Abstract
This study aimed to develop an interpretable machine learning-based scoring system for predicting sepsis and septic shock among febrile patients at emergency department (ED) triage using longitudinal data. This retrospective, single-center study included adult patients, presented to ED of tertiary academic hospital with fever from January 2016 to December 2021. Using the AutoScore framework, we developed a novel scoring system for predicting sepsis and septic shock at the triage stage, incorporating nine variables and a maximum score of 29. The predictive performance of our score was assessed by calculating the area under the receiver operating characteristic curve (AUROC), and its performance was compared with that of two existing scoring systems: the quick Sequential Organ Failure Assessment (qSOFA) and the Modified Early Warning Score (MEWS). Our model incorporated nine variables including initial vital signs, age, baseline platelet count, total bilirubin, and creatinine levels. Among these, initial systolic blood pressure was identified as the most important predictor. AUROC of our model was 0.844 (95% confidence interval [CI], 0.812–0.875) in predicting septic shock and 0.703 (95% CI, 0.687–0.720) for sepsis. Compared to qSOFA ≥ 2 (AUROC: 0.605) and MEWS ≥ 5 (AUROC: 0.678), our scoring system demonstrated superior predictive performance for septic shock. For comparable specificity levels (ranging from 0.50 to 0.95), our scoring system achieved higher sensitivity than MEWS. Our scoring system is an interpretable and practical scoring tool for predicting sepsis and septic shock among patients with documented fever at ED triage using patient’s longitudinal data.
Supplementary Information
The online version contains supplementary material available at 10.1007/s10916-026-02443-9.
Keywords: Machine learning, Septic shock, Sepsis, Emergency medicine, Triage
Introduction
Sepsis and septic shock remain leading causes of high mortality among patients presenting to emergency department (ED) [1, 2]. Early recognition and timely initiation of treatment are critical determinants of patient outcomes [1, 3]. Traditional scoring systems such as the Sequential Organ Failure Assessment (SOFA) score are widely used to assess organ dysfunction, but they rely on laboratory results, which may delay early identification of sepsis [4, 5]. As a result, there is increasing interest in screening tools that leverage data readily available at the triage stage [6, 7].
At present, quick Sequential Organ Failure Assessment (qSOFA) and Modified Early Warning Score (MEWS) are widely adopted for triage-level severity assessment [3, 8]. While qSOFA has benefit for its simplicity, its predictive accuracy for septic shock is limited [9, 10]. MEWS, designed for detecting clinical deterioration, is not sepsis-specific [11]. Numerous studies have highlighted the limitations of these scores and explored machine learning (ML)-based approaches that have shown improved performance in predicting sepsis and septic shock [12–15].
However, most existing tools—including ML models—are based on data from a single time point at ED presentation and often did not incorporate longitudinal patient data [16]. This is a critical gap, as recent studies have reported that patients with underlying chronic conditions have a higher risk of developing sepsis and progressing to septic shock [17–19]. Despite the fact that many patients revisit the same hospitals where prior data is accessible, few models integrate historical health data to enhance early sepsis prediction in the ED.
In this study, we developed a novel interpretable ML-based scoring system – Septic Shock Index at Triage score (SSIT) – to predict sepsis and septic shock among patients with documented fever at the triage stage using both current and longitudinal patient data. Although sepsis can occur without fever, febrile presentations represent a clinically meaningful subgroup, and thus this study specifically targeted patients presenting with documented fever at triage [20, 21]. We further compared the performance of this model with conventional scoring systems to evaluate its effectiveness in real-world clinical settings.
Materials and methods
Study Design and Data Source
This study is a single-center retrospective study to develop a predictive scoring system for sepsis and septic shock for the patients who presented fever when they arrived at ED. Data were extracted from the clinical database of the authors’ institution, a tertiary academic hospital in South Korea with an average of 200 ED patient visits per day. The data were de-identified to ensure sufficient anonymization. This study was approved by Institutional Review Board (IRB) of our institution (IRB No. 2023-06-073). Informed consent was waived due to the retrospective nature of the study.
