ESTRO 2021 Abstract Book

S1067

ESTRO 2021

Twelve patients were treated with curative intent.Five (11% of all patients with positive results, 42% of all treated with curative intent) were disease-free and alive at least one year after the procedure (median 26 months, maximal 107 months). Three patients treated with curative pelvic surgery did not have malignancy confirmed on pathological examination. Conclusion Follow-up with PET-CT results in actionable insights. In the majority of patients, findings lead to further diagnostic or therapeutic interventions. There is evidence, that in some patients, this may have an impact on overall survival. We urgently need an evaluation of the proposed strategies in a prospective setting. PO-1297 Machine Learning Identifies Predicators of Disease-Free Survival in Patients with Cervical Cancer E. Yoshida 1 , W. Arbelo Gonzalez 1 , G. Valdes 1 , E. Hirata 1 , O. Morin 1 , I. Hsu 1 1 University of California San Francisco, Radiation Oncology, San Francisco, USA Purpose or Objective We hypothesize that by using machine learning statistical methods, novel pretreatment predictors for disease- free survival may be identified from our matured dataset of patients treated with definitive chemoradiotherapy for cervical cancer. Materials and Methods Patient records of all cervical cancer patients treated with definitive chemoradiotherapy from 2003 to 2014 by a single provider were retrospectively reviewed under IRB approval. Clinical data from initial history and physical exam coded for this study included: AJCC staging, tumor size, number of lymph nodes > 1 cm in short axis, location of enlarged nodes, smoking status, substance use, parity, patient-reported weight change, and number of prescription medications. A total of 283 patients were included in the study with a median follow up time of 55.1 months. The dataset was split into two cohorts: a training cohort of 226 patients, and a testing cohort of 57 patients. A multivariate Cox regression analysis, with Lasso regularization to provide feature selection, was performed to predict the hazard ratios. An internal 5-fold cross-validation was performed on the training dataset to determine the regularization parameter, which was used to re-train this dataset. For each patient, a partial hazard (risk) was predicted using the Cox regression obtained. The training cohort was then recursively partitioned into two groups according to the value of the hazard ratio and a log- rank test was performed with Kaplan Meier survival analysis corresponding to each group. The cutoff value of the predicted hazard ratio that separated two populations with the most divergent Kaplan-Meier curves was determined. Results Concordance indexes of 0.74 and 0.75 were obtained in the training and testing datasets, respectively. Stage and weight gain were found to be significantly associated with disease-free survival (Figure 1). The effect of metastasis on disease-free survival nearly reached significance. A partial hazard of 1.15 was determined as the threshold that yielded the most disparate (lowest p-value) survival curves for the two groups in the training dataset. This threshold also divided the patients in the testing cohort into two distinct groups with divergent survival outcomes. Incorporation of weight gain and metastasis into our predictive model significantly improved the discriminatory ability of stage to identify higher- versus lower-risk patients. This improvement can be estimated by the degree of separation between the higher-risk and lower-risk survival curves (Figure 2), and the considerable difference in p-values from the log-rank test (1.176 x 10 -3 versus 8.973 x 10 -7 ).