Self-Aware SGD: Reliable Incremental Adaptation Framework for Clinical AI Models.

Area of Science:

• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Applications
• Deep Learning for Medical Data Analysis

Background:

• Healthcare data is dynamic, leading to distribution shifts that degrade clinical AI model performance.
• Deployed AI models require continuous adaptation to maintain accuracy and relevance.
• Incremental learning offers adaptation but risks model integrity due to unreliable data.

Purpose of the Study:

• To introduce a novel incremental deep learning algorithm, self-aware stochastic gradient descent (SGD).
• To enhance the reliability of adapting clinical AI models to evolving data distributions.
• To ensure model integrity is maintained during the incremental training process.

Main Methods:

• Developed self-aware SGD, an incremental deep learning algorithm.
• Integrated a contextual bandit-like sanity check to analyze and filter gradient updates.
• Utilized the sanity check to isolate and remove unreliable gradients, ensuring only trustworthy modifications are applied.

Main Results:

• Self-aware SGD effectively balances incremental training with model integrity.
• The algorithm demonstrated reliable incremental updates in experimental evaluations.
• Performance was validated on Oxford University Hospital datasets, particularly under label noise conditions.

Conclusions:

• Self-aware SGD provides a robust solution for adapting clinical AI models to distribution shifts.
• The method enhances the trustworthiness of AI in dynamic healthcare environments.
• This approach mitigates risks associated with compromised or mislabeled data in incremental learning.