Early prediction of vasopressor initiation in ICU sepsis patients using an interpretable EHR-based ML model

Abstract

Background

Early identification of septic patients who will require vasopressor support could provide a critical window for hemodynamic optimisation, yet current bedside cues often appear only when shock is imminent.

Objective

We aimed to develop and validate an interpretable electronic health record (EHR)–based machine-learning model that predicts vasopressor initiation several hours before therapy in intensive care unit (ICU) patients with sepsis.

Methods

We conducted a retrospective study using the MIMIC-IV v2.2 database (2008–2019). We screened adult Sepsis-3 ICU stays and labeled the patients who commenced continuous vasopressor infusions 6 to 48 hours after admission as cases; we defined controls as sepsis patients with ICU stays ≥48 hours and no vasopressor exposure. We performed one to one nearest neighbour matching on age, sex, Charlson index, SOFA score with the cardiovascular component removed, weight, and early lactate/hematocrit availability to minimise confounding. We engineered demographic, physiological, and laboratory features measured from −6 to −2 hours relative to vasopressor initiation (or a matched time point) under multiple parameter combinations. We trained seven algorithms with Monte Carlo cross validation and evaluated performance on an independent validation set. We assessed model interpretability with Shapley values.

Results

We analyzed 1,539 cases and 1,431 controls; the independent validation set comprised 751 stays (~25%). A Random Forest classifier achieved an area under the receiver operating characteristic (AUROC) of 0.75 (95% CI, 0.72–0.79), a sensitivity of 0.74 (95% CI, 0.69–0.78), a specificity of 0.65 (95% CI, 0.60–0.70), a precision of 0.70 (95% CI, 0.66–0.74) and a F1 score of 0.72 (95% CI, 0.68–0.75) at the Youden’s index threshold. The model outperformed simple surrogates—mean blood pressure (AUROC, 0.68; 95% CI, 0.64–0.72) and modified shock index (AUROC, 0.65; 95% CI, 0.62–0.69)—and a reproduced bidirectional LSTM (AUROC, 0.73; 95% CI, 0.70–0.77). Key predictors included declining mean blood pressure at − 2 to −4 hours, elevated lactate ( > 2.5 mmol/L), and hematocrit outside 30–37%. Model alerts would occur two to four hours before vasopressor initiation, providing actionable lead time for clinicians.

Conclusions

This proof-of-concept study shows that routinely collected ICU data can predict impending vasopressor initiation with clinically interpretable outputs. However, these findings reflect internal validation only and should be interpreted with caution. External validation on multi-center retrospective cohorts, followed by silent-mode prospective evaluation, is warranted to confirm generalisability and to assess the real-world impact on time-to-vasopressor, fluid balance, and patient outcomes.

Supplementary information

The online version contains supplementary material available at 10.1186/s12911-025-03274-3.

Introduction

Sepsis is a leading cause of mortality in intensive care units (ICUs), characterized by life-threatening organ dysfunction resulting from a dysregulated host response to infection [1]. Despite improvements in critical care, sepsis continues to pose a substantial clinical and economic burden, with an estimated 189 hospital-treated cases per 100,000 person-years and a 90-day mortality rate of 32.2% in high-income countries [2]. In its most severe form, sepsis can progress to septic shock. This condition is defined by persistent hypotension, elevated serum lactate levels ( > 2 mmol/L), and profound vascular hyperpermeability [3, 4]. These pathophysiological changes lead to tissue hypoperfusion, multi-organ failure, and an even higher 90-day mortality rate, reaching 38.5% [2].

Vasopressors—medications used to constrict blood vessels, improve contractility of the heart and elevate blood pressure—are the cornerstone of hemodynamic support in septic shock [5]. Timely initiation of vasopressors is critical, as delays in reversing septic shock can result in irreversible organ damage [6]. However, clinicians often lack reliable early indicators that a sepsis patient’s blood pressure will crash and require vasopressors. The widespread adoption of Electronic Health Records (EHRs) enables the development of predictive algorithms that continuously monitor patient data and provide real-time alerts to clinicians [7]. Such systems could help reduce treatment delays, shorten ICU stays, and mitigate downstream complications [8].

Building on this opportunity, several groups have developed machine-learning systems to forecast impending vasopressor therapy in the ICU. Kwak et al. trained an attention-based bi-directional long short-term memory (LSTM) on hourly vital-sign sequences from Medical Information Mart for Intensive Care (MIMIC)-III and eICU Collaborative Research Database (eICU‑CRD); their cohort simply excluded patients who had already received a vasopressor in the first three hours and required at least 24 h of monitored data, yielding an area under the receiver operating characteristic (AUROC) of 0.83 for 2-hour prediction [9]. Holder et al. built a neural Weibull–Cox survival model in three U.S. hospitals; eligible patients were adults with ICU stays ≥8 h who had not developed shock in the first four hours, and the model achieved AUROCs of 0.80–0.81 when applied to external cohorts across 4- to 24-hour horizons [10]. Although these studies underscore the promise of data-driven prediction, their relatively broad inclusion criteria and the absence of case-control matching on baseline severity or comorbidity leave room for temporal bias, residual confounding, and optimistic discrimination estimates.

