Nucleated red blood cell distribution in critically ill patients with acute pancreatitis: a retrospective cohort study

Abstract

Objectives: This study examined the potential association between nucleated red blood cell (NRBC) levels and mortality in critically ill patients with acute pancreatitis (AP) in the intensive care unit, due to limited existing research on this correlation.

Methods: This retrospective cohort study utilized data from the MIMIC-IV v2.0 and MIMIC-III v1.4 databases to investigate the potential relationship between NRBC levels and patient outcomes. The study employed restricted cubic splines (RCS) regression analysis to explore non-linear associations. The impact of NRBC on prognosis was assessed using a generalized linear model (GLM) with a logit link, adjusted for potential confounders. Furthermore, four machine learning models, including Gradient Boosting Classifier (GBC), Random Forest, Gaussian Naive Bayes, and Decision Tree Classifier model, were constructed using NRBC data to generate risk scores and evaluate the potential of NRBC in predicting patient prognosis.

Results: A total of 354 patients were enrolled in the study, with 162 (45.8%) individuals aged 60 years or older and 204 (57.6%) males. RCS regression analysis demonstrated a non-linear relationship between NRBC levels and 90-day mortality. Receiver Operating Characteristic (ROC) analysis identified a 1.7% NRBC cutoff to distinguish survivor from non-survivor patients for 90-day mortality, yielding an Area Under the Curve (AUC) of 0.599, with a sensitivity of 0.475 and specificity of 0.711. Elevated NRBC levels were associated with increased risks of 90-day mortality in both unadjusted and adjusted models (all Odds Ratios > 1, P < 0.05). Assessment of various machine learning models with nine variables, including NRBC, Sex, Age, Simplified Acute Physiology Score II, Acute Physiology Score III, Congestive Heart Failure, Vasopressin, Norepinephrine, and Mean Arterial Pressure, indicated that the GBC model displayed the highest predictive accuracy for 90-day mortality, with an AUC of 0.982 (95% CI 0.970-0.994). Post hoc power analysis showed a statistical power of 0.880 in the study.

Conclusions: Elevated levels of NRBC are linked to an increased mortality risk in critically ill patients with AP, suggesting its potential for predicting mortality.

Keywords: Acute Pancreatitis; Cohort study; Critical care; Nucleated red blood cell; Risk factors.

Fig. 1

Correlation analysis between NRBC and disease scores

Fig. 2

Association of NRBC with 90-Day Mortality, 28-Day Mortality, and ICU Mortality (a-c) RCS regression; (d-f) Grouped regression; (g-i) ROC curve. NOTE: The high, medium, and low groupings of NRBC in Fig. 2d-f are established according to the tertile divisions of NRBC levels

Fig. 3

Effect size of NRBC on 90-day mortality in prespecified and exploratory subgroups

Fig. 4

The NRBC-based risk score efficiently predicts 90-day Mortality (a) ROC curves for the performance of various machine learning models in predicting 90-day mortality are as follows: GBC model AUC (95% CI) = 0.982 (0.970–0.994); DTC model AUC (95% CI) = 0.947 (0.928–0.967); RF model AUC (95% CI) = 0.822 (0.779–0.865); GNB model AUC (95% CI) = 0.683 (0.627–0.739); (b) Precision-Recall curves for various machine learning models illustrate performance metrics, with a greater curve towards the upper right corner indicating superior model performance.; (c) Ranking of feature variable importance in the GBC models; (d) SHapley Additive Explanation visualization illustrating the individual feature impacts on the outcomes of the GBC prediction model. The color spectrum represents the degree of influence, where redder hues indicate a higher association with the risk of 90-day Mortality, while bluer hues suggest a lower association with the risk of 90-day Mortality; (e) Survival curve analysis in subgroups with different risk scores. NOTE: The three subgroups for the survival curve analyses were established based on the tertiles derived from the mortality probability generated by the GBC prognostic model constructed using the 9 variables