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

Area of Science:

• Oncology
• Biostatistics
• Clinical Research

Background:

• Brain metastases affect 30% of solid tumor patients, significantly impacting clinical outcomes.
• Accurate prognostic models are essential for personalized treatment strategies in brain metastasis patients.
• Existing prognostic tools may benefit from integration of readily available clinical data.

Purpose of the Study:

• To develop and validate a precise prognostic model for patients with brain metastases.
• To evaluate the utility of hematological parameters in enhancing prognostic accuracy.
• To compare the performance of different statistical modeling techniques for survival prediction.

Main Methods:

• Hematological parameters were collected from 1,385 brain metastasis patients (970 training, 415 validation).
• Four models were constructed: univariate Cox regression, stepwise regression, LASSO regression, and random survival forest (RSF).
• The best-performing model (Model-HP, based on RSF) was merged with the Graded Prognostic Assessment (GPA) to form Model-GPAH. Model performance was assessed using AUC, IDI, and cNRI.

Main Results:

• Model-HP (RSF-based) demonstrated superior performance compared to univariate Cox, stepwise, and LASSO regression models (AUC 0.71 vs. 0.65, 0.63, 0.64 respectively).
• Model-GPAH significantly improved prognostic prediction compared to Model-HP and GPA alone (AUC 0.70 vs. 0.67 and 0.61 respectively).
• Model-GPAH showed consistent performance across patients receiving diverse treatment modalities.

Conclusions:

• Integrating hematological parameters into the GPA model substantially enhances prognostic prediction for brain metastasis patients.
• Blood tests serve as a valuable source of biomarkers for predicting outcomes in oncology.
• The developed Model-GPAH offers a more precise and clinically feasible tool for personalized patient management.