Κωστούλας Θεόδωρος

Περίληψη:
Background Risk stratification in intensive care unit (ICU) stroke patients remains challenging. Machine learning (ML) can leverage medical data to improve prognostication and identify novel biomarkers. Inflammatory and nutritional markers, including red cell distribution width (RDW), albumin, and their ratio (RAR), have demonstrated potential prognostic value in ICU. Aims To build and externally validate ML models for mortality prediction in ICU stroke patients and evaluate RDW and RAR as prognostic biomarkers. Methods Clinico-laboratory data from stroke patients in freely accessible MIMIC-IV and eICU databases were extracted via ICD codes, in accordance with Health Insurance Portability and Accountability Act. Time-series features were summarized in overlapping 3-hour windows over the first 48 hours (mean, median, min, max). Only first ICU stays ≤10 days were included. RAR was defined as mean RDW/mean albumin in the first 24 hours. Decision tree and XGBoost models were developed; SHAP assessed interpretability and decision curve analysis clinical utility. A patient-level 90/10 development–test split with 5-fold StratifiedGroupKFold cross-validation was used for tuning, and performance evaluated on independent test and external validation sets using AUC, PR-AUC, and Brier scores. Mortality outcomes were in-hospital, 30-, 180-, and 365-day. Results Patients’ demographic and clinical characteristics are presented in Table 1. XGBoost demonstrated the best performance for predicting in-hospital (first 3hour model, AUC = 0.96), out of hospital 30-day (AUC = 0.88), 180-day (AUC = 0.87), and 365-day mortality (AUC = 0.90). External validation in the eICU cohort showed modest but acceptable discrimination (in-hospital, 48-hour model, AUC = 0.81). Key predictors of in-hospital mortality included Glasgow Coma Scale, age, glucose, RDW, inspired O2 fraction, anion gap, SOFA, white blood cells, liver enzymes, BMI, and blood pressure. Conclusions ML models can effectively predict short- and long-term mortality in ICU stroke patients. RDW emerged as important prognostic marker, reflecting systemic inflammation. External validation supports the generalizability of the approach. However, key clinical variables such as NIHSS scores and neuroimaging data were unavailable in these databases, and late mortality outcomes could not be externally validated in the eICU cohort. Further research is needed to clarify the biological mechanisms underlying RDW and RAR as prognostic indicators.

Περίληψη:
On a daily basis, medical decisions play a key role in healthcare services and patient well-being. This study aims to develop and validate machine learning (ML) models to predict both length of stay (LOS) and mortality in critically-ill stroke patients using the clinical-laboratory data available during the first 48 hours from admission, as those can be retrieved from the electronic health record of a patient. For this purpose, we utilized the MIMIC-IV database extracting clinical, laboratory, and demographic data. To capture time changes we grouped data in three-hour over-lapping observation windows for the first 48 hours of the patient’s stay in the Intensive Care Unit (ICU). The results obtained indicate that the root mean square error and the mean absolute error ranged from 1.35 to 2.07 days and 1.04 to 1.5 days, respectively for LOS and Area Under the Curve between 0.802 and 0.819 for mortality. This study highlights the importance of using ML techniques for predictions in the ICU, for patients suffering from stroke. The experimental evaluation demonstrated the significant potential of the XGBoost model in predicting both LOS and mortality, demonstrating the potential for efficient resource allocation and patient management. Our findings can contribute to optimizing clinical decision-making processes and to the overall improvement of quality of care in intensive care environments by providing personalized treatment and care based on the intended outcome.

Περίληψη:
Ischemic stroke is a medical emergency that requires hospitalization and occasionally, specialized care at the Intensive Care Unit. Mortality prediction in the ICUs has been a challenge for intensivists, since prompt identification could impact medical clinical practices and allow efficient allocation of health resources in the ICUs, which are extremely restricted, especially in the era of COVID-19 pandemic. Clinical decision support systems based on machine learning algorithms are taking advantage of the vast amount of information available in the ICUs and are becoming popular in the medical predictive analysis. This study aims to explore the feasibility of interpretable machine learning models to predict mortality in critically-ill patients suffering from stroke. To do so, a vast variety of clinical and laboratory information stored in the electronic health record, are pre-processed to allow taking into account the temporal characteristics of a patient’s stay. An 8-hour sliding observation window was utilized. For the experimental evaluation we used the Medical Information Mart for Intensive Care Database (MIMIC-IV). Results indicate sufficient ability to predict mortality at the end of a given day during the patient’s stay. Moreover, attribute evaluation highlights the important indicators to consider when following up with a patient.