Benchmarking deep learning models on large healthcare datasets.

Area of Science:

• Clinical informatics
• Artificial intelligence in healthcare
• Biomedical data science

Background:

• Deep learning models (Deep Neural Networks) have shown success in various fields and are increasingly applied to clinical healthcare.
• Benchmarking deep learning against traditional machine learning and prognostic scoring systems on public healthcare data is limited.
• The Medical Information Mart for Intensive Care III (MIMIC-III) dataset provides a valuable resource for such comparisons.

Purpose of the Study:

• To benchmark the performance of deep learning models against state-of-the-art machine learning and established prognostic scoring systems.
• To evaluate these models on clinical prediction tasks including mortality, length of stay, and ICD-9 code group prediction.
• To assess the impact of using 'raw' clinical time series data as input features.

Main Methods:

• Utilized the publicly available MIMIC-III (v1.4) dataset, containing ICU patient data from 2001-2012.
• Implemented and compared Deep Learning models, an ensemble of machine learning models (Super Learner algorithm), SAPS II, and SOFA scores.
• Performed benchmarking across multiple clinical prediction tasks.

Main Results:

• Deep learning models demonstrated superior performance compared to all other evaluated approaches.
• This outperformance was particularly pronounced when utilizing 'raw' clinical time series data as input.
• The findings highlight the potential of deep learning for complex clinical prediction tasks.

Conclusions:

• Deep learning models offer a powerful approach for clinical prediction tasks, surpassing traditional methods.
• The use of raw clinical time series data is key to unlocking the full potential of deep learning in healthcare.
• Further research and application of deep learning are warranted in clinical informatics and patient outcome prediction.