Utility based approach in individualized optimal dose selection using machine learning methods.

Area of Science:

• Biostatistics
• Machine Learning
• Personalized Medicine

Background:

• Personalized medicine aims to tailor treatments using patient data for improved outcomes.
• Optimal dose selection requires balancing treatment efficacy and toxicity.
• Existing methods may not fully leverage patient characteristics and biomarkers for dose optimization.

Purpose of the Study:

• To develop and evaluate machine learning models for personalized optimal dose selection.
• To balance efficacy and toxicity outcomes using patient features and biomarkers.
• To incorporate clinician-defined utilities for decision-making.

Main Methods:

• Utilized flexible machine learning models (Random Forest, Gaussian Process) to predict binary efficacy and toxicity.
• Employed copula models to jointly analyze outcomes, enforcing non-decreasing dose-response relationships.
• Incorporated elicited clinical utilities to maximize posterior mean for optimal dose selection.
• Investigated alternative approaches with toxicity constraints and utility uncertainty.

Main Results:

• Gaussian process models demonstrated superior performance compared to Random Forest.
• Monotonicity constraints offered minor improvements in model performance.
• Modeling the correlation between efficacy and toxicity had minimal impact on results.
• The methods were successfully illustrated in a liver cancer patient cohort.

Conclusions:

• Machine learning, particularly Gaussian processes, offers a robust framework for personalized optimal dose finding.
• The proposed methods effectively balance efficacy and toxicity, guided by clinical utilities.
• These approaches hold promise for optimizing cancer treatment, such as stereotactic body radiation therapy for liver cancer.