Routine Admission Variables Add Only Modest Predictive Value to APACHE II in Acute Pancreatitis

Routine variables collected when patients arrive at hospital may add some predictive information to a widely used severity score in acute pancreatitis, but a new retrospective study suggests the gain is modest and not yet ready for clinical practice. The research, published in Clinical and Experimental Medicine, examined whether everyday admission measurements improve on the Acute Physiology and Chronic Health Evaluation II score, known as APACHE II, for predicting intensive care unit admission and systemic complications.

The study team assessed 264 patients with acute pancreatitis in a retrospective cohort design. Rather than building a model from scratch, the investigators asked a specific statistical question: does adding routine admission variables to APACHE II measurably improve early prediction of the two endpoints? They evaluated endpoint-specific complete-case logistic models using discrimination, calibration, reclassification metrics, and decision curve analysis, and compared patients who were included with those who had to be excluded.

The analyses were constrained by missing data and small numbers of events. The ICU admission analysis included 160 patients, of whom 34 required intensive care. Patients excluded from this analysis actually had a higher ICU admission rate, 39.4 percent versus 21.3 percent among those included, a signal of selection bias the authors explicitly flagged. A key source of missingness was C-reactive protein: of 104 patients without an exact numeric CRP value, 89 had blank records and 15 had reports at the assay’s detection limit. The systemic complication analysis was larger, covering 226 patients with 89 events.

On the surface, the combined models looked strong. The apparent area under the receiver operating characteristic curve for the Combined model versus APACHE II alone was 0.947 versus 0.928 for ICU admission, and 0.968 versus 0.962 for systemic complications. However, neither difference reached statistical significance, with P values of 0.0958 and 0.1536 respectively. In other words, the apparent edge of the combined models could not be distinguished from chance in these samples.

Bootstrap resampling offered a more conservative picture. The researchers performed 1,000 bootstrap resamples with fixed predictor sets to correct for optimism, a common problem when models are evaluated on the same data used to fit them. The optimism-corrected AUCs for the Combined model were 0.937 for ICU admission and 0.962 for systemic complications, confirming that both approaches performed well in absolute terms but that the added variables changed little.

Reclassification metrics told a more nuanced story. Apparent category-free net reclassification improvement was 0.877 for ICU admission and 0.674 for systemic complications, while integrated discrimination improvement was 0.090 and 0.027 respectively. These numbers suggest the combined models shifted some patients across risk thresholds, but the authors caution that such apparent gains must be weighed against the absence of a significant AUC improvement and the small event counts.

C-reactive protein, one of the routine variables of interest, showed limited selection stability across the resampling procedure, meaning its contribution to the models was inconsistent. This instability, combined with the large share of blank or assay-limited CRP records, further weakens any claim that this biomarker reliably adds early predictive information in this setting.

The authors frame their conclusions carefully. Routine variables showed what they describe as preliminary incremental information, without a statistically significant AUC improvement over APACHE II. They identify selection bias, limited events, and endpoint heterogeneity as factors that constrain interpretation of the results.

The practical takeaway is one of restraint. The authors state that the exploratory Combined models require independent validation before clinical use.

For now, the study stands as a methodologically transparent example of prediction model evaluation: it reports apparent and optimism-corrected performance side by side, quantifies missing data, and treats reclassification gains with appropriate skepticism.

Subject of Research: Whether routine admission variables add predictive value to APACHE II for early prediction of ICU admission and systemic complications in acute pancreatitis

Article Title: Incremental value of routine admission variables added to clinical scores for early prediction of ICU admission and systemic complications in acute pancreatitis

Article References: Xu, L., Lin, Z., & Cheng, F. (2026). Incremental value of routine admission variables added to clinical scores for early prediction of ICU admission and systemic complications in acute pancreatitis. Clinical and Experimental Medicine. https://doi.org/10.1007/s10238-026-02356-4

Image Credits: AI Generated

DOI: 10.1007/s10238-026-02356-4

Keywords: acute pancreatitis, APACHE II, ICU admission, prediction model, C-reactive protein, retrospective cohort, systemic complications, reclassification, bootstrap validation, clinical scoring, Incremental, value