Abstract Book

ESTRO 37

S391

Sydney, Australia 8 MAASTRO Clinic, Department of Radiation Oncology, Maastricht, The Netherlands Purpose or Objective Clinical decision support systems (DSS) combine multiple variables in a statistical analysis to predict treatment outcomes from a particular treatment. Radiotherapy (RT) treatment of non-small cell lung cancer (NSCLC) is variable in practice, with different proportions of curative versus palliative RT use. We developed an overall survival DSS model based on data from patients treated with curative RT in two institutions using a distributed learning algorithm, whereby data remains at each centre. We aim to assess the feasibility of building DSSs for NSCLC radiotherapy utilizing data obtained from routine practice at multiple institutions. Material and Methods Clinical information and RT planning computed tomography (CT) imaging data for 607 patients with inoperable, Stage I-III NSCLC treated with RT between 2003 and 2017 were compiled locally at two cancer institutions. There were 511 NSCLC patients at Institute 1 and 96 at Institute 2. Of these patients 315 received curative RT (dose > 48Gy). A support vector machine (SVM) model for predicting two-year survival was trained across the two institutions using distributed learning software on a randomly-selected half of the cohort. Assignment to one of three risk groups was adjusted towards providing treatment selection decision support between curative and palliative treatment based on the training cohort, where based on past experience from this cohort, future patients falling in the predicted high- risk group might be managed with non-curative intent. The model was tested on the remaining half of the cohort. The attributes used in the model were age, gender, performance status, tumour volume and lung function.

been eligible for the clinical trials that inform standard treatment guidelines. References [1] Dehing-Oberije C. et al. doi:10.1016/j.ijrobp.2008.08.052

Poster: Clinical track: Upper GI (oesophagus, stomach, pancreas, liver)

PO-0758 favorable vs unfavorable prognosis by post- CRT PET scan in cN+ esophageal SCC treated with dCRT W.K. Yap 1 , T.M. Hung 1 1 ChangGung Memorial Hospital, Radiation Oncology, Taoyuan City, Taiwan Purpose or Objective To examine the prognostic value of post-chemoradiation PET scan based on the presence or absence of FDG-avid metastatic lymph node(s) and metabolic response of the primary tumor in patients with clinically node-positive esophageal squamous cell carcinoma (cN+ ESCC) treated with definitive chemoradiotherapy (dCRT) Material and Methods We identified 108 eligible patients treated by chemoradiotherapy (CRT) with or without resection from our prospectively collected database. Absence of FDG- avid metastatic lymph node with at least partial response of the primary tumor on PET scan after initial CRT was defined as Post-CRT PET favorable group (yPET-F), and otherwise as unfavorable group (yPET-U). Kaplan-Meier method and Cox regression were performed for survival analyses and multivariable analysis, respectively. Results The study cohort comprised of 59 patients receiving dCRT. Forty-five patients receiving trimodality therapy (TMT) comprised the comparative group and 4 patients excluded from further analyses for developing interval distant metastasis detected on post-CRT PET scan. The median follow-up for the study cohort was 41 months. On K-M analysis of the study cohort, yPET-F was found to have significant better OS (2-year: 72.5% vs 13.7%, p < 0.01) and DMFS (2-year: 71.6% vs 36.6%, p = 0.01) than yPET-U. In multivariable analysis, yPET-F remained as a strong independent favorable prognosticator on both OS (HR 0.08, p < 0.01) and DMFS (HR 0.14, p = 0.02) for dCRT cohort. Comparing with TMT cohort, for yPET-U patients, TMT had better OS (p = 0.03) than dCRT-Operable and dCRT-Operable had superior OS (p = 0.04) than dCRT- Unresectable. For yPET-F patients, there was no difference in both OS (p > 0.99) and DMFS (p = 0.92) between these three groups

Results The SVM model had an AUC of 0.7 on the withheld data from both institutes. A survival analysis for low and high risk groups is displayed in Figure 1, now also including the past patients who received palliative treatment. Here, 17% (n=26) of patients treated with palliative RT resided in the predicted low risk group and observed significantly reduced survival (logrank, p<0.01) over curative treatment. There is the potential for better outcomes if similar patients are treated with curative RT in the future. Alternatively, 21% (n=39) of curatively treated patients belonged to the predicted high risk group and did not observe increased survival with curative RT (logrank, p=0.6). The model published in [1] had an AUC of 0.66 in our cohort, and according to the nomogram assigned risk a large percentage of patients (63%) were considered high risk compared to 29% according to the reported DSS. Conclusion The DSS exhibited improved performance in comparison to previous models particularly in terms of the calibration of risk groups. This is critical to the generalizability of a DSS. We believe developing models with this approach will provide DSS tools for patients who would not have