Incremental Discrimination of a Logistic-Regression Model Using Physiological and Missingness Predictors Beyond Six-Hour Mean Arterial Pressure for Sustained Hypotension: A FHIR-Based Pilot Study

Background: Multivariable early-warning models can integrate physiological signals and missingness patterns, but added complexity may not provide meaningful value over simple predictors. We evaluated whether a single prespecified L2-regularized logistic-regression model using physiological and missingness predictors improved the prediction of sustained hypotension beyond recent mean arterial pressure (MAP). Methods: We conducted a retrospective pilot prognostic-model study using the MIMIC-IV Clinical Database Demo on FHIR 2.1.0. Predictors were derived from the preceding 12 h, and sustained hypotension was defined as MAP < 65 mmHg for at least two consecutive hours during the following six hours. A single prespecified L2-regularized logistic-regression model using eight physiological or derived predictors and ten missingness-related predictors was compared with six-hour mean MAP. Performance was estimated using patient-grouped nested cross-validation with five outer and three inner folds. The primary comparison was the difference in area under the precision–recall curve (AUPRC); advancement required an absolute gain of ≥0.020. Paired uncertainty was estimated using 1000 patient-cluster bootstrap replicates. A prespecified sensitivity analysis recalculated AUPRC after assigning equal total weight to each patient, irrespective of the number of eligible windows contributed. Results: Ninety-two patients contributed 8970 eligible windows, of which 1774 (19.8%) were positive. The candidate achieved an AUPRC of 0.629 versus 0.622 for the comparator and an AUROC of 0.845 versus 0.842. The AUPRC difference was +0.0075 (95% bootstrap interval, +0.0005 to +0.0171). The prespecified patient-equal sensitivity analysis yielded a larger delta AUPRC point estimate of +0.0156; both the patient-equal and window-weighted estimates remained below the +0.020 advancement margin. In a post hoc analysis, 81.6% of positive windows had observed MAP below 65 mmHg during t − 5 through t. Among windows without observed recent hypotension, the delta AUPRC was +0.0030 (95% interval, −0.0288 to +0.0238). The 8970 prediction windows were overlapping repeated observations nested within 92 patients and should not be interpreted as independent samples. Conclusions: In this pilot internal-validation study, the single prespecified L2-regularized logistic-regression model using eight physiological or derived predictors and ten missingness-related predictors produced a small improvement in discrimination over six-hour mean MAP but did not meet the prespecified development-advancement margin. These findings apply only to this model specification, dataset, and outcome and do not establish the performance of multivariable vital-sign models generally. Neither strategy was validated for clinical use.