The objective of this study was to develop and internally validate an interpretable machine‑learning model using routinely collected EHR data from the MIMIC‑IV database to predict continuous vasopressor initiation several hours before therapy in ICU patients with sepsis.

Materials and methods

Data source

This study utilized retrospective electronic health record data from the MIMIC-IV v2.2 database [11–13]. MIMIC-IV is a large, freely available database comprising deidentified health-related data associated with patients who stayed in care units of the Beth Israel Deaconess Medical Center in Boston, Massachusetts, USA. This database comprises 431,231 hospital admissions from 2008 to 2019.

Study population and cohort creation

We investigated vasopressor initiation in ICU patients with sepsis. Sepsis patients were identified using the Sepsis-3 criteria from the MIT Laboratory for Computational Physiology (MIT‑LCP) GitHub derived table “sepsis 3” [14, 15]. Exclusion criteria were chosen to eliminate clinical contexts in which vasopressor use is driven by mechanisms other than septic vascular leakage and vasoplegia. We therefore removed admissions with (i) recent major surgery (to avoid transient, anaesthesia-related hypotension), (ii) International Classification of Diseases (ICD) codes for haemorrhagic shock (dominant hypovolaemia), and (iii) “do-not-resuscitate” or comfort-measures status (treatment-limitation bias), see Supplementary Table 1.

After applying these exclusions, we focused the case cohort on vasopressor initiations 6–48 h after ICU admission as a proxy for the onset of septic shock (signifying likely vascular leak rather than immediate postoperative hypotension).

The control population comprised admissions that (i) remained in the ICU for ≥ 48 h, permitting alignment of the first 48 h of physiological data with the case observation window, (ii) received no vasopressor therapy during their stay, and (iii) underwent no surgical procedure from 18 h before to 48 h after ICU admission.

To mitigate baseline confounding, each case was matched without replacement to a single control using nearest-neighbor Mahalanobis distance. Matching variables included age, sex, weight, Charlson Comorbidity Index, and the admission-day Sequential Organ Failure Assessment (SOFA) score with the cardiovascular component removed (to avoid circularity, as this sub-score is determined by vasopressor dose). Availability of hematocrit and lactate measurements within the first 48 h was also enforced as a matching criterion to ensure comparable data completeness. Encounters lacking any recorded predictor values within the designated feature window were excluded from analysis. We evaluated the balance between cases and controls using two complementary approaches. First, we compared baseline clinical characteristics using Chi-square tests for categorical variables and Wilcoxon rank-sum tests for continuous variables. Second, we calculated standardized mean differences (SMD) for all matching variables to quantify residual imbalance. An SMD below 0.1 was considered indicative of adequate balance.

Both the CCI and SOFA scores were computed based on MIT-LCP code and the derived tables [14, 15].

Clinical characterization

The clinical characterization of the study population involved the extraction and analysis of various baseline characteristics, parameters at admission, and outcomes. Baseline characteristics included demographic data such as age, gender, weight, height, body mass index (BMI) computed from weight and height when available, and the Charlson Comorbidity Index (according to MIT-LCP GitHub derived table “charlson” [14, 15]). Parameters at admission encompassed clinical scores and indices such as Simplified Acute Physiology Score II (SAPS II) on Day 1 (MIT-LCP GitHub derived table “sapsii” [14, 15]), SOFA scores at different time points (Day 0, Hour 5, and Day 1, according to MIT-LCP GitHub derived table “sofa” [14, 15]), and the Vascular Leak Index (following Chandra et al. [16]). Outcomes measured included the length of ICU stay, length of hospital stay, number of ventilator days, and mortality rates at different time points (14-day, 28-day, and 1-year mortality). Additionally, the rates of ICU discharge alive and in-hospital mortality were also extracted. Continuous data were controlled following min/max filters introduced in Supplementary Table 2.

Features for prediction

A set of features was selected a priori based on existing knowledge with input from two experienced physicians to ensure clinical relevance [1, 16–18]. These features included demographic data (age, gender, weight, height, and Body Mass Index (BMI)), comorbidity data (CCI), severity of illness data (SOFA score 5 hours after ICU start – i.e. before vasopressor initiation that we aimed to predict), vital signs (heart rate (HR), mean blood pressure (MBP), and respiratory rate (RR)), capillary refill time (CRT), laboratory data (albumin, hematocrit, lactate, C-reactive protein (CRP), and total protein), and cardiac output data (measured by thermodilution). Data were collected from −6 to −2 hours before vasopressor initiation, which served as the reference time window. Other time windows were also tested, as described later [18].

We experimented with multiple strategies for handling missing values and summarizing the −6 to −2 h time-series (full details in Supplementary Methods 1). In brief, we varied four parameters: whether to carry forward recent values for missing labs, how to aggregate measurements within the window, how to summarize high-frequency vitals, and which method to use for final imputation of missing entries.

