Application of interpretable machine learning algorithms to predict distant metastasis in osteosarcoma.

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Osteosarcoma is the most frequent bone cancer in pediatric and adolescent populations.
• Prognosis and treatment strategies differ significantly between localized and metastatic osteosarcoma at diagnosis.

Purpose of the Study:

• To identify potential risk factors associated with metastatic osteosarcoma.
• To develop accurate predictive models for distant metastasis in osteosarcoma patients.

Main Methods:

• Utilized the Surveillance, Epidemiology, and End Results (SEER) database (2004-2015) for patient data.
• Developed and compared six machine learning (ML) models (logistic regression, SVM, GaussianNB, XGBoost, RF, kNN) for metastasis prediction.
• Employed adaptive synthetic (ADASYN) sampling for imbalanced data and Shapley Additive Explanation (SHAP) for model interpretation.

Main Results:

• Machine learning models achieved an average precision (AP) of 0.661-0.781 for predicting distant metastasis.
• The random forest (RF) model demonstrated superior performance with 71.8% accuracy and an AP of 0.781.
• RF model performance was significantly influenced by tumor size, primary surgery status, and patient age; SHAP analysis confirmed key clinical factors.

Conclusions:

• An effective machine learning-based prediction model for osteosarcoma metastasis has been developed.
• This model can assist clinicians in making informed decisions for patient management and treatment planning.