Abstract 896P
Background
Surufatinib shows good efficacy in treating advanced NET. To better assist therapeutic decision-making and improve outcomes, candidates who may benefit from surufatinib need to be identified in advance. This study aimed to seek novel imaging biomarkers for stratifying treatment response by developing a radiomics-based machine learning model.
Methods
Imaging features were extracted from liver lesions in portal phase of pretreatment abdominal contrast-enhanced CT. Two experienced radiologists independently delineated the lesions. Baseline clinical variables and the overall response (RECIST 1.1) were added into model training. Patients were divided into responder and stable disease (SD) groups. Responder was defined as target lesion size reduced by over 15%, and SD as target lesion size changed within -15%∼20%. Data were randomly assigned in 8:2 as training set and validation set. Features were selected by ElasticNet. The model was comprised of two-layer classifiers. The first-layer classifier (FLC) was composed of random forest, support vector machine, and Bayes algorithms, followed by logistic regression as the second-layer classifier (SLC). The FLC was validated using 5-fold cross-validation. The output from FLC, the predicted probabilities of each cross-validation set, was then used as input to train SLC for predicting responder. The performance of the model was evaluated in validation set with area under the curve (AUC), specificity, sensitivity, and accuracy.
Results
A total of 217 patients in 3 trials (NCT02267967, NCT02588170, NCT02589821) were included (responder vs SD: 95 vs 122). Among 55 radiomics features and 6 clinical factors, ten variables were selected, including coarseness, compactness, skewness, kurtosis, busyness, long-run matrix homogeneousness, cumulative intensity, feature homogeneity and mean HU value of lesion, diagnosis-to-medication time. The model performance achieved an accuracy of 68%, with an AUC of 0.67, sensitivity of 63%, and specificity of 71%.
Conclusions
This preliminary result demonstrated the radiomics-based machine learning model could predict treatment responses in patients with advanced NET.
Clinical trial identification
NCT02267967, release date March 4, 2019; NCT02588170, release date September 20, 2020; NCT02589821, release date September 20, 2020.
Editorial acknowledgement
Legal entity responsible for the study
Hutchison MediPharma.
Funding
Hutchison MediPharma.
Disclosure
J. Xu: Financial Interests, Personal and Institutional, Principal Investigator: The Fifth Medical Center of Chinese PLA General Hospital. C. Zhao, J. Zhou, X. Luo, S. Fan, W. Su: Financial Interests, Institutional, Sponsor/Funding: Hutchison MediPharma. K. Nie, C. Lin, J. Yang: Financial Interests, Institutional, Funding: Taimei Medical Technology.