Multiple combinations of these feature engineering parameters were tested to identify the optimal feature set for predicting vasopressor need, based on AUROC performance. We used the machine-learning pipeline described hereafter with 50 Monte Carlo cross-validation and the following three algorithms: Flexible Discriminant Analysis (FDA), Generalized Linear Model with Elastic Net Regularization (GLMNet), and High Dimensional Discriminant Analysis (HDDA). For each of the three methods, the top-performing combinations, defined as those with a difference within 0.03 of the maximum AUROC on the validation set, were retrieved. Then, only the combinations that were common across all three methods were kept. The most frequently occurring parameter options among these were identified as the optimal feature engineering set. After the feature engineering step, we assessed potential redundancy among predictors by quantifying pairwise associations using Spearman’s ρ and screening for multicollinearity using variance inflation factors (VIF) computed on standardized predictors. We pre-specified |ρ| ≥ 0.80 and VIF ≥ 5 as flags for potential problematic redundancy.

Outcome: definition of vasopressor initiation

Vasopressor initiation was defined as the first continuous infusion—recorded in the inputevents table—of any of the following agents: norepinephrine, epinephrine, phenylephrine, vasopressin.

Machine-learning model training and evaluation

Seven machine-learning algorithms from different families are explored for this vasopressor prediction problem, including Multi-Layer Perceptron (MLP) Weight Decay, Random Forest, Naive Bayes, SVM Radial, FDA, GLMNet and HDDA. The cohort was divided into a training set (75%, n = 2,219) and an independent hold-out validation set (25%, n = 751). Monte Carlo cross-validation was performed on the training set using a 70% training and 30% test split, with 100 resamples. Grid search was then applied to the resampled training sets to identify the optimal hyperparameters listed in Supplementary Table 3 maximizing the average AUROC across the corresponding test sets. Finally, all machine-learning algorithms were fitted with optimized hyperparameters on the entire training set and applied to the validation set. The best model was selected based on the highest AUROC on the independent validation set. To monitor overfitting, we compared the median AUROC between training and validation sets for each candidate model and considered a ΔAUROC < 0.03 as acceptable. The probability threshold for final label estimation is chosen by maximizing the Youden’s index on the probabilities estimated on the training set.

Benchmarking against existing literature

Benchmark performance was assessed against three complementary comparators. First, we re-implemented the attention-based bidirectional LSTM model of Kwak et al. [9], adapting the publicly available code of Kaji et al. [19] under Python 3.10.4 because the original implementation was not released; cohort definition, exclusion rules, feature engineering, scaling and train–validation partitioning were kept identical to those used for our primary models. Hourly vital-sign trajectories were linearly interpolated to a 60-min grid, and static covariates were repeated across time steps so that the network, which accepts sequential inputs only, received the complete feature set; network depth, hidden-unit size, batch size and learning-rate schedule followed the original specifications. Second, mean blood pressure (MBP) was extracted at 5, 4, 3 and 2 h before vasopressor initiation and receiver-operating-characteristic curves were generated—MBP provides the single-variable benchmark because a sustained MBP < 65 mm Hg is the guideline threshold that typically prompts vasopressor therapy in septic shock [5]. Third, the modified shock index, defined as heart rate divided by MBP, was calculated at the same four time points and its discrimination characterized in an identical manner. All comparators were evaluated on the same held-out validation cohort, and confidence intervals for the AUROC were estimated with the same bootstrap procedure applied to the Random Forest model, ensuring that performance differences reflected modelling approach rather than sampling variability.

Model explainability

To ensure the explainability of the machine-learning model, we employed Shapley values to provide a unified measure of feature importance, allowing us to understand the contribution of each feature to the model’s individual predictions.

Impact of time-to-event on prediction performance

To extend beyond the default feature collection window of −6 to −2 hours, we aimed to explore different prediction delays. This analysis investigates the effect of varying prediction delays on the performance of our models. Specifically, we repeated the same feature engineering and machine-learning pipeline using data measured from −6 hours before the reference time to different time limits, namely −4 hours, −3 hours, and −1 hour. We compared the AUROC values of the models trained on these different time windows to assess the effect of window size and distance on predicting vasopressor initiation.

Software and tools

All data analyses were conducted using R version 4.1.3 [20].

Results

A visual summary of the study design and main findings is provided in Supplementary Figure 1.

Case and controls cohort description

The MIMIC-IV database comprises 431,231 hospital admissions, of which 66,239 included at least one ICU stay. Among these, 31,206 admissions were identified as sepsis-3 patients. To focus on predicting vasopressor initiation, we excluded stays labeled as “do not resuscitate” and those involving hemorrhage based on ICD codes. This resulted in a total of 22,083 Sepsis-3 stays, among which 1,678 had vasopressor initiation within 6 to 48 hours of ICU admission and without a surgery event within the 24 hours before (cases), see Fig. 1. To create a control cohort, we performed nearest neighbor matching between cases and the 4,803 Sepsis-3 patients who had an ICU stay of at least 48 hours, no surgical events, and no vasopressor administration during their stay. This matching process resulted in 1,585 case–control pairs, of which 1,539 cases and 1,431 controls had more than one predictor measured; these pairs were retained for analysis.

