Machine learning models to predict the response to anti-cancer therapy in metastatic melanoma patients.

Abstract

e14071 Background: Machine learning methods are new artificial intelligence tools with promising applications in healthcare. We developed and validated 4 machine learning models to predict the response to immunotherapy and targeted therapy in stage IIIc or IV melanoma patients. Methods: This work was conducted on data from 10 centers participating in the French network for Research and Clinical Investigation on Melanoma (RIC-Mel), launched in 2012. Thus, 935 patients, corresponding to 1978 systemic treatments have been extracted from RIC-Mel database. The following data were considered: age, sex, Breslow, melanoma type, ulceration, spontaneous regression, mitotic index, number of invaded lymph nodes, extracapsular extension, mutational status, melanoma stage, number of metastasis sites, lines of treatments, and time between first melanoma excision and metastatic relapse. Treatment response: complete response, partial response, stable disease, defined as class 1 and progressive disease as class 2. We split this cohort/database into a training set (80%) and test set (20%). The algorithm performances were evaluated on the test set by the percentage of treatments correctly classified in class 1 or 2. Four machine learning algorithms (linear model, random forest, XGBoost and LightGBM) were compared in terms of performance and interpretation for both types of treatments. Results: The accuracies of the best models for immunotherapy (LightGBM) and targeted therapy (random forest) were respectively 66% and 65%. The most significant variables for building the models were respectively: stage (IIIc or IV), response to previous treatments lines, age, number of metastasis sites and time between first melanoma excision and metastatic relapse. Conclusions: We present here the first machine learning models to predict the response to immunotherapy and targeted therapy in stage IIIc or IV melanoma patients. The most predictive variables are coherent with the literature. Future development will include data from 18FDG-PET/CT imaging and other predictive markers recently identified, as circulating DNA to improve the models performance.