A Framework for Locally Imputing and Predicting Biomarker Trajectories Under Irregular Monitoring: Application to...

Area of Science:

• Biomedical data science
• Clinical informatics
• Longitudinal data analysis

Background:

• Irregular patient monitoring and missing data hinder the use of longitudinal biomarkers in clinical practice.
• Accurate tracking of biomarker trajectories is crucial for effective disease management, particularly in conditions like chronic myeloid leukemia (CML).

Purpose of the Study:

• To develop a generalizable framework for creating complete biomarker trajectories and forecasting future values from irregularly collected data.
• To enhance the utility of longitudinal biomarkers in real-world clinical settings by addressing data gaps.

Main Methods:

• Developed a framework combining interval-aligned preprocessing, localized multiple imputation, and machine learning (RNN, XGBoost) for forecasting.
• Applied the method to BCR::ABL1 monitoring data in CML, aligning measurements to 90-day intervals.
• Utilized a windowed, uncertainty-propagating imputation strategy and trained models to predict values 3 and 6 months ahead.

Main Results:

• Achieved low Root Mean Square Errors (RMSEs) of 1.22-1.24 for 3-month BCR::ABL1 predictions, below observed variability.
• Maintained prediction accuracy even when the most recent patient visit data was excluded, simulating extended follow-up.
• Demonstrated the framework's ability to preserve local temporal structure and support individualized monitoring.

Conclusions:

• The proposed framework effectively generates complete biomarker trajectories and enables accurate future value prediction from irregular data.
• This approach enhances the reliability and utility of longitudinal biomarkers in routine clinical practice for diseases like CML.
• The framework is adaptable to other continuous biomarkers measured under real-world, non-ideal schedules.