Machine Learning-based Prognostic Model for Brain Metastasis Patients: Insights from Blood Test Analysis

Purpose: Brain metastases, affecting 30% of solid tumor patients, have a substantial impact on clinical outcomes. Developing a clinically feasible and precise prognostic model is crucial for personalized and comprehensive treatment. Methods: Parameters from blood test were collected from brain metastases patients, and were used to construct the four models, including univariate Cox regression, stepwise regression, LASSO regression, and random survival forest (RSF). Model-HP (based RSF), identified as the best-performing, was chosen. Model-GPAH was formed by merging Model-HP risk scores and GPA (Graded Prognostic Assessment). AUC, IDI, and cNRI were used to evaluate different models. Results: A cohort of 1,385 patients was included, with 970 patients assigned to the training cohort and 415 patients were to the validation cohort. Compared to the other models, the Model-HP built on the RSF demonstrated superior performance (compared with RSF: AUC = 0.71 [0.66, 0.77], Univariate Cox regression: AUC = 0.65 [0.59, 0.71], P = 0.011; Stepwise regression: AUC = 0.63 [0.57, 0.69], P = 0.001; LASSO regression: AUC = 0.64 [0.58, 0.70], P < 0.001). Compared with Model-HP and GPA, Model-GPAH significantly enhanced the performance of prognosis prediction (compared with Model-GPAH: AUC = 0.70 [0.67, 0.73], GPA: AUC = 0.61 [0.57, 0.64], P = 0.001; Model-HP: AUC = 0.67 [0.64, 0.70], P < 0.001). Model-GPAH performed favorably across patients receiving diverse treatments. Conclusions: Integrating hematological parameters into the GPA model significantly enhanced prognostic prediction for brain metastasis patients, highlighting blood tests' crucial role in identifying biomarkers for outcomes.