| dc.description.abstract | The rise in publicly available healthcare databases, such as MIMIC and the eICU, now make it possible to revolutionize medical care when paired with modern machine learning techniques. The MIMIC-IV critical care database allows us to explore these techniques in the ICU setting using data from thousands of patients. One area that can be improved upon in the medical domain is prediction of events in the ICU setting, such as whether a patient will have a heart attack during their stay. Through improved prediction of events, hospitals can be more efficient and better allocate resources to patients who need it most, saving both lives and costs. In the ICU setting, there has been previous work for prediction of events via machine learning classification models. However, we believe time-to-event models may offer more accuracy and interpretability than these classification models. We also believe that current, popular time-to-event models are limited in their scope, either not being able to deal with dynamic data, being too slow to use in real time, or having to make assumptions about the underlying structure of the data. Some models also require restructuring of the data into specific formats which leads to information loss. These kinds of restrictions are not desirable in the ICU, where measurements come in frequently at irregular intervals and requires fast prediction of events. It follows, then, that we need dynamic, lightweight, time-to-event models for prediction of events that do not make assumptions about the data’s structure. In this paper, we use BoXHED, a lightweight, dynamic, boosting, time-to-event model and compare it to other time-varying models in the ICU. To evaluate the different models’ performances, we used time series data from the MIMIC-IV database by refactoring code used previously for MIMIC-III preprocessing by Harutyunyan et al. We then compared the different models’ accuracy in predicting mortality in the ICU, both as new data measurements became available and using measurements within the first 48 hours of the patients’ stays. We then evaluated the models based on an approach inspired by TREWScore where patient risk scores were compared to given thresholds to obtain each models’ AUC-ROC scores. | en |