Study Population
Adult patients (aged ≥ 19) presenting with fever (≥ 38.0 °C) at ED triage between January 1, 2016 to December 31, 2021 were included. We excluded patients with trauma related visits, those in cardiac arrest, and those who were dead on arrival. Fever was defined as a recorded tympanic temperature ≥ 38.0 °C at ED triage. We selected this threshold because ≥ 38.0 °C is commonly adopted in ED practice and prior sepsis screening studies [22, 23]. This cutoff provides a clinically actionable cutoff rather than physiological definition of fever. The dataset was then randomly split into non-overlapping sets of training set (70%), validation set (10%) and testing set (20%).
Study Outcomes and Candidate Variables
The target outcome was the development of sepsis or septic shock within 24 hours from ED arrival. Sepsis was defined according to the Sepsis-3 criteria as suspected infection accompanied by an acute increase of two or more points in the total SOFA score [3]. Septic shock was defined as persistent hypotension requiring vasopressors despite adequate fluid resuscitation, along with a serum lactate level ≥ 2 mmol/L, in addition to meeting the criteria for sepsis [3].
Candidate variables for model derivation included patient demographics, initial vital signs (systolic blood pressure [SBP], diastolic blood pressure [DBP], mean blood pressure [MBP], pulse rate [PR], respiratory rate [RR], body temperature [BT], and peripheral oxygen saturation [SPO2]), initial mental status, mode of arrival (Emergency Medical Service, private ambulance, walk-in, others and unknown), underlying medical conditions identified using International Classification of Disease, 10th revision codes, and the most recent laboratory test results within one year prior to ED visit. These baseline laboratory values included white blood cell count (WBC), hemoglobin level (Hb), platelet count (PLT), total bilirubin (TB), C-reactive protein (CRP), and creatinine (Cr).
Missing values were imputed according to type of values. Missing values in the triage stage were imputed using multiple imputation by chained equations, whereas missing baseline laboratory values were replaced with lower-to-mid values within the normal reference range of our institution, consistent with SOFA score framework. Initial laboratory test results obtained after ED arrival were excluded, given that the model was designed for use at the triage stage—before laboratory data become available. The frequency and distribution of missing data requiring imputation are summarized in Supplementary Table 1.
Autoscore Framework for Model Derivation
We used Autoscore Framework, a machine learning based scoring generator, composed of 6 modules based on clinical data. Developed by Xie et al. in Singapore, this framework includes modules for variable selection, generation of a point-based scoring model, clinical fine-tuning, and model validation [24]. In Module 1, random forest was applied to rank all candidate variables according to their feature importance, using 100 decision trees. A parsimony plot of the area under the receiver operating characteristic curve (AUROC) was then generated to identify the optimal set of top-ranking variables. Module 2 was employed to transform continuous variables into categorical ones for better interpretability. Modules 3 and 4 were used to develop a point-based scoring model using logistic regression and to determine the number of variables to include in the model. Considering previous studies and clinical settings, researchers selected the final variables and tuned the intervals of continuous variables in Module 5. In Module 6, we evaluated the model performance with a test dataset. Detailed information about the Autoscore algorithm can be found in the original Autoscore publications [24, 25].
Performance Evaluation
We evaluated model performance by calculating AUROC with the test dataset. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were obtained by analyzing AUROC. The optimal threshold for SSIT was derived from Youden. Regarding the effectiveness of our SSIT model, we compared the sensitivity, PPV and NPV at the same specificity with conventional index as qSOFA and MEWS score with testing dataset.
We performed Decision Curve Analysis (DCA) to assess the clinical utility and net benefit of the SSIT score across various threshold probabilities.
Conventional Index
We compared the SSIT performances with qSOFA, and MEWS scoring system. qSOFA is scoring system composed by 3 criteria (RR > 22, presence of altered mental status, and SBP < 100). By meeting two out of the three diagnostic criteria, poor prognosis is predicted in sepsis [3]. MEWS consists of 5 variables, each scoring between 0 and 3: RR, mental status (accessed using the Alert-Verbal-Pain-Unresponsive system), SBP, BT, and PR. A total score of 5 or more indicates a high risk of catastrophic deterioration [8]. In line with previous studies, a qSOFA score ≥ 2 and a MEWS score ≥ 5 were used as cutoff points to compare the performance of the scoring systems.
