Augmentation of Electronic Medical Record Data for Deep Learning.

Area of Science:

• Machine Learning
• Medical Informatics
• Data Science

Background:

• Class imbalance is a significant challenge in developing machine learning models, especially for predicting critical clinical outcomes.
• Existing methods like over/under-sampling and loss weighting are common, but data augmentation lacks standardization for discrete time-series data.
• Electronic Medical Records (EMR) generate discrete time-series data vital for clinical predictions.

Purpose of the Study:

• To propose and demonstrate a simple, effective data augmentation strategy for discrete time-series data.
• To address the lack of consensus on data augmentation techniques for this data type.
• To improve the performance of machine learning models dealing with imbalanced clinical datasets.

Main Methods:

• A novel data augmentation strategy specifically designed for discrete time-series data derived from EMR.
• Application and validation of the proposed strategy on a publicly available dataset.
• Proof-of-concept demonstration for the augmentation technique's utility.

Main Results:

• The proposed strategy offers a viable method for augmenting discrete time-series data.
• Demonstrates the potential to enhance machine learning model performance on imbalanced clinical datasets.
• Provides a foundational method for future research in this area.

Conclusions:

• The developed data augmentation strategy is a practical solution for handling imbalanced discrete time-series data.
• This method can improve the reliability of machine learning models predicting important clinical endpoints.
• Further research can build upon this strategy for broader applications in medical informatics.