| Title | A Comparative Machine Learning Modelling Approach for Patients' Mortality Prediction in Hospital Intensive Care Unit |
|---|
| Authors | Aldraimli, M., Nazyrova, N., Djumanov, A., Sobirov, I. and Chaussalet, T.J. |
|---|
| Editors | Sotirov, S.S., Pencheva, T., Kacprzyk, J., Atanassov, K.T., Sotirova, E. and Staneva, G. |
|---|
| Type | Conference paper |
|---|
| Abstract | Mortality prediction in a hospital Intensive Care Unit (ICU) is a challenge that must be addressed with high precision. Machine Learning (ML) is a powerful tool in predictive modelling but subject to the problem of class im-balance. In this study, we tackle class imbalance with combining new features, data re-sampling, ensemble learning and an appropriate selection of evaluation metrics in a clinical setting. We built and evaluated 126 ML mod-els to predict mortality in 48546 ICU admissions extracted from the Medical Information Mart for Intensive Care III (MIMIC-III) repository.
In our study design, six mortality prediction datasets are extracted; five of which are legacy dataset sets while the remainder is our new constructed dataset. For our combined data models, when testing on isolated data, our selection of features enhanced the prediction performances beyond those for the traditional legacy sets used in research. The legacy datasets are the Simplified Acute Physiology Score (SAPS II), the Sequential Organ Failure Assessment score (SOFA), the Glasgow Coma Scale (GCS), Elixhauser Comorbidity Index (ECI) and Demographics & Disease Groups (DDG). Our approach has a considerable impact on the classification; it resulted in an improvement in the mortality status prediction. For evaluation, we implement a comparative multi-stage evaluation filter for binary classification to compare all models. The best models are identified. The Area Under Receiver Operator Characteristic curves of the tested models range from 0.57 to 0.94. These encouraging results can guide further development of models to allow for more reliable ICU mortality predictions. |
|---|
| Keywords | Classification |
|---|
| Imbalanced Data |
|---|
| MIMIC III |
|---|
| Tomek links |
|---|
| SMOTE |
|---|
| Ensembles |
|---|
| ICU Mortality Prediction |
|---|
| Year | 2022 |
|---|
| Conference | International Symposium on Bioinformatics and Biomedicine |
|---|
| Publisher | Springer |
|---|
| Accepted author manuscript | File Access Level Open (open metadata and files) |
|---|
| Publication dates |
|---|
| Published | 12 Mar 2022 |
|---|
| Journal | Lecture Notes in Networks and Systems |
|---|
| Journal citation | 374, pp. 16-31 |
|---|
| ISSN | 2367-3389 |
|---|
| 2367-3370 |
|---|
| Book title | The International Symposium on Bioinformatics and Biomedicine: BioInfoMed 2020: Contemporary Methods in Bioinformatics and Biomedicine and Their Applications |
|---|
| Digital Object Identifier (DOI) | https://doi.org/10.1007/978-3-030-96638-6_2 |
|---|