Fig. 1.

Flowchart of cases and controls selection from MIMIC-IV. The following inclusion criteria were applied to construct the study cohort based on ICU stays from the MIMIC-IV v2.2 database: diagnosis of sepsis according to sepsis 3 definition, no ICD codes indicating “do not resuscitate” or hemorrhage (see Supplementary Table 1). Cases (yellow, left box): ICU patients who received vasopressors 6–48 h after admission and had no major surgery within the prior 24 h. Controls (blue, right box): ICU patients ≥48 h length of stay, no vasopressor use, and no surgery within 18 h before/48 h after admission. We performed a 1:1 nearest neighbor matching using mahalanobis distance to obtain the final cases and control cohorts. Abbreviations: DB: database, ICU: intensive care unit, init.: initiation, LOS: length of stay, h: hour, nHadm: number of unique hospital admissions

Table 1 confirms successful matching: there were no significant differences in the matched covariates (marked with * in the table) between cases and controls. Upon admission, the median SAPS II score computed over the first 24 hours was slightly higher for cases (41 [34–52]) compared to controls (36 [29–44]). This difference is expected, as the SAPS II score includes features such as systolic blood pressure, ventilator use, PaO2, and FiO2, which are typically more severe in sepsis patients requiring vasopressor therapy. Notably, cases had substantially higher 28-day mortality (20% vs 8%) and longer ICU stays (5.4 [3.0–9.9] days vs 4.0 [2.8–6.6] days) than controls, reflecting the greater illness severity of those who required vasopressors. Covariate balance was further assessed using SMD, which showed all variables below the 0.1 threshold after matching (Supplementary Figure 2). These clinical characteristics remained consistent even after the training testing split, as shown in Supplementary Table 4.

Table 1.

Cohort baseline and outcomes characteristics

	Case		Control
Patient count	1539 (52%)		1431 (48%)		p.adjust
	n	stat	n	stat
Baseline characteristics
Age*	1539	67.0 [56.0–76.0]	1431	66.0 [56.0–76.0]	0.964
Gender (Female)*	1539	657 (43%)	1431	609 (43%)	0.971
Weight*	1539	80.0 [67.6–96.3]	1430	80.0 [67.4–97.0]	0.798
Height	1205	170.0 [163.0–178.0]	799	170.0 [160.0–178.0]	0.971
BMI	1205	27.9 [24.3–32.7]	799	28.2 [24.4–33.0]	0.459
CCI*	1539	5.0 [3.0–7.0]	1431	5.0 [3.0–7.0]	0.964
Parameters at admission
SAPS II Day 1	1539	41.0 [34.0–52.0]	1431	36.0 [29.0–44.0]	< 0.001
SOFA Day 0	1539	1.0 [0.0–3.0]	1431	1.0 [0.0–2.0]	0.032
SOFA Day 0 w/o Card.*	1539	0.00 [0.00–2.00]	1431	0.00 [0.00–2.00]	0.597
SOFA Hr 5	1539	3.0 [2.0–5.0]	1431	3.0 [2.0–5.0]	0.597
SOFA Day 1	1539	6.0 [4.0–9.0]	1431	4.0 [3.0–6.0]	< 0.001
Vascular Leak Index	1131	−1.0 [−2.4–0.1]	742	−1.2 [−3.7–0.5]	0.241
Outcomes
Length of ICU stay	1539	5.4 [3.0–9.9]	1431	4.0 [2.8–6.6]	< 0.001
Length of hosp. stay	1539	11.4 [6.8–20.2]	1431	9.8 [6.6–15.7]	< 0.001
Vent Days	1262	2.6 [1.0–6.6]	736	2.1 [0.9–4.6]	< 0.001
D14 Mortality	1539	245 (16%)	1431	72 (5%)	< 0.001
D28 Mortality	1539	311 (20%)	1431	117 (8%)	< 0.001
Y1 Mortality	1539	549 (36%)	1431	391 (27%)	< 0.001
ICU Discharge Alive	1539	1279 (83%)	1431	1381 (96%)	< 0.001
In-Hosp. Mortality	1539	295 (19%)	1431	81 (6%)	< 0.001

Data are presented as median [interquartile range] or n (%), and percentages are rounded to the nearest whole number, in “stat” column. Numerical values are rounded to one decimal place. The column ‘n’ indicates the number of non-missing data points used for statistical computation. This table is computed before data imputation, and “p.adjust” column represents the p-value adjusted by the Benjamini Hochberg correction. Parameters used for cohort matching are indicated by a “*”

Abbreviations: BMI: body mass index; CCI: Charlson Comorbidity Index; D: Day; Hr: Hour; Hosp: Hospital; ICU: intensive care unit; SAPS II: Simplified Acute Physiology Score II; SOFA: Sequential Organ Failure Assessment; Vent: ventilator; Y: year; w/o Card.: without cardiovascular component

Prediction performances for feature window [−6 h, −2 h]

Among the candidate predictors, albumin, total protein, and cardiac output were excluded due to < 1% availability. Most remaining variables were present in > 80% of patients, except CRT, hematocrit, and lactate (~35% availability), which often co-occurred as missing (31% lacked all three; Supplementary Figures 3 and 4).

