Survival Machine Learning Methods Improve Prediction of Histologic Transformation in Follicular and Marginal Zone...

Area of Science:

• Hematology
• Oncology
• Computational Biology

Background:

• Follicular lymphoma (FL) and marginal zone lymphoma (MZL) are indolent B-cell lymphomas with a risk of transformation to aggressive subtypes.
• Predicting histologic transformation (HT) is challenging due to data complexities and class imbalances.
• Existing prognostic indices do not specifically target HT risk.

Purpose of the Study:

• Develop and validate machine learning models for predicting HT in FL and MZL.
• Compare the performance of survival-based and traditional classification models.
• Investigate the impact of next-generation sequencing (NGS) data on HT prediction.

Main Methods:

• Utilized a multicenter retrospective dataset (n=1068) for model development.
• Compared various survival models (e.g., XGBoost-Cox, Lasso-Cox) and classification models (e.g., Logistic Regression, Random Forest).
• Validated best-performing models on an independent test set (n=92) and incorporated NGS data.

Main Results:

• Survival models outperformed traditional classifiers in predicting HT.
• XGBoost-Cox demonstrated the highest predictive performance (accuracy 85.3%, AUC 0.795).
• NGS data integration improved model accuracy and specificity, identifying key genes like TP53, BLM, and RAD50 associated with HT risk.

Conclusions:

• Survival-based machine learning models offer significant clinical value for HT risk stratification.
• Integration with NGS data enables personalized risk assessment for FL and MZL patients.
• This approach can aid in clinical decision-making for managing these lymphomas.