Developing predictive models using electronic medical records: challenges and pitfalls.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Data Science

Background:

• Electronic Medical Records (EMR) offer rich patient data but contain systematic biases.
• Standard machine learning algorithms often fail due to violations of their underlying assumptions when applied to EMR data.

Purpose of the Study:

• To discuss critical issues and potential pitfalls in building predictive models from EMR.
• To emphasize the importance of aligning EMR data characteristics and model clinical utility.
• To present methods for training and evaluating EMR-based predictive models.

Main Methods:

• Analysis of systematic biases within EMR data.
• Evaluation of machine learning algorithm suitability for EMR.
• Development and application of specific approaches for EMR model training and validation.
• Case study: Early prediction of septic shock.

Main Results:

• Failure to account for EMR characteristics and clinical use can result in less practical predictive models.
• Specific methodologies are required to overcome data biases and improve model reliability.
• The proposed approaches demonstrate effectiveness in the context of septic shock prediction.

Conclusions:

• Building effective predictive models from EMR necessitates a nuanced understanding of data properties and clinical application.
• Careful methodological choices are crucial for mitigating biases and ensuring model utility in real-world healthcare settings.
• The study provides a framework for developing reliable predictive models from complex EMR data.