Optimal Machine Learning Models for Developing Prognostic Predictions in Patients With Advanced Cancer

Context: Accurate prognosis prediction for cancer patients in palliative care is critical for clinical decision-making and personalized care. Traditional statistical models have been complemented by machine learning approaches; however, their comparative effectiveness remains underexplored.

Objectives: To assess the prognostic accuracy of statistical and machine learning models in predicting 30-day survival in patients with advanced cancer using objective data, such as the result of the blood test.

Methods: A secondary analysis of the Japan-Prognostic Assessment Tools Validation (J-ProVal) study was performed from September 2012 to April 2014. We used data from 58 palliative care services in Japan and enrolled 915 patients. Four models, fractional polynomial (FP) regression, Kernel Fisher discriminant analysis (KFDA), Kernel support vector machine (KSVM), and XGBoost, were compared using 17 objective clinical characteristics. Models were evaluated with the area under the receiver operating characteristic curve (AUC) as the primary metric.

Results: The KSVM model demonstrated the highest predictive accuracy (AUC: 0.834), outperforming the FP model (AUC: 0.799). XGBoost showed comparatively lower performance; however, it was likely limited by the size of the dataset.

Conclusions: Machine learning, particularly KSVM, has high predictive accuracy in palliative care when sufficient data are available. However, our findings suggest that traditional statistical models offer advantages in stability and interpretability, underscoring the importance of tailored model selection based on data characteristics.