AI04 A novel machine learning software application for prognosticating cutaneous malignant melanoma based on data from 78 351 patients from the Surveillance, Epidemiology, and End Results (SEER) database

Abstract Cutaneous malignant melanoma (CMM) is one of the most common cancers in incidence and mortality. Artificial intelligence and machine learning (ML) are useful in developing cancer-related applications. Current prognosticating systems for CMM, such as tumour–nodes–metastasis (TNM) staging, lack some of the influential survival-related factors. This study aims to develop an ML software application that improves CMM prognostication. Patients diagnosed with CMM during 2011–2021 were extracted from SEER and were split into training and testing sets (80 : 20 ratio) after excluding patients with missing data. The Survival eXtreme Gradient Boosting (XGBoost) ML model was used to explore the variables affecting their survival, which were visualized using Shapley Additive exPlanations (SHAP). These variables included age, sex, ethnicity, income, rurality, site, histology, ulceration, Breslow thickness, mitotic rate, number of positive lymph nodes, metastasis type and TNM stage. The concordance index (C-index) was calculated for this model (model 1) and compared with a model using TNM stage only (model 2) in determining prognosis. Model 1 was used to predict risk scores for all of the patients, which were grouped into five categories (0–20, 20–40, 40–60, 60–80 and 80–100 percentiles). In total 78 351 patients with CMM were included in the analysis. The mean age, Breslow thickness and number of positive lymph nodes were 59.2 years, 1.3 mm and 0.1, respectively. Most patients were White (91.0%), male (54.4%), of average income (33.6%) and living in metropolitan areas (88.5%). Only 13% had ulcerated lesions and the most common site was the trunk (33.8%). The most common histology and stage were melanoma not otherwise specified (41.9%) and stage I (76.1%), respectively. SHAP showed that the five most important factors affecting survival were Breslow thickness, age, income, stage and histology. ML model 1 had C-indexes of 81% and 80% for the training and testing sets, respectively. ML model 2 had C-indexes of 66.2% and 66.6% for the training and testing sets, respectively. A Kaplan–Meier plot showed that the five survival categories of patients do not overlap. A software application was developed using model 1 that risk stratifies patients into one of the five categories after inputting the 13 clinical and demographic variables, and shows the Kaplan–Meier plot for these categories. There are important tumour-related and demographic factors that affect the prognosis of patients with CMM, some of which are not used in the current TNM staging system. ML could be used to incorporate these factors into a software application that guides clinicians to prognosticate patients in a more personalized way. Further research is required to validate this application on other demographic groups and healthcare systems and improve this clinical application tool, given the limitations of the SEER database.