Deep Learning: A Heuristic Three-Stage Mechanism for Grid Searches to Optimize the Future Risk Prediction of Breast...

Area of Science:

• Computational biology
• Machine learning in healthcare
• Deep learning for cancer research

Background:

• Grid search is effective for optimizing deep learning models but faces time management challenges.
• Low-budget grid searches require efficient strategies to avoid excessive computation time.
• Breast cancer metastasis risk prediction is a critical area for clinical decision-making.

Purpose of the Study:

• To introduce a heuristic three-stage mechanism for managing low-budget grid search time in deep learning.
• To apply sweet-spot grid search (SSGS) and randomized grid search (RGS) strategies for optimizing breast cancer metastasis risk prediction models.
• To enhance the prediction performance of deep feedforward neural network (DFNN) models for 5-, 10-, and 15-year metastasis risk.

Main Methods:

• Developed deep feedforward neural network (DFNN) models for breast cancer metastasis risk prediction.
• Implemented an eight-cycle, three-stage grid search process to optimize hyperparameters.
• Utilized SHAP analyses to interpret model predictions and hyperparameter contributions.

Main Results:

• Grid searches improved 5-, 10-, and 15-year breast cancer metastasis risk prediction by 18.6%, 16.3%, and 17.3% respectively, compared to RGS.
• The three-stage mechanism proved effective for feasible and manageable low-budget grid searches.
• SHAP analyses identified key clinical risk factors and important DFNN hyperparameters for prediction.

Conclusions:

• Grid search significantly enhances deep learning model prediction performance.
• The proposed three-stage mechanism effectively manages grid search time and improves model accuracy.
• SHAP analyses provide valuable insights into both clinical factors and model hyperparameters for breast cancer metastasis prediction.