Incremental logistic regression for customizing automatic diagnostic models.

Area of Science:

• Medical Informatics
• Machine Learning in Medicine
• Statistical Learning for Healthcare

Background:

• Clinical Decision Support Systems (CDSS) increasingly use statistical learning for automatic diagnostic models.
• Developing these models requires extensive data, posing challenges in collection, preprocessing, validation, and ethical approval.
• External validation of CDSS models often results in underperformance due to data discrepancies between centers.

Purpose of the Study:

• To introduce an incremental learning algorithm based on Bayesian inference for developing and updating diagnostic models.
• To enable model creation with limited initial data and facilitate incremental updates as new data becomes available.
• To allow for efficient recalibration of models for different healthcare centers using a reduced number of cases.

Main Methods:

• Developed an incremental learning algorithm utilizing Bayesian inference.
• Employed benchmark datasets and a real-world brain tumor dataset for evaluation.
• Compared the proposed algorithm against a previous incremental algorithm and a non-incremental Bayesian model.

Main Results:

• The proposed Bayesian incremental learning algorithm demonstrated good convergence.
• The algorithm proved to be independent of the data model, offering flexibility.
• Effective performance was shown in updating models with new data and recalibrating for different centers.

Conclusions:

• The novel incremental learning algorithm offers an efficient approach to developing and maintaining diagnostic models in medicine.
• This method addresses data limitations and ethical concerns associated with traditional model development.
• The algorithm's adaptability makes it suitable for dynamic healthcare environments and multi-center applications.