Machine Learning Predicts ICU Deaths in PAD Patients, but Hemodynamics Add Little

An explainable machine-learning model predicted 28-day mortality among critically ill patients carrying a peripheral artery disease diagnosis with good discrimination when tested on data from a later calendar period, but a detailed set of early hemodynamic measurements failed to add measurable predictive value beyond standard severity scores and routine clinical variables. That is the central finding of a secondary analysis of the MIMIC-IV intensive care database, published in Artery Research, which also found that individual hemodynamic measures such as shock index and pulse pressure were associated with mortality as risk markers without being validated treatment targets.

The study drew on Medical Information Mart for Intensive Care IV version 3.1, a deidentified electronic health-record database from Beth Israel Deaconess Medical Center in Boston. The researchers screened adults aged 18 or older with at least one prespecified ICD-9 or ICD-10 code compatible with lower-extremity peripheral artery disease and retained the first intensive care unit stay of the first eligible hospital admission. Importantly, this definition identifies critically ill patients who carry PAD diagnosis codes; it does not establish PAD as the reason for admission or confirm symptomatic disease. From 3029 candidate first ICU stays, the team excluded 81 deaths within 24 hours and 489 short or missing-length stays, leaving a landmark cohort of 2459 patients.

The design used a 24-hour landmark: all predictors were derived only from data available in the first 24 hours after ICU admission, and patients who died within that window or whose stays were shorter than 24.0 elapsed hours were excluded. The primary outcome was all-cause death from the landmark through day 28. Of the 2459 patients, 435, or 17.7 percent, died by day 28. Admissions from 2008 to 2016 formed the development cohort of 1924 patients with 329 deaths, while admissions from 2017 to 2022 formed an untouched temporal test cohort of 535 patients with 106 deaths. The models developed on the earlier period were evaluated in the later period without refitting, a form of internal temporal validation rather than validation at a different institution.

The full Extreme Gradient Boosting, or XGBoost, classifier used 47 predictors, including age, sex, coded comorbidities, the Sequential Organ Failure Assessment score, the Simplified Acute Physiology Score II, the Acute Physiology Score III, the Charlson Comorbidity Index, organ support, first laboratory values, and summaries of vital signs. Four hemodynamic variables were included: mean shock index, mean pulse pressure, the standard deviation of mean arterial pressure, and the fraction of valid mean arterial pressure measurement points below 65 mmHg, a measurement fraction rather than a duration-weighted time fraction. The model used fixed regularized settings with 300 trees, a learning rate of 0.03, and a maximum depth of 4.

In the 2017–2022 test cohort, the full model achieved an area under the receiver-operating-characteristic curve of 0.830 (95 percent confidence interval 0.786–0.871), an area under the precision-recall curve of 0.602 (0.524–0.682), a Brier score of 0.117 (0.104–0.130), and an expected calibration error of 0.029. Calibration was assessed with particular care because the test cohort had a modestly higher mortality rate than the development cohort, 19.8 percent versus 17.1 percent, a difference that may reflect temporal shifts in case mix or severity. The calibration intercept was 0.097 (−0.149 to 0.379) and the slope was 0.879 (0.728–1.064). For comparison, a parsimonious logistic regression reached an AUROC of 0.826, while the SOFA score alone reached 0.742 (0.688–0.796).

The key incremental-value test came from paired ablation comparisons with 2000 bootstrap confidence intervals. Removing the four hemodynamic variables produced a numerically higher AUROC of 0.837, giving a paired difference of −0.007 (95 percent CI −0.024 to 0.010) for the full model minus the ablated model. The AUPRC difference was 0.001 (−0.035 to 0.039) and the Brier-score difference was less than 0.001 (−0.005 to 0.006). The authors interpret this carefully: the point estimate favored the ablated model, but the interval was imprecise and compatible with small changes in either direction, so the hemodynamic block did not demonstrate measurable incremental predictive value. Removing the severity scores themselves produced only small, imprecise changes, and Shapley additive explanation values showed that APS III, SAPS II, the Charlson index, and age dominated the full-model ranking; after severity-score removal, age, red-cell distribution width, anion gap, blood urea nitrogen, respiratory rate, and temperature were most prominent, with mean pulse pressure and shock index remaining lower-ranked contributors.