Testing multiple feature-engineering strategies revealed several parameter combinations that achieved AUROC values within 0.03 of the maximum across all algorithms (Supplementary Figure 5). The most frequent options among these top-performing sets were: no forward-filling of sporadic labs (CRT, hematocrit, lactate), aggregation of other vitals/labs by their minimum over 4 hours, retention of hourly values for dense vitals (HR, RR, MBP), and imputation of remaining missing data via Random Forest. After feature engineering, Spearman correlation and VIF analyses (Supplementary Figures 6 and 7) revealed no problematic redundancy among predictors: all pairwise ρ values were below 0.80 and all VIFs were below 5 except for the expected high correlations among hourly measurements of HR, which reflect their temporal structure rather than true multicollinearity.

Prediction of vasopressor initiation

Among the seven explored algorithms, Random Forest performed best on our data set both on training and validation set, with validation AUROC of 0.75 (95% CI: 0.72–0.79), as shown in Fig. 2A, B, and C. The AUROC difference between training and validation sets for this model was 0.027 (0.779 vs. 0.752), which is below the predefined threshold of 0.03 and indicates an absence of substantial overfitting. Using the Youden’s index threshold (optimized on training data, the model achieved a sensitivity of 0.74 (95% CI, 0.69–0.78) and a specificity of 0.65 (95% CI, 0.60–0.70) in validation, corresponding to 103/336 missed cases (false negatives) and 124/415 false alarms (false positives). This equates to a PPV of 0.70 (95% CI, 0.66–0.74), a NPV of 0.69 (95% CI, 0.64–0.74) and a F1 score of 0.72 (95% CI, 0.68–0.75) as shown in Figure 2D and Supplementary Figure 12. Error analysis revealed no major demographic differences across prediction strata (Supplementary Tables 5 and 6). In the validation set, false negatives exhibited slightly lower severity than true positives (median SAPS II 38 [31–45] vs. 42 [33–53]; SOFA Day 1 5 [3–6] vs. 7 [4–9]), whereas false positives resembled true negatives (SAPS II 35 [28–44] vs. 35 [28–42]). Although these differences reached statistical significance (adjusted p < 0.001), the absolute gaps were small. Mortality patterns followed a similar gradient: 28-day mortality was 14% for false negatives and 22% for true positives, compared with 12 and 7% for false positives and true negatives, respectively. These findings suggest that misclassification primarily occurred in patients with intermediate severity rather than at extremes of illness. Calibration on the validation set was acceptable, with observed event rates closely tracking predicted probabilities across deciles and a Brier score of 0.209 (Supplementary Figure 8).

Fig. 2.

Best-performing model for vasopressor initiation prediction. Selection process of the best performing model to predict vasopressor initiation, using features measured from −6 to −2 hours before initiation. A. AUROC performances of 7 ML algorithms, sorted by AUROC on validation sets. Median and 95% CI of 2000 stratified bootstrap resamples AUROC is depicted in yellow for the train set and green for the validation set. B. ROC of the best-performing ML model (random forest) along with sensitivity’s 95% confidence area based on 2000 stratified bootstrap resample. Train set ROC is depicted in yellow, with sensitivity on the ordinate and specificity on the x-axis, and validation set in green. The sensitivity and specificity of the final threshold calculated on the train set by maximizing Youden’s index, is indicated on both curves by a black point. 2000 stratified bootstrap resample median AUROC along with 95% confidence interval are reported in the legend. C. Violin and boxplot of the estimated probabilities (y-axis) of being a case in both training (left) and validation (right) sets grouped by the true status of patients (x-axis, “case” in yellow, “control” in blue). Youden’s index decision threshold is indicated as a dashed horizontal line, and the number of patients per violin is reported on the x-axis. D. Confusion matrix obtained on validation set through the best ML model (random forest). y-axis corresponds to the true status, and x-axis to the predicted status. Count of patients per category (TP, TN, FP, FN) is reported in each case. Abbreviations: AUROC: area under the receiver operating characteristic; CI: confidence Interval; HDDA: high-dimensional discriminant Analysis; FDA: Flexible Discriminant Analysis; GLMnet: generalized linear model net; ML: machine-learning; MLP: multilayer Perceptron; SVM: support vector Machine; TP: true positive; FN: false negative; FP: false positive; TN: true negative

Subsequently, we compared our Random Forest model with the simpler clinical benchmarks and the reproduced model based on attention mechanism and bi-LSTM. MBP and the modified shock index achieved their best discrimination 2 h before vasopressor initiation, with respectively AUROCs of 0.68 (95% CI, 0.64–0.72) and 0.65 (95% CI, 0.62–0.69), see Supplementary Figure 9—both falling short of our Random Forest AUROC of 0.75. The bi-LSTM reached an AUROC of 0.73 (95% CI: 0.70–0.77), closely approaching but still marginally below that of the Random Forest, while demanding substantially greater training time and hyper-parameter tuning (Supplementary Figure 10).

