OC-0257: A Bayesian network model for acute dysphagia prediction in the clinic for NSCLC patients

OC-0257A Bayesian network model for acute dysphagia predictionin the clinic for NSCLC patientsA.T.C. Jochems1MAASTRO clinic, Radiotherapy, Maastricht, The Netherlands1, T.M. Deist1, E. Troost2, A. Dekker1, C.Faivre-Finn3, C. Oberije-Dehing1, P. Lambin12Helmholtz-Zentrum, Radiooncology, Dresden-Rossendorf,Germany3The Christie NHS Foundation Trust & University ofManchester, Radiation Oncology, Manchester, UnitedKingdomPurpose or Objective: Acute dysphagia is a frequentlyobserved toxicity during concurrent chemo-radiation (CRT) orhigh-dose radiotherapy (RT) for lung cancer. This toxicity canlead to hospitalizations, treatment interruptions and ESTRO 35 2016 S119______________________________________________________________________________________________________consequently reduce chances of survival. Models to predictacute dysphagia are available. However, these models werebased on limited amounts of data and the performance ofthese models needs improvements before implementationinto routine practice. Furthermore, Bayesian network modelsare shown to perform better than conventional modelingtechniques on datasets with missing values, which is acommon problem in routine clinical care. In this work, wetrain a Bayesian network model on a large clinical datasets,originating predominantly from routine clinical care, toaccurately predict acute dysphagia in NSCLC patients duringand shortly after (C)RT.Material and Methods: Clinical data from 1250 inoperableNSCLC patients, treated with radical CRT, sequential chemoradiation or RT alone were collected. The esophagus wasdelineated using the external esophageal contour from thecricoid cartilage to the GE junction. A Bayesian networkmodel was developed to predict severe acute dysphagia ( ≥Grade 3 according to the CTCAEv3.0 or v4.0). The modelutilized age, mean esophageal dose, timing of chemotherapyand N-stage to make predictions. Variable selection andstructure learning was done using the PC-algorithm. Themodel was trained on data from 1250 patients. The model’sperformance was assessed internally and on an externalvalidation set (N=218) from the United Kingdom. Modeldiscriminative performance was expressed as the Area Underthe Curve (AUC) of the Receiver Operating Characteristic(ROC). ROCs were compared using the method proposed byDeLong and colleagues. Model performance was also assessedin terms of calibration. Calibration refers to the agreementbetween the observed frequencies and the predictedprobabilities and is expressed as the coefficient ofdetermination (r2).Results: One-hundred forty patients (11,2%) developed acutedysphagia (≥ Grade 3 according to the CTCAEv3.0 or v4.0).The model was first validated internally, by validating on thetraining cohort (N=1250, AUC = 0.77, 95% CI: 0.7325-0.8086,r2 = 0.99). Subsequently, the model was externally validatedon a UK dataset (N = 218, AUC = 0.81, 95% CI: 0.74-0.88, r2 =0.64). The ROC curves were not significantly different (p =0.28).Conclusion: The Bayesian network model can make accuratepredictions of acute dysphagia (AUC = 0.77, 0.81 in theinternal and external validation respectively), making it apowerful tool for clinical decision support.