LSTM-Based Decision Support System for ICU Discharge

Background: Whether a patient can be discharged from an intensive care unit (ICU) or not is usually decided by the treating physicians based on their clinical experience. However, nowadays limited capacities and growing socioeconomic burden of our health systems increase the pressure to discharge patients as early as possible, which may lead to higher readmission rates and potentially fatal consequences for the patients.

Method: To help physicians decide whether a patient can be safely discharged from the ICU, we developed a deep learning model based on long short-term memory (LSTM), which estimates a probability of the patient's readmission to cardiovascular ICU based on their clinical parameters. We compared the performance of our LSTM to a standard deep learning model called “Feed Forward Neural Network (FNN)” and logistic regression (LR). Additionally, we trained our LSTM based model on patient histories of all ICU types. For the training test and evaluation of all models we used the MIMIC-III dataset.

Results: The area under the receiver operator characteristic curve (AUCROC) was 0.818 for the LSTM, as compared with 0.716 for the LR and 0.739 for the FNN models, indicating that models that capture temporal dependencies provide better results. Additionally, the training of our LSTM based model on patient histories from all types of ICUs shows an AUCROC of 0.855, indicating that the performance of our model can be further improved by a larger dataset.

Conclusion: The deep learning solution presented in this paper can help physicians to decide on patient discharge from the ICU. This may not only help increase the quality of patient care but may also help reduce costs and to optimize ICU resources. Further, the presented LSTM-based approach may help improve existing and develop new medical machine learning prediction models. However, further studies are warranted to test and improve the model with real life data.

Publikationsverlauf

Artikel online veröffentlicht:
03. Februar 2022