DBNet: a novel deep learning framework for mechanical ventilation prediction using electronic health records

The outbreak of the Coronavirus disease (COVID-19) pandemic has caused millions of deaths and put immense pressure on the health care system, especially the supply of mechanical ventilators. It is critical for clinicians to identify the patients in a timely manner whose status may deteriorate in the near future and therefore need mechanical ventilators. We propose a prediction model to estimate the probability of requiring mechanical ventilation for in-hospital patients at least 24 hours after their admission. Our model is a multi-modal encoder-decoder attention model that takes full usages of the electronic health record (EHR) database. The EHR database consists of heterogeneous data tables of different formats (diagnosis, drug administrations, medicine prescriptions, lab tests, vital sign observations, clinical procedures, and demographics). We leverage the attention mechanism to increase model performance and promote result interpretability. The attention mechanism also serves the role of the missing data imputation technique, which is often used on irregularly sampled temporal data. We name the model as DBNet as the model takes the database as input. DBNet is evaluated on a large cohort of COVID-19 patients and the result shows it outperforms the state-of-the-art baseline deep learning models in predicting the future requirement of mechanical ventilation. It also outperforms several machine learning models even with sophisticated feature engineering. Due to its ability to handle multiple tables and also longitudinal data, the DBNet can also is not limited to this single application and can be generalized to other healthcare prediction tasks.