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

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Accurate prognosis is vital for cancer patient care, especially in palliative settings.
• Machine learning (ML) models are increasingly used, but their comparison with traditional statistical models for cancer prognosis is underexplored.

Purpose of the Study:

• To compare the prognostic accuracy of statistical and ML models for predicting 30-day survival in advanced cancer patients.
• To evaluate model performance using objective clinical data, including blood test results.

Main Methods:

• Secondary analysis of the Japan-Prognostic Assessment Tools Validation (J-ProVal) study (2012-2014).
• Included 915 patients from 58 palliative care services in Japan.
• Compared four models: fractional polynomial (FP) regression, Kernel Fisher discriminant analysis (KFDA), Kernel support vector machine (KSVM), and XGBoost, using 17 objective clinical characteristics.
• Primary evaluation metric was the area under the receiver operating characteristic curve (AUC).

Main Results:

• Kernel Support Vector Machine (KSVM) achieved the highest predictive accuracy (AUC: 0.834).
• KSVM outperformed fractional polynomial (FP) regression (AUC: 0.799).
• XGBoost showed lower performance, potentially due to dataset size limitations.

Conclusions:

• Machine learning, specifically KSVM, demonstrates high predictive accuracy for palliative care survival when sufficient data is available.
• Traditional statistical models offer benefits in stability and interpretability.
• Model selection should be tailored based on specific data characteristics for optimal prognostic prediction.