External validation of AI-based scoring systems in the ICU: a systematic review and meta-analysis

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Prediction Models

Background:

• Machine learning (ML) models are increasingly used for predicting clinical deterioration in intensive care units (ICUs).
• However, these models often exhibit poor performance in new hospital settings due to overfitting.
• External validation is crucial for assessing the reliability of ML-based risk scores before clinical implementation.

Purpose of the Study:

• To systematically review the frequency of external validation for ML-based risk scores in ICU patient deterioration prediction.
• To evaluate the impact of external validation on the performance of these ML models.

Main Methods:

• A systematic literature search was conducted across MEDLINE, Web of Science, and arXiv for studies predicting ICU patient deterioration using ML.
• Included studies were published in English before December 2023.
• The frequency of external validation was summarized, and changes in performance (Area Under the Receiver Operating Characteristic curve - AUROC) were analyzed using linear mixed-effects models.

Main Results:

• Out of 572 included studies, only 14.7% underwent external validation, rising to 23.9% by 2023.
• External validation was more common in studies using open-source data, particularly MIMIC and eICU datasets.
• On average, AUROC decreased by -0.037 in external data, with nearly half of studies showing a performance drop greater than 0.05.

Conclusions:

• External validation of ML risk scores in ICUs is infrequent, despite its increasing trend.
• The observed performance decline in external datasets raises concerns about the generalizability and reliability of many proposed ML scores.
• Over-reliance on limited datasets and exclusive use of AUROC hinder robust interpretation and validation.