Predicting of anaphylaxis in big data EMR by exploring machine learning approaches

Area of Science:

• Medical Informatics
• Computational Biology
• Allergy and Immunology

Background:

• Anaphylaxis is a severe allergic reaction requiring epidemiological study for prevention and treatment strategies.
• Electronic Medical Records (EMR) offer rich data for anaphylaxis epidemiology but manual review is labor-intensive.
• Automating EMR analysis is crucial for efficient and cost-effective epidemiological research.

Purpose of the Study:

• To explore machine learning techniques for automatic classification of anaphylaxis cases in EMR data.
• To reduce the manual effort required for large-scale anaphylaxis epidemiological studies.
• To compare the effectiveness of various text classification methods and document representations for EMR data.

Main Methods:

• Utilized diverse machine learning classifiers including Logistic Regression, Linear SVM, Multilayer Perceptron, Random Forest, and Convolutional Neural Networks (CNN).
• Compared document representations such as Bag of Words (BoW) and word embedding models (average, bag of centroids).
• Implemented a novel clustering-based undersampling technique to address the class imbalance in EMR data (anaphylaxis cases <1%).

Main Results:

• Most classifiers and representations achieved high performance (F1 score >90%).
• Logistic Regression, Linear SVM, Multilayer Perceptron, and Random Forest demonstrated F1 scores around 95%, with linear methods showing faster training times.
• Convolutional Neural Networks (CNN) achieved a slightly superior F1 score of 95.6%.

Conclusions:

• Machine learning, particularly CNNs, offers an effective and efficient solution for automatically identifying anaphylaxis cases in EMR.
• Automated classification significantly reduces the time and cost associated with anaphylaxis epidemiological studies.
• The developed methods provide a valuable tool for advancing anaphylaxis research and public health strategies.