ESTRO 38 Abstract book

S1032 ESTRO 38

A web-based interface was then created to incorporate this model and make it accessible to physicians. A Java service was built at the back-end of the interface to allow direct communication between the model and the EHR system (HiX, Chipsoft, The Netherlands). Communication was performed by SOAP (Simple Object Access Protocol) queries through a business intelligence software suite (BusinessObjects) which acted as an intermediary. This allowed for automated filling of the input parameters and subsequent calculations. Results The internally validated AUC was 0.71 and provides the probability of surviving 26 weeks after the start of PCI radiation. Input parameters included the tumour N- and M-staging, gender, and concentrations of lactate dehydrogenase (LDH) and haemoglobin (HB) in the blood upon diagnosis. A web-based interface was created in which the model was plugged in. The architecture of the interface was set up in such a way that other models could be plugged in easily. Because of the direct communication with the HiX EHR system, real-time retrieval of the most recent information was possible so that the model input parameters are up-to-date at any time. Additionally, even though the input parameters are filled in automatically, clinicians are able to adjust all input parameters, after which the model instantly recalculates the outcome. Finally, accessibility of the model was improved by adding it as an option inside the HiX EHR system, allowing physicians to access the model directly from within the system that they use during the consultation. The interface is currently being tested in a prospective clinical trial. Fig 1: Screenshot of the interface inside the HiX EHR system

Material and Methods 94 patients with primary lung tumor treated with SBRT from September 2010 to December 2016 were retrospectively analysed (mean follow-up time= 2.6 years). All patients were treated with a 3DCRT technique and a dose prescription of 60 Gy in 3, 5 or 8 fractions. Three dosimetric/radiobiological variables of the PTV, ITV and GTV were used in the analysis (BED 2 , BED mean , BED 98 , with α/β=10) as well as their volumes. Treatment outcomes analysed were local recurrence (LR, 13 cases), nodal recurrence (NR, 13 cases), distant recurrence (DR, 33 cases) and death from disease (DD, 23 cases). Three different machine learning techniques were used in this work: logistic regression, linear Support Vector Machines (SVM) and decision trees (DT). Validation of the machine learning techniques were performed with 10-fold cross validation due to the limited data available, while Lasso regularization was employed for feature selection. The objective was to use no more than two variables for the model. In all outcomes but DR, Synthetic Minority Over-sampling Technique (SMOTE) was employed. This technique creates new instances of positive cases from real ones in order to balance the number of positive and negative cases. Results Areas under receiving operating characteristics curve (AUC) of different outcomes with the three methods are shown in Table 1. Best results of AUC are highlighted. For LR, the features used where PTV volume and GTV BED mean and the best method was logistic regression with AUC close to 0.75. In case of NR, selected features were ITV volume and GTV BED mean , and the method with a best performance was again logistic regression. In the case of DR, none of the methods were able to predict it, as values of AUC close to 0.5 are obtained. Finally, the best AUC obtained for DD was 0.67 for both Decision Trees and SVM using PTV volume and PTV BED 2 . In this case, we can see that the method clearly benefits from using SMOTE.

For those outcomes where we found AUCs>0.7, we will be able to develop models for treatment outcome probabilities. In this case, we developed models for LR and NR with the method with a better performance, which is logistic regression. To do this, we trained the model with the whole dataset using the hyperparameters and features obtained during cross validation. Models are shown in Table 2.

Conclusion The interface described in this abstract allows for automated generation of model outcomes at a moment’s notice with up-to-date information and within the EHR system. By increasing the usability of a model using this interface, clinicians are encouraged even more to utilize the added benefit that predictive models can bring in routine clinical practice. EP-1899 Prediction of outcomes in lung SBRT with dosimetric variables and machine learning techniques D. Sevillano 1 , C. Martín 2 , C. Vallejo 3 , M. Martín 3 , R. Colmenares 1 , R. Morís 1 , B. Capuz 1 , J.D. García 1 , M. Cámara 1 , A. Martínez 1 , F. Orozco 1 , M.J. Béjar 1 , D. Prieto 1 , S. Sancho 3 , F. García-Vicente 1 1 Hopital Universitario Ramón y Cajal, Medical Physics, Madrid, Spain ; 2 ETSIT. Universidad Politécnica de Madrid, Biomedical Engineering, Madrid, Spain ; 3 Hopital Universitario Ramón y Cajal, Radiation Oncology, Madrid, Spain Purpose or Objective To find models based on dosimetric variables that allow prediction of outcomes in primary lung tumor SBRT.

Conclusion Machine learning techniques permit to obtain fair predictions of LR and NR in our clinical practice. Furthermore, those models yield reasonable conclusions regarding tumor volume (volume of PTV and ITV) and dose prescription (GTV BED mean ). Logistic regression was the most accurate method, being also the one with a closest relationship to recurrence probability. These results should be considered as preliminary considering that only cross validation was performed, the