(1785) Ensemble Model With Inherently Good-Quality Data Improves Early Mortality Prediction in ICUs

Introduction: Critical Care Units are technologically advanced, enabling data driven approaches towards precision medicine. Previous studies have developed prediction models for outcomes using databases like MIMIC, but random and non-random data missingness in them require temporal harmonisation of variables. This is challenging, thus, conventional imputation techniques are utilized, either inherently or as preprocessing, which may not be appropriate for heterogenous, complex data.
Implementing data quality metrics can help; however a pragmatic approach may entail identifying the inherent good quality data subset for improving model performance, as explored in this study.

Methods: Patients under 18, over 80 years old, or with ICU LOS < 48 hours from MIMICIV were excluded. ICU readmissions within 48h were clubbed. Mortality within 48h of ICU discharge was considered in-ICU mortality. Data was binned into 1h bins taking median for multiple values. For early prediction of mortality data was truncated to 48h and the binary outcome was in-ICU mortality within 7 days of admission. Thresholds for measurements, procedures and drug selection were based on a graphical interpretation of sparsity. Domain based filtering of measurements was performed. Missingness per visit was used to identify a cohort with good quality data (cGQD).
Class balanced XGBoost models were trained using the total cohort(XG-T) and cGQD(XG-GQ) for each hour bin, and classification was performed using ensemble. Test set was undersampled to ensure class balance.

Results: We included 5 static measurements, 30 'chartevents', 26 'inputevents' and 31 'procedureevents' after filtering. Total cohort had 22493 and cGQD had 8702 visits. With classification done by '0' for threshold <=0.3 and '1' for >=0.7, and minimum ensemble votes=10, XG-T had an accuracy of 74.4%, F1 of 74.1% and missed classifying 81/902 visits. In contrast, XG-GQ classified all samples and had an accuracy of 80.8% and F1of 81.7%.

Conclusions: Using cGQD, XG-GQ showed improved performance in predicting mortality across the dataset, showing potential of inherent good quality data. Future work requires adopting better strategies for handling variables of different missingness patterns, drawing trajectories, optimising hyperparameters and validation in external datasets.