An Adaptive Deep Ensemble Learning Method for Dynamic Evolving Diagnostic Task Scenarios.

Area of Science:

• Machine Learning
• Medical Informatics
• Computer-Aided Diagnosis

Background:

• Complex machine learning models can be difficult for physicians to interpret.
• Variations in data across diagnostic tasks and institutions lead to inconsistent model performance.

Purpose of the Study:

• To propose an adaptive deep ensemble learning method (TPE-DEM) for dynamic diagnostic scenarios.
• To create a model that is understandable to physicians and adaptable to varying data characteristics.

Main Methods:

• Combined Deep Ensemble Model (DEM) with tree-structured Parzen Estimator (TPE).
• Utilized TPE to aggregate simpler, interpretable models.
• Enabled dynamic selection of optimal model layers and basic learners based on task data.

Main Results:

• The TPE-DEM model demonstrated superior performance compared to baseline models across diverse datasets.
• The model achieved optimal performance by adapting its structure to different diagnostic tasks.
• Tested on hospital and multiple UCI datasets.

Conclusions:

• The TPE-DEM offers a novel, understandable machine learning approach for computer-aided diagnosis.
• This method addresses challenges of variable datasets and feature sets in diagnostic applications.
• Findings have significant implications for deploying machine learning in clinical settings.