Statistical Analysis
All continuous and categorical variables were described as a mean with standard deviations (SD) or median with interquartile range (IQR), as appropriate, and as a number with percentages. A normality test was performed to compare the baseline characteristics of three groups (sepsis, septic shock, and others). If the data followed a normal distribution, one-way ANOVA was used; if not, the Kruskal-Wallis test was used. A two-sided p-value < 0.05 was considered statistically significant for all statistical analyses. The statistical analysis was performed using R Software (version 4.1.2, the R Foundation for Statistical Computing, Vienna, Austria, https://www.R-project.org).
Results
Study Population
A total of 415,914 patients were screened during the study period. We excluded 78,635 patients aged < 19 years, 69,420 with injury-related visits, 14,484 who visited the ED for a medical certificate or left without being seen, 580 who were dead on arrival (DOA), 347 with cardiac arrest, and 226,064 whose initial BT at triage was < 38.0 °C. Consequently, a total of 26,384 febrile adult patients were included in the final dataset. Among them, 5,238 (19.9%) were diagnosed with sepsis, and 887 (3.4%) with septic shock (Fig. 1).
Fig. 1.
Patient selection diagram. Among 415,914 patients who presented to ED, 26,384 were included to the study. Of these, 887 had septic shock, 5,238 had sepsis, and 20,259 presented with fever without sepsis. The 26,384 patients were randomly split into three groups: 70% for the training set, 10% for the validation set, and 20% for the testing set. Abbreviation: ED, emergency department
Patients with sepsis and septic shock were older (median age 63 [IQR, 53–73] and 66 [58–74], respectively) and more likely to be male (58.2% and 61.3%) than those with fever only (55 [38–66], 44.4%). At ED triage, sepsis and septic shock patients were less likely to be mentally alert, and had lower SBP, SpO2, but higher PR and RR, than patients with fever only. As clinical severity increased, the prevalence of underlying cardiovascular, cancer, and pulmonary disease increased, while no significant difference was observed in the prevalence of underlying neurological disease. Laboratory findings showed decreasing level of baseline Hb, PLT, WBC, and increasing level of CRP, lactate, and Cr with greater severity of infection. Both MEWS and qSOFA scores increased with clinical severity, supporting the observed physiological progression across groups (Table 1).
Table 1.
Baseline characteristics of patients
| Characteristics | Others (n = 20,259) | Sepsis (n = 5,238) |
Septic shock (n = 887) |
p-value |
|---|---|---|---|---|
| Age, median [IQR] | 55 [38;66] | 63 [53;73] | 66 [58;74] | < 0.005 |
| Male, n (%) | 8,999 (44.4%) | 3,049 (58.2%) | 544 (61.3%) | < 0.005 |
| Initial vital sign at ED, median [IQR] | ||||
| Systolic blood pressure | 123 [109;138] | 118 [101;137] | 106 [88;128.5] | < 0.005 |
| Diastolic blood pressure | 74 [65;84] | 69 [58;80] | 60 [52;74] | < 0.005 |
| Pulse rate | 107 [95;119] | 110 [96;124] | 120 [104;137] | < 0.005 |
| Respiratory rate | 18 [18;20] | 20 [18;20] | 20 [18;24] | < 0.005 |
| Body Temperature | 38.5 [38.2;39.0] | 38.6 [38.2;39.1] | 38.8 [38.3;39.3] | < 0.005 |
| SpO2 | 98 [96;99] | 97 [95;98] | 96 [94;98] | < 0.005 |
| Initial Mental status, n (%) | < 0.005 | |||
| Awake | 19,950 (98.5%) | 4,912 (93.8%) | 772 (87.0%) | |
| Verbal | 106 (0.5%) | 121 (2.3%) | 50 (5.6%) | |
| Pain | 9 (0.0%) | 25 (0.5%) | 6 (0.7%) | |
| Unresponsive | 194 (1.0%) | 180 (3.4%) | 59 (6.7%) | |
| Mode of arrival, n (%) | < 0.005 | |||
| EMS | 1,812 (8.9%) | 820 (15.7%) | 222 (25.0%) | |
