ESTRO 2020 Abstract book

S251 ESTRO 2020

To develop normal tissue complication probability (NTCP) model for predicting risk of radiation-induced liver disease (RILD) using clinical and dosimetric factors and to assess

Integrating clinical and dosimetric factors and assessing uncertainty lead to reliable risk estimation.

for model uncertainty. Material and Methods

PD-0427 Optimization of preselection process in model-based selection for proton in head and neck cancer M. Tambas 1 , J.G.M. Van den Hoek 1 , R.G.J. Kierkels 1 , D. Scandurra 1 , A. Wolters 1 , D. Mulder 1 , E. Oldehinkel 1 , T.W.H. Meijer 1 , S. Both 1 , R.J.H.M. Steenbakkers 1 , J.A. Langendijk 1 1 University of Groningen- University Medical Center Groningen, Radiation Oncology, Groningen, The Netherlands Purpose or Objective In the Netherlands, head and neck cancer (HNC) patients qualify for proton therapy based on model-based selection using NTCP-models to assess ΔNTCP (expected difference in toxicity rates between photons and protons), which requires a plan comparison. However, performing planning comparisons in all patients is labor intensive. Therefore, tools to identify patients that are more likely to be selected for protons are needed. The aim of this study was to develop an advanced preselection tool (PST) to reduce unnecessary plan comparisons. Material and Methods According to Dutch National Indication Protocol, NTCP- models for grade≥2 dysphagia and xerostomia, and tube feeding dependence are used for the model-based selection. The ΔNTCP-thresholds are ≥10% for grade ≥2, ≥5% for grade ≥3, and/or ≥15% for summed risk reduction (ΣΔNTCP) of grade 2 toxicities. If the NTCP value of photon plan exceeds any of these ΔNTCP thresholds (basic PST), then plan comparison is indicated and proton plan is created. For the development of advanced PST, an OAR_Overlap structure (OAR ∩ PTV+5 mm) was created for all five OARs within the NTCP models. Using the percentage of OAR_Overlap as predictor, D mean of OAR for proton was predicted witha linear regression model. The estimated NTCP values were calculated using the predicted OAR D mean values and compared with those from photon plan. If estimated ΔNTCP exceeded thresholds, a plan comparison was indicated (Figure 1). Then, the positive and negative predictive value (PPV and NPV) of PST was investigated. The primary endpoint was: patient qualifies for proton therapy. Results Using basic PST, 141 patients with HNC who were eligible for plan comparison were included in the study. Thus, NPV of the basic PST was 100%. Eventually, 81 patients qualified for proton, so PPV was 57%. However, in 60 patients (43%), plan comparisons could potentially have been avoided. Using advanced PST, unnecessary plan comparisons were avoided in 47(78%) of these 60 patients, whereas 24 patients were wrongfully denied (i.e. false negative). So the PPV and NPV of advanced PST were 82% and 66%, respectively. As this was considered unethical, the predicted OAR proton dose was reduced by a factor to ensure that no patient was denied a plan comparison based on the result of PST, who eventually qualified for protons. In our patients cohort, the optimal reduction factor was 1.2. The optimized predicted proton dose with this reduction factor resulted in prevention of 30(50%) unnecessary plan comparisons with no false negatives. The sensitivity and specificity of the adjusted new PST were 100% and 50%, while the PPV and NPV values were 73% and 100%, respectively.

We retrospectively collected data of primary liver cancer patients who were treated with radiotherapy: hepatocellular carcinoma (HCC) (n=215) and intrahepatic cholangiocarcinoma (n=107). A total of 322 patients were randomly assigned into training and test dataset which were used for model development and validation, respectively. Study endpoint was RILD. Normal liver dosimetry was reduced to equivalent uniform dose and normalized using reference fraction size=2 Gy/fraction, volume effect=1, and α/β ratio=2 Gy, referred as mean liver dose (MLD). The most predictive variables were selected using a multivariate logistic regression analysis with bootstrapping method. Model performance and validation were evaluated. Subsequently, Monte Carlo simulation method was performed to assess model uncertainty by using multivariate normal distribution of regression coefficients and covariance matrix of model parameters. Results The incidence of RILD was 33.5% (108/322 patients), 32.3% in training dataset and 38.5% in test dataset (p=0.35). In multivariate analysis, diagnosis of HCC, Child-Pugh class B or C, positive hepatitis status and MLD were significant predictors for RILD. The NTCP value for the individual patient can be calculated using the equation: NTCP = 1/[1+e^(-s)], in which s= -3.33 + [diagnosis of HCC (yes=1, no=0)* 1.44] + [CP-B or C (yes=1, no=0)* 1.04] + [viral hepatitis infection status (yes=1, no=0)* 0.77] + [MLD (in Gy) * 0.04]. Model performance was good with an area under the receiver operative characteristics curve of 0.77 and 0.86 in the training and test dataset, respectively. For subgroups of patients stratified by clinical factors, NTCP curves with uncertainty (95% confidence interval) were illustrated (figure 1).

Conclusion The multivariable NTCP model has been developed for predicting risk of RILD for primary liver cancer patients.