Artificial intelligence improves survival prediction in patients with brain metastases submitted to radiosurgery

Stereotactic radiosurgery (SRS) is effective for non-small cell lung cancer (NSCLC) brain metastases in deep or eloquent brain regions. Identifying predictors of treatment failure is crucial. Artificial intelligence (AI) models may improve prediction, but data on NSCLC BM are scarce. A retrospective study analyzed NSCLC patients with single brain metastases treated with SRS (Elekta Gamma Knife) from 2010 to 2015, with up to 10 years of follow-up. Clinical, radiological, and histological data were collected. Kaplan-Meier and Cox proportional hazards models assessed survival. Decision tree and random forest (RF) models predicted treatment failure, with feature importance analyzed. Among 133 patients (mean age 61.6, 56.4% male), most tumors were grade 1 (56.4%) and in the right hemisphere (60.2%). The mean tumor volume was 1.84 cm³. Decision trees identified metastasis volume and location as key predictors (AUC = 0.85). RF models improved prediction (AUC = 0.92). Tumor volume, diameter, and age were major predictors. AI models effectively identified patients at risk of treatment failure. A random forest based artificial intelligence model presented an excellent predictive ability for stereotactic radiosurgery success/failure in a population with NSCLC brain metastases, with an area under curve of 0.92. This predictive ability was superior to a decision tree or a simple diameter-to-volume ratio.