| Hospital or private ambulance | 1,237 (6.1%) | 816 (15.6%) | 216 (24.4%) | |
| Others | 17,210 (84.9%) | 3,602 (68.8%) | 449 (50.6%) | |
| Underlying medical conditions, n (%) | ||||
| Cardiology disease | 3,406 (16.8%) | 1,229 (23.5%) | 231 (26.0%) | < 0.005 |
| Cancer | 11,356 (56.1%) | 2,431 (46.4%) | 345 (38.9%) | < 0.005 |
| Neurology disease | 65 (0.3%) | 25 (0.5%) | 3 (0.3%) | 0.234 |
| Pulmonary disease | 1,026 (5.1%) | 364 (6.9%) | 63 (7.1%) | < 0.005 |
| Baseline laboratory findingsa, median [IQR] | ||||
| White blood cell, x 10³/µL | 7.9 [5.7;11.8] | 7.9 [3.3;12.6] | 7.0 [ 2.0;13.5] | < 0.005 |
| Hemoglobin, g/dL | 12.4 [10.7;13.4] | 11.2 [9.4;13.2] | 10.6 [8.9;12.2] | < 0.005 |
| Platelet, x 10³/µL | 217 [162;260] | 123 [67;191] | 114 [52;198] | < 0.005 |
| Total bilirubin, mg/dL | 0.6 [0.4; 0.9] | 1.2 [0.6; 2.0] | 1.1 [0.7; 2.2] | < 0.005 |
| CRP, mg/dL | 3.0 [0.6; 8.2] | 7.3 [2.6;15.2] | 10.3 [3.6;20.2] | < 0.005 |
| Creatinine, mg/dL | 0.8 [0.6; 0.9] | 0.9 [0.7; 1.3] | 1.1 [0.8; 1.7] | < 0.005 |
| Lactic acid, mmol/L | 1.3 [0.8; 1.8] | 1.7 [1.2; 2.4] | 3.4 [2.4; 5.2] | < 0.005 |
| MEWS score, median [IQR] | 3 [2; 5] | 4 [3; 6] | 6 [4; 7] | < 0.005 |
| qSOFA score, median [IQR] | 0 [0; 0] | 0 [0; 1] | 1 [0; 1] | < 0.005 |
aBaseline results were defined as laboratory results obtained within one year prior to ED visit. If unavailable, normal values were imputed for scoring
IQR interquartile range, ED emergency department, SpO2 pulse oximetry value, EMS emergency medical service, CRP C-reactive protein, MEWS Modified Early Warning Score, qSOFA quick Sequential Organ Failure Assessment
SSIT Scoring Systems
For the development of the scoring system, nine variables were selected: age, initial values of SBP, DBP, PR, RR, and BT, and baseline values of platelet count, creatinine, and bilirubin levels (Table 2). Variable selection was guided by feature importance pre-determined by clinicians and model parsimony, as visualized in the parsimony plot (Supplementary Fig. 1). Cutoff values for each variable were optimized for sepsis prediction using the Youden index within the AutoScore framework. The SSIT score ranges from 0 to 29. The SSIT scoring system is presented in Table 2.
Table 2.
SSIT score for septic shock prediction among patients with documented fever at ED triage
| Variables | score | Variables | score |
|---|---|---|---|
| Initial systolic blood pressure, mmHg | Initial body temperature, °C | ||
| < 90 | 5 | < 39.0 | 0 |
| 90–103 | 2 | 39.0–39.4 | 1 |
| ≥ 104 | 0 | ≥ 39.5 | 2 |
| Initial pulse rate, bpm | Baseline creatinine, mg/dLa | ||
| < 120 | 0 | < 2.0 | 0 |
| 120–139 | 2 | ≥ 2.0 | 1 |
| ≥ 140 | 4 | Baseline total bilirubin, mg/dLa | |
| Baseline platelet count x 10³/µLa | < 2.0 | 0 | |
| < 150 | 2 | 2.0–2.9 | 2 |
| 150–399 | 0 | ≥ 3.0 | 4 |
| ≥ 400 | 1 | Initial respiratory rate, rpm | |
| Initial diastolic blood pressure, mmHg | < 16 | 1 | |
| < 60 | 3 | 16–19 | 0 |
| 60–79 | 1 | 20–23 | 1 |
| ≥ 80 | 0 | ≥ 24 | 3 |
| Age (years) | |||
| < 40 | 0 | ||
| 40–59 | 3 | ||
| ≥ 60 | 5 |
a Baseline results were defined as laboratory results obtained within one year prior to ED visit. If unavailable, normal values were imputed for scoring
SSIT Septic Shock Index at Triage, ED emergency department, bpm beats per minutes, rpm rates per minutes
Model Performance
Table 3 summarizes the predictive performance of the SSIT, qSOFA, and MEWS scoring systems. For septic shock prediction, the SSIT score (cutoff ≥ 9) showed the highest AUROC at 0.844 (95% CI, 0.812–0.875), with a sensitivity of 0.807 (0.742–0.865) and specificity of 0.734 (0.722–0.746). In comparison, MEWS ≥ 5 had an AUROC of 0.678 (0.658–0.697), and qSOFA ≥ 2 had the lowest AUROC of 0.605 (0.585–0.625) (Fig. 2a).
