Patient Mortality Prediction Using Clinical Notes

Predicting various aspects of medical diagnosis and prognosis is an ever expanding discipline in machine learning. In this paper we present a comparison of the predicting capability of conventional algorithms, traditional deep learning algorithms and fine tuned deep learning models on a prognosis prediction task using clinical reports from various medical professionals. We used the MIMIC III dataset containing various types of medical reports on hospital admissions and predicted the mortality of the patients discharged from the hospital. The objective was to find out the best model to predict the mortality of patients being admitted and particularly, how accurately we could do this at the beginning of admission compared to when the patient got discharged. Additionally, with all the hype around deep learning, we also wanted to compare various conventional and the deep learning classifiers in order to determine how well the deep learning ones performed compared to the conventional ones. The results showed that mortality prediction with multiple clinical notes at the discharge time was only marginally better then the predictions with only the first clinical note. We also show that the conventional classifiers perform as good as the traditional deep learning algorithms such as LSTM and ConvNet. However, the fine tuned BERT models fine tuned on Pubmed data performed much better on all aspects of the prediction task.