort were 62%, 78.3%, 55.2% and 67.2%, respectively. Median pretreatment SUVmax for the primary tumor and lymph nodes was 11.85 (range 3.1e25.8) and 5.4 (range 1.8e18), respectively. Univariate analysis showed that patients with KPS <80% vs 80% (P<0.001), TNM stage IV vs III (P Z 0.037) and patients undergoing radiation therapy vs surgery (P Z 0.042) were significantly associated with lower 2-year OS (0% vs 67.9%; 37.5% vs 87.5% and 48.6% vs 85.7%, respectively). Patients with KPS <80% vs 80% (P Z 0.003) or age 65 years vs < 65 years (P Z 0.007) presented lower 2year LRC (40% vs 81% and 59% vs 90.6%, respectively). However, other parameters including SUVmax of the primary tumor or in the lymph nodes were not associated with OS or LRC. The KPS was the only factor associated with decreased 2-year DFS (0% and 60.4% for patients with KPS <80% and KPS 80%, respectively) (P Z 0.001). On univariate analysis nodal SUVmax was the only factor associated with decreased DMFS. Patients with pretreatment tumor SUVmax that exceeded the median value of the cohort demonstrated inferior 2-year DMFS relative to patients with SUVmax £ the median value of the cohort, 53.1% vs. 80.9%, respectively, P Z 0.016. Conclusion: The pretreatment SUVmax of nodal disease in patients with locally advanced HNSCC is prognostic for DMFS. Author Disclosure: J. Cacicedo: None. I. Fernandez: None. O. del Hoyo: None. A. Gomez-Iturriaga: None. L. Martinez-Indart: None. A. Frias: None. J. Gomez: None. E. Roden˜o: None. O. Rodriguez: None. J. Mendiola: None. A. Sancho: None. R. Ortiz de Zarate: None. F. Perez: None. F. Casquero: None. P. Bilbao: None.
2945 2944 Prognostic Value of Maximum Standardized Uptake Value Measured by Pretreatment 18F-FDG PET/CT in Locally Advanced Head and Neck Squamous Cell Carcinoma: Adding Value to Clinical Staging J. Cacicedo,1 I. Fernandez,2 O. del Hoyo,3 A. Gomez-Iturriaga,4 L. Martinez-Indart,1 A. Frias,1 J. Gomez,1 E. Roden˜o,2 O. Rodriguez,5 J. Mendiola,1 A. Sancho,1 R. Ortiz de Zarate,1 F. Perez,1 F. Casquero,2 and P. Bilbao2; 1Cruces University Hospital, Bilbao, Spain, 2Hospital Universitario Cruces, Barakaldo, Spain, 3Cruces University hospital, Bilbao, Spain, 4Cruces University Hospital/ Biocruces Health Research Institute, Barakaldo, Spain, 5Cruces university hospital, Bilbao, Spain Purpose/Objective(s): To evaluate the prognostic significance of maximum standardized uptake value (SUVmax) in locally advanced headand-neck squamous cell carcinoma (HNSCC) in patients undergoing pretreatment [F-18] fluoro-D-glucose-positron emission tomography/ computed tomography (FDG-PET/CT) imaging. Materials/Methods: This study is a subanalysis of a previous prospective study performed in our institution to determine the incremental staging information provided by PET/CT and its impact on management plans in patients with untreated stage IIIe IV HNSCC. Patients with HNSCC who underwent PET/CT before a radical treatment with definitive radiation therapy (concomitant chemotherapy) or surgery + postoperative (chemo) radiation were analyzed. Patients undergoing treatment with palliative intent or presenting with distant metastasis at diagnosis where excluded. FDG uptake by both the primary lesion and the neck node was measured using the SUVmax value. The effects of clinicopathological factors (age, gender, tumor location, tumor stage, Karnofsky Performance Status (KPS), and treatment strategy) including primary tumor SUVmax and nodal SUVmax on overall survival (OS), disease-free survival (DFS), locoregional control (LRC), and distant metastasis-free survival (DMFS) were evaluated. Kaplan-Meier survival curves were generated and compared via the