Table 3.
Model performance for AUC value in predicting sepsis and septic shock compared to the MEWS and qSOFA
| Outcome | AUROC (95% CI) | Sensitivity (95% CI) | Specificity (95% CI) | |
|---|---|---|---|---|
| Septic Shock prediction | qSOFA ≥ 2 | 0.605 (0.585, 0.625) | 0.241 (0.182, 0.301) | 0.968 (0.963, 0.973) |
| MEWS ≥ 5 | 0.678 (0.658, 0.697) | 0.703 (0.626, 0.768) | 0.652 (0.640, 0.665) | |
| SSIT ≥ 9 | 0.844 (0.812, 0.875) | 0.807 (0.742, 0.865) | 0.734 (0.722, 0.746) | |
| Sepsis Prediction | qSOFA ≥ 2 | 0.544 (0.535, 0.552) | 0.104 (0.088, 0.121) | 0.983 (0.979, 0.987) |
| MEWS ≥ 5 | 0.576 (0.568, 0.585) | 0.476 (0.448, 0.506) | 0.676 (0.661, 0.690) | |
| SSIT ≥ 7 | 0.703 (0.687, 0.720) | 0.713 (0.688, 0.739) | 0.570 (0.555, 0.586) |
AUC area under the curve, MEWS Modified Early Warning Score, qSOFA quick Sequential Organ Failure Assessment, AUROC area under the receiver operating characteristics, CI confidence interval, SSIT Septic Shock Index at Triage
Fig. 2.
Receiver operating characteristics (ROC) curve for the different prediction scoring systems on test data set: (a) for predicting septic shock (b) for predicting the sepsis (a) ROC curves were plotted using predefined cutoffs for MEWS ≥ 5 (AUC 0.68) and qSOFA ≥ 2 (AUC 0.60), and the optimal cutoff for SSIT that yielded the maximum AUC (≥ 9, AUC 0.84) for predicting septic shock (b) ROC curves were plotted using predefined cutoffs for MEWS ≥ 5 (AUC 0.58) and qSOFA ≥ 2 (AUC 0.54), and the optimal cutoff for SSIT that yielded the maximum AUC (≥ 7, AUC 0.70) for predicting sepsis Abbreviations: ROC, receiver operating characteristics; AUROC, area under the receiver operating characteristic curve; AUC, area under curve; MEWS, Modified Early Warning Score; SSIT, Septic Shock Index at Triage; qSOFA, quick Sequential Organ Failure Assessment
For sepsis prediction, the SSIT score (cutoff ≥ 7) outperformed the other tools with an AUROC of 0.703 (0.687–0.720), sensitivity of 0.713 (0.688–0.739), and specificity of 0.570 (0.555–0.586). MEWS ≥ 5 demonstrated an AUROC of 0.576 (0.568–0.585), and qSOFA ≥ 2 showed limited discriminatory power with an AUROC of 0.544 (0.535–0.552) (Fig. 2b).
Overall, the SSIT score demonstrated superior performance in predicting both sepsis and septic shock compared to existing triage-level scores.
To compare performance at equivalent specificity levels, thresholds of SSIT and MEWS were matched across predefined specificity cutoff points (0.5, 0.6, 0.7, 0.8–0.9, and 0.95). As shown in Table 4, the SSIT consistently demonstrated higher sensitivity than MEWS across all specificity cutoffs while maintaining comparable negative predictive values (NPV).
Table 4.
