Automatic Prediction of Recurrence of Major Cardiovascular Events: A Text Mining Study Using Chest X-Ray Reports.

Area of Science:

• Cardiology
• Medical Informatics
• Artificial Intelligence

Background:

• Cardiovascular diseases remain a leading cause of mortality worldwide.
• Accurate prediction of cardiovascular events is crucial for timely intervention.
• Electronic Health Records (EHR) contain valuable unstructured data, such as radiology reports, that can potentially improve risk prediction.

Purpose of the Study:

• To develop and evaluate a deep learning-based multimodal architecture for predicting the recurrence of major cardiovascular events.
• To assess the added value of integrating unstructured text data from radiology reports with clinical predictors.

Main Methods:

• Utilized EHR data from the Second Manifestations of ARTerial disease (SMART) study.
• Developed a text mining pipeline integrating neural text representation with clinical predictors.
• Employed various machine learning models including logistic regression, SVM, MLP, CNN, LSTM, and BiLSTM.
• Evaluated models in two scenarios: integration with clinical predictors and integration with age and sex only.

Main Results:

• The multimodal BiLSTM (MI-BiLSTM) model achieved an AUC of 84.7% when integrating radiology reports with clinical predictors.
• The Support Vector Machine (SVM) model with bag-of-words representation showed the lowest misclassification rate (12.2%) in the first scenario.
• In the absence of clinical predictors, the MI-BiLSTM model achieved an AUC of 74.5% using radiology reports and demographic data.

Conclusions:

• Integrating text features from chest X-ray reports with classical predictors enhances cardiovascular risk prediction.
• Deep learning models, specifically MI-BiLSTM with word embeddings, demonstrate strong performance in this multimodal approach.
• Text data from radiology reports, when combined with clinical and laboratory values, offer superior predictive power compared to models using only known clinical predictors.