(371) Augmentation of Mortality Predictive Models in ICU Setting Through Synergistic Feature Engineering

Introduction: In complex environment of Intensive Care Units (ICU), predicting patient mortality is of paramount importance to guide therapeutic and family decisions. Contemporary prediction models often utilize single physiological indicators potentially missing critical integrative and synergistic information. This study hypothesizes that a combined measure of patient's biochemical fragility represented by Serum Albumin, and acute illness severity, quantified by the Apache score, would yield superior mortality predictive power.

Methods: This study leveraged a dataset of 148,000 electronic health records from the eICU Collaborative Research Database and it included adult ICU admissions with common diagnoses. The feature set consisted of lab values including age, sodium, diabetes status, bilirubin, glucose, lactate, BUN, albumin, and bicarbonate levels, as well as net fluid balance. To assess potential interactions, two engineered features were created: the product of albumin and Apache score (ALBUMIN_APACHE), and their ratio (ALBUMIN/APACHE). Missing data was imputed using multivariate imputation. Gradient boosting machine models were developed using the LightGBM implementation to predict in-hospital mortality and performance was evaluated using F1 score on a held-out test set. Feature importance was assessed using SHAP values. The primary evaluation metric was F1 score, reflecting the balance of precision and recall.

Results: Gradient boosting machine model with engineered features demonstrated improved predictive performance, with an F1 score of 0.303 versus 0.245 without engineered features. Analysis of feature importance revealed the constructed interaction variables ALBUMIN_APACHE and ALBUMIN/APACHE were highly ranked, exceeding standalone albumin in predictive importance for mortality.

Conclusions: This investigation underscores the significant potential of feature engineering in enhancing the predictive accuracy of mortality models in ICU environments. Integration of information capturing a patient's biochemical resilience, epitomized by serum Albumin, with an acute illness severity index, such as the Apache score, facilitated a superior predictive capacity. Future research endeavors could explore additional combinations to further augment the performance of mortality prediction models in the ICU.