Comparison of accuracy of SSIT and MEWS for septic shock with similar specificity points on test data set
| Specificity (cutoffs) | Threshold | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|
| 0.5 | SSIT ≥ 7 | 0.916 | 0.520 | 0.055 | 0.995 |
| MEWS ≥ 4 | 0.832 | 0.484 | 0.064 | 0.990 | |
| 0.6 | SSIT ≥ 8 | 0.877 | 0.623 | 0.067 | 0.994 |
| MEWS ≥ 5 | 0.703 | 0.652 | 0.058 | 0.986 | |
| 0.7 | SSIT ≥ 9 | 0.807 | 0.734 | 0.084 | 0.992 |
| MEWS ≥ 6 | 0.548 | 0.836 | 0.091 | 0.984 | |
| 0.8–0.9 | SSIT ≥ 12 | 0.510 | 0.913 | 0.151 | 0.984 |
| MEWS ≥ 7 | 0.348 | 0.931 | 0.132 | 0.979 | |
| 0.95 | SSIT ≥ 21 | 0.445 | 0.941 | 0.185 | 0.982 |
| MEWS ≥ 11 | 0.200 | 0.975 | 0.196 | 0.976 |
SSIT Septic Shock Index at Triage, MEWS Modified Early Warning Score, PPV positive predictive value, NPV negative predictive value
In the DCA, the SSIT model demonstrated a superior net benefit compared to both ‘treat all’ and ‘treat none’ strategies across a wide range of threshold probabilities for both septic shock and sepsis (Fig. 3). For septic shock (Fig. 3a), the net benefit remained positive even at low threshold probabilities, with the optimal threshold (Score ≥ 9, probability = 0.031) positioned within the clinically relevant range. A similar trend of clinical utility was observed for sepsis prediction (Fig. 3b) at the optimal threshold (Score ≥ 7, probability = 0.206).
Fig. 3.
Decision curve analysis of SSIT for predicting septic shock (a) and sepsis (b) (a) For septic shock, SSIT showed higher net benefit than ‘treat all’ and ‘treat none’ strategies across the displayed threshold probability range. A SSIT score of ≥ 9 corresponded to a threshold probability of 0.031. (b) For sepsis, SSIT showed higher net benefit than ‘treat all’ and ‘treat none’ strategies across the displayed threshold probability range. A SSIT score of ≥ 7 corresponded to a threshold probability of 0.206. Abbreviation: SSIT, Septic Shock Index at Triage
Discussion
In this study, we derived sepsis and septic shock prediction model, SSIT, based on information achievable at ED initial triage stages incorporating the patient’s longitudinal data before they arrived at the ED. Importantly, this model was developed specifically for patients with documented fever (≥ 38.0 °C) at triage. This cohort-based design allowed targeted prediction rather than the entire ED population. SSIT demonstrated superior predictive performance for septic shock and for sepsis, comparing with qSOFA ≥ 2 or MEWS ≥ 5. This advantage became even more prominent when sensitivity was compared at similar specificity levels [26]. Furthermore, our DCA results provide evidence of the clinical utility of the SSIT score. By showing a higher net benefit than traditional strategies, the model supports more informed decision-making at ED triage, allowing for the early identification of high-risk patients while minimizing unnecessary interventions.
Although longitudinal laboratory data often being available for patients with chronic illness who regularly visit same hospital for outpatient care, most triage models fail to incorporate this information. The SSIT model addresses this gap by integrating historical laboratory data to enhance the predictive performance at triage. This study underscores the clinical value of utilizing pre-existing patient data for early identification of septic shock.
SSIT, the prediction tool specifically focused on sepsis and septic shock with documented fever at triage, is applicable in ED due to its clinical convenience. It allows clinicians to predict both sepsis and septic shock in one scoring system. The model stratifies infection severity based on the score compared to the selected cutoff. Despite lower specificity compared to previously reported models, the SSIT score has the advantage of both simplicity and accuracy. Moreover, it is less time-consuming compared to the SOFA score-based system recommended by the current Sepsis-3 guidelines, which relies on laboratory results after admission.
One notable aspect of our study is the inclusion of baseline platelet count as a predictor of sepsis and septic shock. As noted in the parsimony plot (supplementary Fig. 1), the 4 most important variables are initial vital signs (SBP, PR, and DBP), and baseline platelet. Baseline platelet count is a unique factor that has not been widely considered in predicting septic shock in the ED. The inclusion of baseline platelet count as a predictive factor for sepsis and septic shock aligns with previous studies suggesting the platelets and factors associated with platelet function may play critical roles in cytokine regulation and host immunity [27–30].
