Machine Learning Assisted Discovery of Novel Predictive Lab Tests Using Electronic Health Record Data

Area of Science:

• Biomedical Informatics
• Computational Biology
• Epidemiology

Background:

• Identifying biological markers for disease diagnosis is crucial but challenging.
• Current methods are often time-consuming, expensive, and require significant expertise.
• Improving the quality of initial marker hypotheses is essential for efficient research.

Purpose of the Study:

• To develop a high-throughput computational pipeline for generating and ranking hypothesized laboratory tests for disease diagnosis.
• To enhance the efficiency and reduce the cost of identifying novel diagnostic markers.
• To improve the accuracy of initial hypotheses for epidemiological studies.

Main Methods:

• A pipeline combining machine learning models for hypothesis generation.
• Text mining techniques for filtering and ranking hypotheses based on novelty.
• Logistic regression analysis to corroborate final candidate hypotheses.
• Validation using electronic health record data and the PubMed corpus.

Main Results:

• The pipeline successfully generated a large number of candidate laboratory test-diagnosis hypotheses.
• Novelty assessment using text mining effectively filtered and ranked potential markers.
• Logistic regression confirmed several promising candidate hypotheses.
• The approach demonstrated feasibility on real-world datasets.

Conclusions:

• The proposed high-throughput pipeline offers an efficient and effective method for identifying high-quality hypothesized diagnostic markers.
• This approach can accelerate the discovery of new laboratory tests for disease diagnosis.
• The integration of machine learning, text mining, and statistical analysis provides a robust framework for biomarker discovery.