Robust Disease Prognosis via Diagnostic Knowledge Preservation: A Sequential Learning Approach.

Area of Science:

• Medical Artificial Intelligence
• Deep Learning in Healthcare
• Prognostic Modeling

Background:

• Accurate disease prognosis is crucial but often limited by insufficient long-term data.
• Deep learning models can leverage large diagnostic datasets for improved prognostic capabilities.
• Naive fine-tuning for prognosis risks 'catastrophic forgetting,' degrading essential diagnostic accuracy.

Purpose of the Study:

• To explore deep learning training strategies for enhancing disease prognosis prediction.
• To investigate methods for pretraining models on diagnostic data without compromising diagnostic accuracy.
• To evaluate a sequential learning strategy with experience replay for prognostic tasks.

Main Methods:

• Utilized large diagnostic datasets (radiographs, MRIs, mammograms) for pretraining.
• Applied a sequential learning strategy with experience replay to mitigate catastrophic forgetting.
• Evaluated model performance on predicting disease progression in knee osteoarthritis, Alzheimer's disease, and breast cancer.

Main Results:

• Diagnostic pretraining significantly improved prognostic performance (e.g., AUROC, AUPRC) compared to baselines.
• The sequential learning approach achieved comparable prognostic accuracy to single-task models.
• This method successfully preserved diagnostic accuracy, unlike simpler multitask approaches prone to forgetting.

Conclusions:

• Leveraging large diagnostic datasets via pretraining is an effective and data-efficient strategy for enhancing prognostic models.
• Sequential learning with experience replay enables robust prognostic prediction while retaining critical diagnostic skills.
• This approach offers a safer and more reliable method for clinical AI deployment in disease prognosis.