Model explanation

To interpret the best model’s predictions, we applied Shapley values and identified several key features that significantly influence the results (Figs. 3A, 3B). The most important features identified were Mean Blood Pressure (MBP) at various intervals (−2, −3, and −4 hours), Lactate levels, and Hematocrit values. Among these, MBP recorded at different intervals was particularly significant. A high MBP reduces the likelihood of vasopressor initiation, whereas a low MBP increases this likelihood. This trend remained consistent across all time points, with the range of contribution being more pronounced closer to the initiation time and diminishing over time. Lactate levels above 2.5 mmol/L were found to elevate the probability of vasopressor initiation. This finding is consistent with the clinical understanding of septic shock, where elevated lactate levels often necessitate vasopressor support. For Hematocrit, values within the range of 30% to 37% did not significantly affect the contribution, while values outside this range heightened the risk of vasopressor initiation. These findings collectively demonstrate a strong alignment between the model’s predictions and clinical reality, thereby reinforcing the model’s reliability in predicting the probability of vasopressor initiation.

Fig. 3.

Best performing model explanation. This figure explains the performance of the best model (random forest) for the −6 to − 2 hours window, trained on the training set and analyzed using Shapley values. A. Patient example illustrating the contribution of individual features (arrow-box) to the model’s prediction probability of vasopressor initiation. Yellow features increase the probability, while blue features decrease it. The x-axis represents the impact on model output (probability impact is reported within each arrow-box). The patient value of each feature is reported below each arrow-box. The decision threshold (derived using Youden’s index) and the overall patient predicted probability of receiving vasopressor (i.e. being a case) are also reported. B. Hexagon plot showing the association between each feature (y-axis) and their contribution to the probability of vasopressor initiation as derived by Shapley value analysis (x-axis). To highlight the most impactful features, only those with more than 5% contribution based on the ratio of mean absolute shap values are shown. Each hexagon contains a certain number of patients, with the true case percentage within each hexagon reported using a color gradient from blue (0%) to yellow (100%), allowing for visual observation of the association between Shapley-derived contribution of being a case and true status. A black line outlines hexagons with more than 15 samples. Abbreviations: MBP: mean blood Pressure; BMI: Body Mass Index; HR: heart rate; RR: respiratory rate; CRT: capillary refill time; - x h: value summarized in the time interval of (-x-1, -x] hour from vasopressor initiation; FDA: Flexible Discriminant Analysis

Impact of time-to-event on prediction performance

Examining different prediction delays ranging from −6 hours to −1 hour, −6 hours to −3 hours, and −6 hours to −4 hours, we found that the Random Forest model consistently outperforms other models in terms of Validation set’s AUROC, other algorithms followed closely within ~0.05 AUROC, see Supplementary Figure 11. We noted a roughly linear decline in predictive performance as the gap to vasopressor initiation increased: for data up to 1 h before initiation, validation AUROC was ~0.76 (95% CI, 0.73–0.80), which fell to ~0.74 (95% CI, 0.71–0.78) for 3 h before, and ~0.69 (95% CI, 0.65–0.73) for 4 h before (Fig. 4A and Supplementary Figure 12). Not surprisingly, the MBP value closest to vasopressor start (e.g., at −1 h for the shortest window, or at −2 h for the standard window) was the dominant feature in all scenarios (see Fig. 4B). Lactate and hematocrit remained the next most influential features across windows, though their relative ranking swapped in some cases. Other variables (such as earlier MBP readings, heart rate, etc.) contributed progressively less at longer horizons.

Fig. 4.

Prediction performance and feature importance across different time windows. This figure re-applies the same pipeline of feature engineering and prediction across different measurement windows (−6 h to −4 h, −6 h to −3 h, −6 h to −2 h – the original window of interest, and −6 h to −1 h from vasopressor initiation). Only the performance (A) and model explanation (B) of the best performing model are reported per time window. A. performance of the best models per measurement windows. vasopressor initiation, using features measured from −6 to − 2 hours before initiation. Median and 95% CI of 2000 stratified bootstrap resample’s AUROC is depicted in yellow for the train set and green for the validation set. B. The relative contribution of each feature, based on their mean absolute SHAP values, is displayed in a bar plot. Only features with a mean absolute Shapley value greater than 0 are included. Those contributing less than 5% are combined and labeled as “others”. Features are listed from top to bottom in order of decreasing percentage, except for the “others” category, which is always shown at the bottom. Abbreviations: AUROC: area under the receiver operating characteristic; h: hours; MBP: mean blood pressure; BMI: Body Mass Index; HR: heart rate; RR: respiratory rate; CRT: capillary refill time

