IDENTIFICATION OF CLINICAL PARAMETERS PREDICTING HEMODYNAMIC INSTABILITY IN ICU SETTING

Abstract

Background. Haemodynamic instability is a leading determinant of morbidity and mortality in the intensive care unit (ICU). A clinically important but poorly characterised phenotype is the patient who progresses from haemodynamic stability to instability without receiving vasopressor support, whether through delayed recognition or resource constraint. South Asian ICU populations are markedly underrepresented in the haemodynamic-instability literature. We characterised the baseline physiological differences and serial deterioration trajectories of such patients in a Pakistani tertiary-care cohort and developed an interpretable machine-learning early-warning system on the same data. Methods. We studied 253 adult ICU patients at the Sindh Institute of Urology and Transplantation (SIUT), Karachi. Patients were classified as haemodynamically unstable (n = 118) if they developed sustained systolic blood pressure (SBP) < 90 mmHg and/or mean arterial pressure (MAP) < 65 mmHg without any vasopressor or vasoactive agent during deterioration, or stable (n = 135). Twenty clinical parameters were recorded at 2-hour intervals over a maximum 48-hour window (6,071 observations). Groups were compared using Mann-Whitney U and chi-square/Fisher tests. From 125 engineered features (vital-sign time-series dynamics, patient-level aggregates, composite clinical indices, and admission laboratory values), 50 were selected by SHAP importance and used to train a weighted ensemble of LightGBM, XGBoost, and logistic regression, validated by 5-fold patient-level stratified cross-validation and on a held-out test set. Results. At admission, unstable patients had lower SBP (102.3 ± 22.0 vs 120.4 ± 19.6 mmHg, p < 0.001), lower MAP (76.5 ± 15.5 vs 87.7 ± 13.5 mmHg, p < 0.001), higher shock index (1.056 ± 0.360 vs 0.813 ± 0.224, p < 0.001), and higher lactate (2.8 ± 2.4 vs 1.5 ± 0.9 mmol/L, p < 0.001), whereas arterial pH, bicarbonate, creatinine, and haematocrit did not differ significantly, indicating that haemodynamic divergence preceded laboratory derangement. In the unstable group, MAP reached a nadir of 63.6 ± 9.0 mmHg at T+6h, shock index remained ≥ 1.0 and lactate remained > 2.0 mmol/L throughout, all without vasopressor support. The ensemble achieved an AUROC of 0.9941 ± 0.0064, sensitivity 96.67%, specificity 95.59%, and a Brier score of 0.0316 in cross-validation (test-set AUROC 0.9928), and retained an AUROC of 0.9788 using only the first 6 hours of data. Conclusions. In a resource-constrained South Asian ICU, patients who deteriorated without vasopressor support followed a reproducible trajectory in which vital signs—particularly MAP, SBP, and shock index—signalled compromise earlier than laboratory tests. An interpretable physiology-driven ensemble identified these patients with high discrimination and good calibration from the first hours of admission, supporting its use as an early-warning aid where continuous invasive monitoring is unavailable.