JMIR Publications

ABSTRACT

Background:

In recent years, the increased availability of electronic health records (EHRs) alongside the advancement of machine learning algorithms has enabled many interesting applications in the field of medicine like predicting the progression of certain diseases or the length of stay at the hospital. It also allowed scientists to have a better understanding of the patients’ diagnosis and allow for individualized treatment options. For example, understanding the diagnosis of a mechanically ventilated patient can allow the use of better treatment options with the help of machine learning techniques. This proves to be particularly important these days because of the COVID-19 virus that caused a sharp increase in the number of hospitalized patients in need of ventilation. Additionally, with the high number of lab values taken during an ICU stay, the doctors can easily miss some important lab values and their abnormal changes.

Objective:

The goal of this work is to predict future abnormalities in 25 different lab values from past lab values for patients in the intensive care unit (ICU). Detecting anomalies can then be used to support clinicians in their decision-making process in an ICU by drawing their attention to values that can get out of range.

Methods:

The approach presented in our work consists of a pre-processing phase followed by time series prediction based on Deep Neural Networks (DNN). The preprocessing involves data imputation to convert sparse time series data into a normal time series. In the Second Phase, we use four current deep learning algorithms for time series prediction. Finally, we train and test our pipeline on two of the most popular datasets in the medical field: MIMIC-III and eICU.

Results:

The results compare the performance of the developed models on the datasets providing guidance on which is the best model to use on similar medical data. In our case, the deep learning model based on the multi-CNN architecture performed the best on both datasets as well as being the fastest during training and inference. The model can be used in combination with our preprocessing pipeline on real-life EHRs to continuously predict future abnormalities in lab values.

Conclusions:

Using deep learning methods, our prediction model consistently achieved satisfactory results in the prediction of future abnormalities in lab values. The system was trained, validated, and tested on two well-known datasets to ensure that our system bridged the reality gap as much as possible. Finally, the model can be used in combination with our preprocessing pipeline on real-life EHRs to improve patients’ diagnosis and treatment. Clinical Trial: There have been no trials in our work