Discussion

This study aimed to determine whether routinely collected ICU data could anticipate vasopressor initiation in septic patients. By leveraging a rigorously matched cohort and interpretable machine‑learning techniques, we aimed to address limitations in prior studies and provide a more realistic estimate of predictive performance. Using MIMICIV, we constructed a tightly defined Sepsis-3 cohort with exclusions for perioperative, hemorrhagic and “do not resuscitate” contexts, and implemented 1:1 nested riskset matching on age, sex, admission day SOFA (cardiovascular component removed), Charlson index and early biomarker availability; applying an identical −6 h lookback window to cases and controls emulated real-time conditions and helped reduce temporal/selection bias [21, 22]. Within this framework, a Random Forest classifier achieved a validation AUROC of 0.75, outperforming simple hemodynamic surrogates (MBP and modified shock index) and very slightly surpassing the attention-based bi-LSTM we reproduced from the architecture described by Kwak et al. (AUROC 0.73). These findings suggest that interpretable models leveraging routinely available clinical data can provide an actionable lead‑time of approximately 2–4 hours before hemodynamic collapse [23]. Importantly, this predictive window aligns with a critical period for clinical intervention, suggesting that such models could complement existing bedside assessments rather than replace them.

While our results are promising, they must be interpreted in the context of prior work. Recent high profile sepsis alert systems such as TREWS [24] and COMPOSER [25, 26] and the FDA authorized Prenosis Sepsis ImmunoScore [7], have demonstrated the promise of AI driven alerts to improve outcomes when embedded in clinical workflows. However, they target early sepsis recognition, not vasopressor initiation, which is our specific endpoint. By contrast, the few vasopressor focused studies—such as an attention based bi LSTM [9] (AUROC ≈ 0.83 at 2 h) and a Weibull–Cox survival model [10] (AUROC ≈ 0.80–0.81 over 4–24 h)—used broad inclusion criteria without controlling baseline severity, making them prone to temporal bias and overly optimistic performance estimates. Collectively, our matched design, tightly specified cohort and objective EHR endpoint (first continuous vasopressor infusion) therefore provide a more conservative—and likely more reliable—estimate of real-world performance for a vasopressor initiation warning score.

Beyond performance, interpretability remains essential for clinical adoption. Consistent with clinical understanding, our model’s top features – dropping MBP, elevated lactate, and deviant hematocrit – mirror known hallmarks of septic shock [27]. Lactate is a classic marker of hypoperfusion that predicts vasopressor need and mortality [28]. Low MBP reflects the vasoplegia of incipient shock. It is not surprising that dropping blood pressure was a dominant predictor; this aligns with clinical reasoning. However, our model’s incorporation of multiple features aims to alert clinicians to impending shock a bit earlier or in less obvious cases than a blood pressure monitor alone might, which could still confer practical benefit. Hematocrit outside 30–37% is intriguing as an independent risk factor; this corroborates Chandra et al.’s finding of hemoconcentration correlating with capillary leak and fluid accumulation [16], highlighting the often under-recognized role of endothelial permeability in sepsis (also noted in recent reviews [29]). Our results thus reaffirm these physiological insights, now backed by a data-driven model. While the importance of hematocrit is not entirely new, our model independently identified it from EHR data, lending support to routine monitoring of hematocrit as part of sepsis prognostication.

From an implementation standpoint, EHR inputs arrive passively, allowing the algorithm to run continuously and surface color-coded alerts with concise SHAP-based explanations (e.g., MBP 58 mmHg ↓, Lactate 3.1 mmol/L ↑). This early alert on future vasopressor need would leave time to place arterial lines, titrate fluids, or prepare norepinephrine. Early vasopressor initiation has been tied to faster shock control and improved survival, indicating that even modest lead-time could be clinically meaningful [6]. However, the practical impact of such alerts depends not only on timeliness but also on how the model behaves when it errs. If implemented in clinical workflows, understanding misclassification patterns will be critical to anticipate operational and patient safety implications. At the Youden-index threshold, our model would have generated 103/336 false negatives and 124/415 false positives in the validation set (sensitivity 0.70; specificity 0.68; PPV 0.71; NPV 0.66). In a real-world setting, false negatives would not override standard-of-care decisions, which remain clinician-driven, but they could limit opportunities for earlier preparatory actions such as intensified monitoring or timely reassessment before conventional triggers are reached. These cases tended to show less pronounced hemodynamic instability and slightly lower SAPS II scores than true positives, suggesting that patients with more indolent trajectories or partial stabilization might be harder to detect by an algorithm prioritizing recent MBP decline and elevated lactate. Conversely, false positives—often associated with transient hypotension or elevated lactate—might increase alert burden and preparation costs without leading to vasopressor initiation, possibly because of effective fluid resuscitation or clinician discretion. While such alerts could be perceived as noise, they might also serve as early prompts for vigilance in borderline cases. These considerations highlight that predictive models should be deployed as probabilistic decision-support tools rather than deterministic triggers. Any future implementation should include governance mechanisms—such as tiered alerting, dynamic thresholds, and clinician confirmation steps—to mitigate alert fatigue while preserving potential benefits [30–32]. This perspective aligns with evidence from prospective systems like TREWS, where clinical impact depended on workflow integration and provider engagement rather than algorithmic accuracy alone [24].