Missing data were handled with 20 multiple imputations generated by iterative chained equations, since lactate was missing in 33.6 percent of patients and mean shock index in 6.0 percent, and complete-case restriction would have discarded 895 patients. Imputation models included the outcome indicator, the Nelson–Aalen cumulative hazard, all analysis variables, and auxiliary predictors, and the observed pattern was inconsistent with missing completely at random, so conditional missing-at-random was treated as a working assumption rather than a proven mechanism. Sensitivity analyses using a lactate-missing indicator, complete cases, and development-fitted median imputation produced similar temporal performance.

Separately from prediction, landmark Cox models beginning at the 24-hour mark estimated adjusted associations between continuous hemodynamic measures and 28-day mortality, with adjustment for age, sex, Charlson index, SOFA, sepsis status, invasive ventilation, vasopressor exposure, creatinine, hemoglobin, and lactate. Higher mean shock index (per 0.1: hazard ratio 1.19; 95 percent CI 1.13–1.25), lower mean pulse pressure (per 10 mmHg decrease: HR 1.20; 1.12–1.28), and greater mean arterial pressure variability (per 5 mmHg standard deviation: HR 1.20; 1.10–1.30) were associated with mortality, while the MAP-below-65 measurement fraction was not (per 0.10: HR 1.00; 0.96–1.04). Spline tests supported nonlinearity for pulse pressure and gave nominal, cautiously interpreted evidence for shock index. Because global proportional-hazards tests were significant, the Cox coefficients summarize average associations across follow-up, and window-specific models for days 0–7 and 7–27 were reported as sensitivity analyses.

A prespecified exploratory composite instability phenotype—positive for vasopressor exposure, a MAP-below-65 measurement fraction of at least 0.10, or a mean shock index of at least 0.90 during the first 24 hours—separated unadjusted Kaplan–Meier curves (HR 1.72; 1.36–2.19) but was not independently associated with mortality after full adjustment (HR 1.07; 0.82–1.39). The result remained compatible with no association in atherosclerosis-coded, invasive-pressure, and Sepsis-3-negative subsets and in the complete-case analysis, consistent with the composite’s overlap with treatment exposure and illness severity. The authors caution that vasopressor administration is influenced by indication, timing, and clinician choice, and that these data do not validate the composite as an independent prognostic phenotype or treatment trigger.

The study’s limitations are substantial. PAD was identified from discharge diagnosis codes without ankle–brachial index, imaging, limb severity assessment, or admission-reason adjudication, threatening construct validity and restricting interpretation to ICU patients carrying PAD codes. The single-center database limits transportability, missingness was informative, missing-not-at-random bias remains possible, bedside measurements were summarized rather than analyzed as waveforms, and residual confounding and treatment-by-indication affect the association models. The authors stress that temporal validation does not establish clinical utility and that no independent external cohort was available; the model is not a deployment-ready clinical tool. They call for external validation in a multicenter ICU database reproducing the 24-hour landmark and feature definitions, with calibration reported alongside discrimination, and for assessment of whether hemodynamic waveforms or duration-weighted exposure measures add information beyond severity scores.

Subject of Research: Temporal validation of an explainable machine-learning model for 28-day mortality prediction in ICU patients with peripheral artery disease

Article Title: Temporal Validation of Explainable Machine Learning for 28-Day Mortality Prediction in Adults with Peripheral Artery Disease in Intensive Care

Article References: Isleyen, H. B., Yavuz, S. T., Bulut, S., Ozcan, S., Sozen, A. A., Uluganyan, M., & Okuyan, E. (2026). Temporal Validation of Explainable Machine Learning for 28-Day Mortality Prediction in Adults with Peripheral Artery Disease in Intensive Care. Artery Research, 32(1), Article 32. https://doi.org/10.1007/s44200-026-00131-0

Image Credits: AI Generated

DOI: 10.1007/s44200-026-00131-0

Keywords: peripheral artery disease, machine learning, XGBoost, intensive care, mortality prediction, MIMIC-IV, temporal validation, shock index, pulse pressure, hemodynamics, severity scores, explainable AI