An Oversampling-Enhanced Multi-Class Imbalanced Classification Framework for Patient Health Status Prediction Using...

Area of Science:

• Oncology
• Health Informatics
• Machine Learning

Background:

• Patient-reported outcomes (PROs) are vital for managing treatment toxicities in cancer patients undergoing radiation therapy.
• Accurate PRO data analysis is crucial for clinical decision-making and post-treatment survivorship care.
• Raw PRO data often suffers from sparsity and imbalanced toxicity distributions, hindering predictive modeling.

Purpose of the Study:

• To investigate advanced machine learning techniques for predicting patient outcomes like pain and sleep disturbances.
• To address data sparsity and imbalanced distributions in PRO datasets from a cancer therapy center.
• To enhance clinical decision support for managing treatment-related toxicities.

Main Methods:

• Implemented advanced classifiers: Random Forest (RF), XGBoost, Gradient Boosting (GB), Support Vector Machine (SVM), MLP-Bagging, and Logistic Regression (LR).
• Applied oversampling techniques to manage minority class imbalance while preserving class ratios.
• Utilized PRO datasets from a cancer therapy center across three cancer types.

Main Results:

• Random Forest (RF) and XGBoost demonstrated strong generalization capabilities in multi-class imbalance tasks.
• The models successfully categorized post-therapy severity levels for various toxicities.
• These machine learning approaches proved effective in handling challenges posed by PRO data.

Conclusions:

• RF and XGBoost are highly effective for predicting patient-reported outcomes and toxicity severity.
• These models offer valuable clinical decision support for oncologists and care teams.
• Advanced machine learning can overcome limitations in PRO data for improved cancer patient care and survivorship.