Page 62 - AIH-2-2

Artificial Intelligence in Health Predicting ICU mortality: A stacked ensemble model

several algorithmic perspectives can improve prediction dataset (before undersampling), aligned with the SHAP
accuracy, outperforming many previously reported diagram of our final dataset presented in this study. This
findings. suggests that our approach did not overlook any critical
nuances inherent in the majority class.
Many studies use various clinical features as inputs in
their attempts to achieve higher metrics for predicting In conclusion, the findings of this research highlight
mortality in the ICU. Although it can be advantageous to the enormous potential of ML, especially stacked ensemble
expand the scientific search by incorporating a variety of learning for precise mortality prediction among ICU
patient characteristics, direct performance comparisons patients. Medical personnel, including nurses, might find
are difficult due to a plethora of factors in ML models. this method useful in providing information on optimal
Furthermore, the number of attributes used in related patient care procedures and allocating resources in an ICU.
research is sometimes excessive, overlooking the difficulty Further investigation holds potential in examining whether
of measurement that may be unfeasible in many ICUs. this model can be applied to a variety of patient demographics
In our study, we used almost exclusively the clinical and healthcare settings. In essence, our research envisions
parameters of APACHE IV, as this is one of the most enhancing mortality prediction, fostering a better grasp
commonly accepted systems for measuring the severity and of patient dynamics, and ultimately contributing to the
mortality prognosis of ICU patients. However, we did not continuous improvement of ICU practices.
include other parameters, which would potentially affect
the significance of each metric and alter the final results. Acknowledgments
This research has some limitations. The dataset None.
was derived from a database that includes many U.S. Funding
hospitals. The debate about using a more generalized
dataset versus being sensitive to more local data and None.
adapting models to the local characteristics of each
hospital is associated with the aim of each study. The Conflict of interest
first option may produce results that are less specific The authors declare they have no competing interests.
but more flexible across scenarios, and the latter is more
precise but less generalizable to other patient groups. Author contributions
The purpose of this research effort was to define a Conceptualization: Dimitrios Simopoulos, Dimitrios
model architecture that is validated on larger datasets, Kosmidis, George Anastassopoulos
with increased prediction accuracy. Despite the cross- Investigation: Dimitrios Simopoulos, Dimitrios Kosmidis
validation technique that we applied, it is likely that the Methodology: Dimitrios Simopoulos
accuracy we achieved would be different in real local Formal analysis: Dimitrios Simopoulos
datasets such as a specific hospital, or in specific patient Writing–original draft: Dimitrios Simopoulos, Dimitrios
groups. The chronological age of our data was another Kosmidis
limitation. The data was collected between 2014 and Writing–review & editing: George Anastassopoulos,
2015. With the rapid progress in critical care, there have Lazaros Iliadis
been advancements in medical and nursing care, as well
as changes in technology and strategies for admitting, Ethics approval and consent to participate
treating, and discharging patients. Furthermore, our Not applicable.
data do not include patients during the pandemic time,
which may differentiate results and comparisons with Consent for publication
results from other similar studies during the pandemic,
or even after the pandemic period. Finally, although the Not applicable.
undersampling strategy and its advantages were justified, Availability of data
in general, it can also introduce certain limitations. In
particular, removing significant major-class records can The data used in this paper were obtained from the “eICU
lead to overfitting if not properly executed. In addition, Collaborative Research Database” (https://eicu-crd.mit.edu).
bias toward one class in the undersampled data may
be reflected in the final results, leading to non-optimal References
performance on unseen data. Our analysis revealed 1. Simopoulos D, Kosmidis D, Koutsouki S, Bonnotte N,
though that the SHAP values derived from the original Anastassopoulos G. Advanced mortality prediction in adult

Volume 2 Issue 2 (2025) 56 doi: 10.36922/aih.4981