1113P - A machine learning model based on computed tomography radiomics to predict prognosis in subjects with stage IV melanoma

Background

Biomarkers and clinical features don’t currently enable the identification of standardized risk categories for optimal treatment strategies in metastatic melanoma. This issue underlines the need of a more sophisticated and comprehensive prognostic evaluation. The aim of this retrospective observational study is to develop a machine learning model based on pre-therapy Computed Tomography (CT) images to stratify the single-subject prognosis in melanoma patients.

Methods

Images from 60 metastatic lesions were collected, 32 (53.3%) belonged to “favorable prognosis” class and 28 (46.7%) to “unfavorable prognosis” class, according to patients’ prognosis intended as Progression Free Survival (PFS) treatment median PFS. This image-set was used for the training and cross-validation of different radiomic-machine learning models through the Trace4Research software (DeepTrace Technologies srl, Italy). A radiomic approach was applied, under the hypothesis that radiomic feature could capture the disease heterogeneity among the two groups. Three models consisting of 4 ensembles of machine learning classifiers (random forests, support vector machines and k-nearest neighbor classifiers) were developed for the binary classification task of interest (favorable vs unfavorable), based on supervised learning, using prognosis as reference standard.

Results

The best model showed ROC-AUC (%) of 82 (majority vote), 81.6** (mean) [77.9-85.4], Accuracy (%) of 77, 75.4** [74.1-76.7], Sensitivity (%) of 84, 80.5** [78-83], Specificity (%) of 68, 69.6** [66.4-72.9], PPV (%) of 75, 75.2** [73.5-76.9], and NPV (%) of 79, 75.7** [73.8-77.7] (*p<0.05, **p<0.005).

The model was external tested on 20 new patients (N=70 lesions) and the classification of each patient’s prognosis was obtained using the one most frequently assigned by the classifier to the metastatic lesions of the same patient. The results show that the classifier can predict subjects with a favorable prognosis with good accuracy (85%). A third of patients (35%) with unfavorable prognosis were predicted.

Conclusions

These preliminary data underscore the potential of radiomics-based machine learning models in predicting prognosis in patients with metastatic melanoma.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

M. Russillo: Financial Interests, Institutional, Financially compensated role: Pierre Fabre Oncologie, Novartis, MSD, BMS. V. Ferraresi: Financial Interests, Institutional, Financially compensated role: BMS, Novartis, Pierre Fabre Oncologie, MSD. All other authors have declared no conflicts of interest.