Using Active Learning for Speeding up Calibration in Simulation Models.

Area of Science:

• Computational Biology
• Machine Learning in Oncology
• Mathematical Modeling of Cancer

Background:

• Cancer simulation models rely on unobservable parameters for disease onset and tumor growth.
• Estimating these parameters requires lengthy calibration involving numerous simulation runs.
• Efficient calibration is crucial for accurate prediction of cancer incidence and mortality.

Purpose of the Study:

• To demonstrate the application of machine learning (ML) for accelerating cancer simulation model calibration.
• To reduce the number of parameter combinations evaluated during the calibration process.
• To improve the efficiency of developing race-specific cancer models.

Main Methods:

• Developed and applied an active learning algorithm, a type of ML, to guide parameter selection.
• The algorithm interactively identifies parameter combinations likely to yield desired outputs.
• Validated the method using the University of Wisconsin breast cancer simulation model (UWBCS).

Main Results:

• Identified all 69 crucial parameter combinations for a race-specific UWBCS model by evaluating only 5,620 out of 378,000 combinations.
• Achieved a significant reduction in computational effort compared to standard calibration methods.
• Demonstrated that evaluating just 1.49% of all possible parameter combinations was sufficient.

Conclusions:

• Machine learning methods, specifically active learning, can substantially expedite cancer model calibration.
• This approach guides researchers toward more promising parameter combinations, reducing simulation workload.
• The findings highlight the potential of ML to accelerate cancer research and model development.