Creating a database for health IT events via a hybrid deep learning model

Area of Science:

• Health Informatics
• Artificial Intelligence in Healthcare
• Medical Device Safety

Background:

• Poorly designed health information technology (HIT) can increase risks by disrupting workflows and encouraging workarounds.
• Analyzing HIT events is crucial for improving patient safety, but accessible reports are limited.
• The FDA MAUDE database contains valuable information on medical device events.

Purpose of the Study:

• To develop a hybrid deep learning model for accurately identifying health information technology (HIT) event reports.
• To establish the first publicly accessible database of HIT event reports using FDA data.
• To improve the understanding and management of risks associated with HIT implementation.

Main Methods:

• A hybrid deep learning model combining logistic regression, Convolutional Neural Networks (CNN), and Hierarchical Recurrent Neural Networks (RNN) was developed.
• The model was trained and evaluated on 6994 samples from the FDA MAUDE database.
• The optimal model achieved high accuracy (0.903) and AUC (0.954) in identifying HIT events.

Main Results:

• The optimal hybrid model demonstrated superior performance compared to individual models.
• The model achieved an accuracy of 0.862 on an independent dataset.
• An HIT event database containing 48,997 reports was generated from the MAUDE database (1991-2018).

Conclusions:

• The developed hybrid deep learning model effectively identifies HIT event reports.
• The creation of a comprehensive HIT event database facilitates better understanding and management of HIT-related risks.
• This resource can aid healthcare professionals in addressing challenges posed by health information technology.