A three-axis burden score and in-hospital mortality in ICU patients with heart failure and atrial fibrillation: limited incremental discrimination and transportability in a retrospective cohort study

Background ICU patients with heart failure (HF) and atrial fibrillation (AF) have substantial in-hospital mortality. We defined a 0–3 burden score from a low prognostic nutritional index (PNI), a high neutrophil-to-lymphocyte ratio (NLR) and a low dynamic lactate clearance rate (dLCR), and asked whether the count is associated with in-hospital mortality after adjustment and whether it improves discrimination in a second database. Methods Cut-points were the medians of observed values in MIMIC-IV and were applied unchanged in both cohorts. Derivation used MIMIC-IV v3.1 (n = 7,562; 1,222 deaths, 16.2%) and external assessment used eICU-CRD 2.0 (n = 2,588; 414 deaths, 16.0%). Missing values were handled by multiple imputation (m = 10). Associations were odds ratios from logistic models. Incremental discrimination was the change in area under the curve relative to a 43-variable reference model. Results In the imputed cohorts, in-hospital mortality rose from 10.8% to 24.9% in MIMIC-IV and from 10.2% to 24.0% in eICU-CRD across scores 0 to 3. Among patients in whom all three axes were measured, mortality was higher at every score, and the eICU-CRD gradient was not monotonic. In MIMIC-IV each additional point had an adjusted odds ratio of 1.20 (95% CI 1.08–1.33). The corresponding estimate in eICU-CRD was 1.14 (0.97–1.35). Out-of-fold, the score raised the AUC of a reference model that already contained albumin, the white-cell count and lactate by 0.0028 (0.0011–0.0045), less than the gain from the axes as continuous terms. That increment was not detectable when the models were refitted in eICU-CRD. The reference model kept its rank order after transport (AUC 0.711) but was miscalibrated; two-parameter recalibration, evaluated out of fold, lowered the Brier score from 0.140 to 0.124. Conclusions The count is a descriptive stratification tied to the derivation cut-points. The adjusted association is supported in MIMIC-IV and not in eICU-CRD. It is not a validated or transportable prediction tool, and the absolute risks from the reference model are not available from the three measurements alone.