Despite these strengths, several limitations warrant discussion. First, the retrospective design precludes causal inference. Second, absence of vascular-leak biomarkers (albumin, total protein) exceeded 99%; hypo-albuminaemia is strongly associated with worse sepsis outcomes [33] and endothelial dysfunction, so richer data capture may uncover additional predictors. Third, the model was developed and validated exclusively on MIMIC-IV, a single-center dataset; lack of external validation limits generalizability, as inter-hospital practice differences can erode model performance; Holder et al. [10] demonstrated that domain adaptation was necessary to maintain performance across health systems. Fourth, the algorithm treats the 6-h observation window passively: it does not adjust for contemporaneous resuscitative actions—fluid boluses, vasoactive pushes, ventilator or sedation changes—that can transiently normalise vital signs and alter clinicians’ threshold for starting continuous vasopressors. Finally, our imputation strategy implicitly assumes a Missing‑At‑Random mechanism, yet certain laboratory tests (e.g., lactate, hematocrit) may be Missing‑Not‑At‑Random (MNAR) because they are often ordered only in sicker patients. Although we compared multiple preprocessing options and selected the configuration that generalized best to validation, MNAR mechanisms could still introduce bias. To partially mitigate this risk, we required availability of lactate and hematocrit in the matching process, which reduces systematic differences in data completeness between cases and controls and limits the confounding effect of indication‑driven missingness.

The next phase of this work should follow a stepwise approach, beginning with external validation on retrospective cohorts from multiple institutions to confirm generalizability and detect performance drift, and progressing to silent-mode prospective validation, where predictions run in the background without influencing clinical decisions, to evaluate real-world feasibility and alert burden before any clinical deployment.

Conclusion

We provide a transparent, methodologically robust framework for predicting vasopressor need in sepsis. This work should be considered a proof-of-concept rather than a deployable clinical tool. The model integrates a matched-cohort design, SHAP-based interpretability, and competitive accuracy, offering a credible step toward decision-support systems that could shorten time-to-vasopressor, limit fluid overload, and improve outcomes in critical-care sepsis management. However, the AUROC of 0.75 should be interpreted with caution as it reflects internal validation only. Future work must include external validation on retrospective multi-center cohorts, followed by silent-mode prospective evaluation, to confirm generalizability and real-world feasibility; future deployment would also require a formal model update strategy, such as periodic retraining on recent data with drift monitoring, to maintain performance over time.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviations

AUROC

Area Under the Receiver Operating Characteristic

AUPRC

Area Under the Precision-Recall Curve

BMI

Body Mass Index

CCI

Charlson Comorbidity Index

CRT

Capillary Refill Time

CRP

C-Reactive Protein

eICU-CRD

eICU Collaborative Research Database

FDA

Flexible Discriminant Analysis

FN

False Negative

FP

False Positive

F1 Score

Harmonic Mean of Precision and Recall

GLMNet

Generalized Linear Model with Elastic Net Regularization

HDDA

High Dimensional Discriminant Analysis

HR

Heart Rate

ICD

International Classification of Diseases

ICU

Intensive Care Unit

MBP

Mean Blood Pressure

MIMIC-IV

Medical Information Mart for Intensive Care IV

MLP

Multi-Layer Perceptron

NPV

Negative Predictive Value

PPV

Positive Predictive Value

RR

Respiratory Rate

SAPS II

Simplified Acute Physiology Score II

SOFA

Sequential Organ Failure Assessment

SVM

Support Vector Machine

TN

True Negative

TP

True Positive

VIF

Variance Inflation Factor

Author contributions

Conceptualization, A. V., E.D., M. E., J. F., J.-F. L., and M. B.; Methodology, L.D., A. V., F. Z., G. T., S. B., and M.B.; Data Analysis & Machine-learning, L. D., F. Z., G. T., S. B. and M.B.; Investigation, L. D., A.V., F. Z., G. T., S. B., J.F., J.-F.L. and M.B.; Writing – Review & Editing, L.D., A.V, and M.B.

Funding

N/A.

Data availability

The data supporting the results of this manuscript are derived from the MIMIC-IV database. The MIMIC-IV database is publicly available and can be accessed at https://physionet.org/content/mimiciv/2.2/

Declarations

Human ethics and consent to participate

This study was conducted using the MIMIC-IV database, which comprises de-identified health-related data from patients admitted to the Beth Israel Deaconess Medical Center. The use of the MIMIC-IV database for research purposes was approved by the Institutional Review Board (IRB) of the Massachusetts Institute of Technology and the Beth Israel Deaconess Medical Center, with a waiver of informed consent. All methods were carried out in accordance with relevant guidelines and regulations, including the Declaration of Helsinki. No direct patient intervention or contact occurred, and all data were fully anonymized prior to access.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.