Machine learning decision tree models for multiclass classification of prognosis in patient undergoing palliative radiotherapy for bone metastases.

backgroundAccurate estimation of prognosis and life expectancy is essential in patients with advanced cancer, as it guides clinical decision-making and helps avoid unnecessary interventions while facilitating timely integration of palliative and supportive care. Palliative radiotherapy plays a key role within multidisciplinary management, offering effective and well-tolerated symptom relief for complications such as pain, bleeding, and obstruction, with treatment strategies closely tailored to expected survival. Although recent advances in machine learning have improved prognostic accuracy by modeling complex variable interactions, their application in palliative care settings remains limited. PURPOSE: To aid clinical decision-making, we developed a decision tree multi-classifier to predict the mortality at 3, 24, and 52 weeks following palliative radiotherapy for bone metastases.

methodsData from 573 adults diagnosed with metastatic cancer were analyzed. The primary endpoint was the overall survival (OS) defined as the number of months from treatment to death event. Four clinically relevant classes were defined: Class 0 (OS: ≤ 3 weeks), Class 1 (OS: 3-24 weeks), Class 2 (OS: 24-52 weeks) and Class 3 (OS ≥ 52 weeks). Candidate covariate predictors consisted of 65 clinical, dosimetric and laboratory variables. Two supervised decision tree machine-learning models were trained and validated using the Python package. A SHapley Additive exPlanations (SHAP) explanaibility analysis was performed to infer the global and local feature importance.

resultsThe SHAP analysis selected three laboratory variables, the interleukin8, haemoglobin and lymphocytes count as the first three ranked variables representing the major impact on OS in each of the four classes and accounting for more than 80% of contribution. In all classes, higher chance of OS was associated with low values of interleukin8 (IL8) and higher values of haemoglobin (HEM) and lymphocytes count (LYMPH). Pre-treatment values of IL8 > 36.7 relocated more than 50% of patients with survival < 3 weeks and only 1.5% of patient with survival > 52 weeks. On the other hand, pre-treatment values of IL8 < 19 relocated about 92% of patients with survival > 52 weeks. Patients are then additionally separated based on the lymphocytes count (LYMPH). LYMPH values higher than 7.5 will drive the probability of survival > 52 weeks still over 90% while it drops down to 2.1% for LYMPH < 7.5.

conclusionAn explainable machine learning approach based on decision trees is able to predict the survival at different timing after radiotherapy in patients with advanced cancer. This approach provides an intelligible explanation of individualized risk prediction, helping clinicians to identify the best strategy for patient stratification and treatment selection.