Updating methods for artificial intelligence-based clinical prediction models: a scoping review

Area of Science:

• * Medical Artificial Intelligence (AI)
• * Clinical Prediction Modeling
• * Machine Learning in Healthcare

Background:

• * AI-based prediction models are vital in healthcare but may not generalize to new settings.
• * Developing new models for each context is inefficient and wasteful.
• * Updating existing AI models offers a practical solution for improved performance and resource utilization.

Purpose of the Study:

• * To provide a comprehensive overview of methods for updating AI-based clinical prediction models.
• * To categorize and describe existing model updating techniques and their use cases.
• * To guide researchers in adapting AI models to new data and clinical scenarios.

Main Methods:

• * Comprehensive literature search of Scopus and Embase up to August 2022.
• * Focused on AI-based prediction models updated with new data in the medical domain.
• * Excluded regression-based updating methods; categorized identified AI updating methods.

Main Results:

• * 78 articles were included, predominantly focusing on neural network updates (93.6%) using medical images (65.4%).
• * Common use cases include adapting broad models to specialized tasks, addressing data drift, and handling cross-center variations.
• * Identified methods categorized into neural network-specific (92.3%), ensemble-specific (2.5%), and model-agnostic (9.0%).

Conclusions:

• * Numerous methods exist for updating AI-based prediction models across various use cases.
• * Updating methods for non-neural network AI models (e.g., random forests) are under-researched in clinical settings.
• * This review serves as guidance to improve the reuse, quality, and efficiency of AI models in healthcare.