Machine learning-based survival prediction tool for adrenocortical carcinoma

Context: Adrenocortical carcinoma (ACC) is a rare, aggressive malignancy with difficult to predict clinical outcomes. The S-GRAS score combines clinical and histopathological variables (tumor stage, grade, resection status, age, and symptoms) and showed good prognostic performance for patients with ACC. Objective: To improve ACC prognostic classification by applying robust machine learning (ML) models. Method: We developed ML models to enhance outcome prediction using the published S-GRAS dataset (n = 942) as the training cohort and an independent dataset (n = 152) for validation. Sixteen ML models were constructed based on individual clinical variables. The best-performing models were used to develop a web-based tool for individualized risk prediction. Results: Quadratic Discriminant Analysis, Light Gradient Boosting Machine, and AdaBoost Classifier models exhibited the highest performance, predicting 5-year overall mortality (OM), and 1-year and 3-year disease progression (DP) with F1 scores of 0.79, 0.63, and 0.83 in the training cohort, and 0.72, 0.60, and 0.83 in the validation cohort. Sensitivity and specificity for 5-year OM were at 77% and 77% in the training cohort, and 65% and 81% in the validation cohort, respectively. A web-based tool (https://acc-survival.streamlit.app) was developed for easily applicable and individualized risk prediction of mortality and disease progression. Conclusion: S-GRAS parameters can efficiently predict outcome in patients with ACC, even using a robust ML model approach. Our web app instantly estimates the mortality and disease progression for patients with ACC, representing an accessible tool to drive personalized management decisions in clinical practice.