Natural History Following Recurrence After Definitive Locoregional Treatment in 1000+ Cases of Locally Advanced Head and Neck Squamous Cell Carcinoma J.E. Leeman,1 J.G. Li,2 P. Venigalla,1 P.B. Romesser,1 Z.S. Zumsteg,3 S.M. McBride,1 C.J. Tsai,1 D.S. Higginson,1 N. Katabi,1 J.O. Boyle,4 B.R. Roman,1 E.J. Sherman,1 N. Lee,1 and N. Riaz1; 1Memorial Sloan Kettering Cancer Center, New York, NY, 2Jiangxi Cancer Hospital, Nanchang, China, 3Cedars-Sinai Medical Center, Los Angeles, CA, 4 MSKCC, New York, NY, United States Purpose/Objective(s): Squamous cell carcinomas of the head and neck (HNSCC) arising from different subsites have unique biologic features and differing clinical outcomes. Herein, we report on the natural history following recurrence in a large cohort of locally advanced HNSCC patients treated with definitive loco-regional treatment with modern radiation techniques (IMRT). Materials/Methods: We reviewed outcomes from 1062 consecutive locally advanced HNSCC patients treated with definite chemo-RT or surgery with post-operative RT +/- chemotherapy from 2001-2013 at a large cancer center, including 734 cases of oropharyngeal carcinoma (OPC), 155 oral cavity carcinoma (OCC), 126 laryngeal carcinoma (LRX) and 47 hypopharyngeal carcinoma (HPX). We analyzed outcomes across sub-sites with a focus on survival following locoregional failure (LRF) or distant metastasis (DM). Median follow-up was 57.7 months in surviving patients. Results: Crude rates of LRF and DM, respectively, by sub-site were: OPC 9.4% and 12%, OCC 25.8% and 12.9%, LRX 19.8% and 18.2%, HPX 31.9%, 29.8%. Time from LRF to death was not significantly different between sub-sites. Across all sub-sites, patients who underwent salvage surgery or salvage re-irradiation following LRF experienced longer survival (surgery vs no surgery, 21.7 vs 6.0 mo, P < 0.01, re-RT vs no re-RT, 21.9 vs. 10.3 mo, P Z 0.04). Time from DM to death was significantly shorter in OCC compared to other sub-sites (OCC median 3.9 months, OPC 13.4 mo, LRX 17.5 mo, HPX 7.5 mo, P < 0.01). Metastatic patients
E384
International Journal of Radiation Oncology Biology Physics
treated after 2006 experienced longer survival across all sub-sites (13.2 vs 6 mo, interquartile ranges [6.2-28.7] vs [3.2-13.4], P < 0.01). Conclusion: Salvage surgery and re-RT following LRF are associated with improved survival. DM from OCC portends inferior outcomes compared to other sub-sites. Metastatic patients treated after 2006, when cetuximab was FDA approved for HNSCC, experienced longer survival. Understanding the natural history of recurrent/metastatic HNSCC after definitive locoregional treatment suggests several unique features which influence outcomes and may affect interpretation of clinical trials in these patients. Author Disclosure: J.E. Leeman: None. J. Li: None. P. Venigalla: None. P.B. Romesser: None. Z.S. Zumsteg: None. S.M. McBride: None. C. Tsai: None. D.S. Higginson: None. N. Katabi: None. J.O. Boyle: None. B.R. Roman: None. E.J. Sherman: None. N. Lee: Advisory Board; Merck, Pfizer, Vertex. N. Riaz: None.
methods provided higher accuracies in keeping with its known strengths in handling large and incomplete datasets. Ensemble models should be studied in greater detail to further understand the meaning of the data. Ensemble methods obtained better predictive performance than other single algorithms by allowing for a more flexible structure in the models. Diverting from the traditional CART models and other single models may be a more effective way of extracting knowledge, especially with sizeable and often incomplete datasets. Author Disclosure: A.M. Hernandez: None. Z. Cheng: None. X. Hui: None. A.P. Kiess: None. S.P. Robertson: None. J. Moore: None. M.R. Bowers: None. A. Choflet: None. J.W. Wong: None. T.R. McNutt: None. H. Quon: None. L. Burns: None. A. Thompson: None.
