Benchmarking machine learning models on multi-centre eICU critical care dataset.

Area of Science:

• Critical Care Medicine
• Machine Learning
• Health Informatics

Background:

• Tracking machine learning progress in critical care is hindered by a lack of public benchmarks.
• Established benchmarks exist in fields like computer vision and natural language processing.
• Large clinical datasets are now available, enabling benchmark creation.

Purpose of the Study:

• To establish a public benchmark suite for critical care machine learning.
• To address key areas: mortality prediction, length of stay estimation, patient phenotyping, and decompensation risk.
• To compare clinical models against baseline and deep learning models.

Main Methods:

• Developed a benchmark suite for four critical care tasks.
• Utilized the eICU critical care dataset (approx. 73,000 patients).
• Compared clinical models, baseline machine learning, and deep learning models.

Main Results:

• Presented the first public benchmark on a multi-center critical care dataset.
• Compared benchmark performance against clinical standards and predictive models.
• Investigated the impact of numerical and categorical variable handling.

Conclusions:

• The proposed benchmark suite facilitates progress tracking in critical care machine learning.
• Publicly available code allows for reproducibility and further research.
• The benchmark provides a standardized evaluation for predictive models in critical care.