2946 The Role of Ensemble Machine Learning Algorithms to Predict Weight Loss Following Head and Neck Radiation Therapy A.M. Hernandez,1 Z. Cheng,2 X. Hui,3 A.P. Kiess,3 S.P. Robertson,4 J. Moore,2 M.R. Bowers,5 A. Choflet,3 J.W. Wong,3 T.R. McNutt,4 H. Quon,3 L. Burns,6 and A. Thompson6; 1George Washington University (Graduate Student), Washington, DC, 2Johns Hopkins University School of Medicine, Baltimore, MD, 3Johns Hopkins University, Department of Radiation Oncology and Molecular Radiation Sciences, Baltimore, MD, 4 Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University School of Medicine, Baltimore, MD, 5Johns Hopkins University, Baltimore, MD, 6Johns Hopkins Hospital, BSN, Baltimore, MD Purpose/Objective(s): Head and neck cancer (HNC) patients often experience weight loss due to various radiation-related toxicities. Predicting weight loss can assist with interventions. The aim of this study was to evaluate the performance of various Machine Learning (ML) classification algorithms and determine dominant predictors for weight loss building on our prior weight loss prediction model using the Classification and Regression Trees (CART) algorithm. Materials/Methods: HNC patients receiving intensity modulated radiation therapy (RT) were queried from Oncospace. Oncospace aggregates prospective structured outcome data captured during routine clinical care with RT planning data also generated during the routine clinical workflow. Multiple ML algorithms available in the computer algorithm’s Classification Learner (CL): Decision Trees, Discriminant Analysis, Support Vector Machines, Nearest Neighbor Classifier, were used to develop prediction models for the primary end point of weight loss (5kg) at 3-month post RT, and compared to Ensemble Classifiers. The predicted response from a trained ensemble is the average of predictions from individual trees. Ten-fold cross-validation was used to protect data against over fitting. Results: From 2007-14, 326 patients with 729 variables collected during treatment and in follow up were analyzed. The primary outcome was a weight loss of (5kg), despite the use of a PEG tube. The predictors utilized with the CL were determined based on the best performance of various ‘weak learning’ ML classification algorithms (decision trees, Knearest neighbors, etc.), then compared to those with a high-quality ensemble model. The patient reported outcomes (PRO) such as “able to eat foods I like” were used with further expanding trees through Ensemble Bagged Trees and anatomic tumor location by ICD-9 resulting in a learning accuracy of 0.844. A comparison of weight loss of (5kg) with radiation dose and chemotherapy predictors resulted in 0.89 and 0.927 accuracies. Additionally, significant accuracies were also obtained for mucositis (0.862), xerostomia (0.865), T-stage (0.869), pain intensity (0.896), PEG used (0.890), and PRO “content with the quality of life” (0.902). Conclusion: ML algorithms were successfully applied modeling weight loss in our head and neck Oncospace informatics platform. Ensemble
2947 Comorbidity With Age as a Predictor of Survival for Patients With Nasopharyngeal Cancer Following Radiation Treatment: A Nationwide Population-Based Study C.C. Yang,1 P.C. Chen,2 L.C. Lin,1 S.L. Chang,3 and C.C. Lee4; 1 Department of Radiation Oncology, Chi-Mei Medical Center, Tainan, Taiwan, 2Department of Radiation Oncology, Pingtung Christian Hospital, Pingtung, Taiwan, 3Department of Otolaryngology, Chi-Mei Medical Center, Tainan, Taiwan, 4Department of Otolaryngology, Head and Neck Surgery, Kaohsiung Veterans General Hospital, Kaohsiung, Taiwan Purpose/Objective(s): To characterize the impact of comorbidity and age on survival outcomes for patients with nasopharyngeal carcinoma (NPC) post radiation therapy (RT). Materials/Methods: A total of 4095 patients with newly diagnosed NPC treated by RT or RT plus chemotherapy (CT) in the period from 2007 to 2011 were included through Taiwan’s National Health Insurance Research Database. Comorbidities present prior to the NPC diagnosis was obtained and adapted to the Charlson Comorbidity Index (CCI) and Age-Adjusted Charlson Comorbidity Index (ACCI). Overall survival was estimated using the KaplandMeier method and the difference between groups was analyzed using log-rank test. The Cox proportional hazards regression model was used for multivariate analysis. Differences between variables with categorical data were examined using the chi-square test. Receiver Operating Characteristic (ROC) curves was generated to assess the accuracy and the predictive ability of each index for survival. Results: Most of the patients (75%) were male (age 5113 years) and 2470 of them (60%) had at least one comorbid condition. The most common comorbid condition was diabetes mellitus. According to these two different comorbidity index (CCI and ACCI), higher scores were associated with worse overall survival (P< 0.001). The Receiver Operating Characteristic (ROC) curve was used to assess the discriminating ability of CCI and AACI and it demonstrated the predictive ability for mortality with the ACCI (0.693, 95% CI 0.670-0.715) was superior to that of the CCI (0.619, 95% CI 0.593-0.644). Conclusion: Comorbidities with age greatly influenced the clinical presentations, therapeutic interventions, and outcomes of patients with NPC post RT. Higher comorbidity index scores accurately was associated with worse survival. The ACCI seems to be a more appropriate prognostic indicator and should be considered in further clinical studies. Author Disclosure: C. Yang: None. P. Chen: None. L. Lin: None. S. Chang: None. C. Lee: None.
2948 Adjuvant Radiation Therapy Improved Local Control for Primary Mucosal Melanoma of the Head and Neck M. Abugideiri,1 K. Patel,2 J.M. Switchenko,3 K. Magliocca,4 Z.S. Buchwald,1 J. Delgaudio,5 D.H. Lawson,6 H. Danish,2 J.J. Beitler,2,5 and M.K. Khan2; 1Department of Radiation Oncology at Emory University, Atlanta, GA, 2Department of Radiation Oncology, Winship Cancer Institute at Emory University, Atlanta, GA, 3Department of Biostatistics and Bioinformatics, Winship Cancer Institute at Emory