Because SOFA score-based sepsis diagnosis relies on laboratory results obtained after ED admission, it delays timely recognition of sepsis. Since the introduction of InSight in 2018 – a scoring system composed solely of vital signs – various ML-based sepsis prediction models have been developed to facilitate earlier sepsis prediction [31]. DeepSEPS, which incorporated longitudinal clinical information, was specifically designed for ICU settings [32]. The qSEPSIS model, developed using data readily available at triage (including vital signs, symptoms, mode of arrival, sex, and age), achieved an AUROC of 0.766 after internal and external validation [6].
SSIT can be considered a scoring system that incorporates the advantages of previously developed models. While Insight and qSEPSIS offer the benefit of rapid prediction using only triage-level information, SSIT enables similarly timely predictions by automatically extracting historical data from electronic medical records. Importantly, despite the comparable time required for prediction, SSIT demonstrates superior AUROC performance compared to models based solely on triage data, especially on septic shock.
SSIT is a system based on vital signs and baseline laboratory results. If baseline laboratory results are unavailable, we assume that these results are normal for scoring purposes. This approach is consistent with Sepsis-3 guidelines. However, the diagnostic accuracy of SSIT is likely to decrease in such cases. Further research is needed relating to accuracy and applicability of SSIT in clinical setting.
Limitations
There are several limitations of this study. First, the study was conducted in a single tertiary hospital in Korea and lacks external validation, which may limit its generalizability to the broader population. However, given that our hospital provides treatment on chronic patients with severe diseases and is among the top hospitals nationwide in terms of patient volume, the included cohort may still reasonably reflect a diverse clinical spectrum. Second, diagnostic value of our model is limited in patients who lack these past laboratory results. For patients without such results, the model assumes normal values, which may not accurately reflect their health status. Finally, because we included only patients with body temperature ≥ 38.0 °C at ED triage in our study, our scoring systems may not be applicable to cases of sepsis without fever.
Conclusion
SSIT, an interpretable ML-based scoring system for early prediction of sepsis and septic shock, demonstrated high sensitivity and specificity among patients with documented fever at ED triage. Future studies, including external validation and assessment in afebrile presentations, are warranted prior to implementation in practice.
Supplementary Information
Below is the link to the electronic supplementary material.
Author Contributions
Conceptualization, Heo S.; methodology, Heo S.; formal analysis, Lee YR.; data curation, Lee YR.; fund acquisition, Heo S.; writing–original draft preparation, Maeng SJ., and Heo S.; writing–review and editing, Maeng SJ., Lee SU., Yu JY., Choi JW., Lee G., Park JE., Shin TG., Hwang SY., Yoon H., Cha WC., Kim T., Kim M., Chang H., and Heo S.; supervision: Heo S.
Funding
This research was supported by a grant of the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health & Welfare, Republic of Korea (RS-2025-25460025).
Data Availability
De-identified datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.
Declarations
Competing interests
The authors declare no competing interests.
Ethics Approval
This study was approved by Institutional Review Board of our institution (IRB No. 2023-06-073) and conducted in accordance with the ethical standards as laid down in the 1964 Declaration of Helsinki and its later amendments.
Informed Consent
Informed consent was waived due to the retrospective nature of the study.
Clinical Trial Number
Not applicable.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Piedmont S, Goldhahn L, Swart E, Robra B-P, Fleischmann-Struzek C, Somasundaram R, et al (2024) Sepsis incidence, suspicion, prediction and mortality in emergency medical services: a cohort study related to the current international sepsis guideline. Infection 52(4):1325–35. 10.1007/s15010-024-02181-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Namgung M, Ahn C, Park Y, Kwak I-Y, Lee J, Won M (2023) Mortality among adult patients with sepsis and septic shock in Korea: a systematic review and meta-analysis. Clinical and Experimental Emergency Medicine 10(2):157–171. 10.15441/ceem.23.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Singer M, Deutschman CS, Seymour CW, Shankar-Hari M, Annane D, Bauer M, et al (2016) The third international consensus definitions for sepsis and septic shock (Sepsis-3). JAMA 315(8):801–10. 10.1001/jama.2016.0287 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Lambden S, Laterre PF, Levy MM, Francois B (2019) The SOFA score—development, utility and challenges of accurate assessment in clinical trials. Critical Care 23(1):374. 10.1186/s13054-019-2663-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Prasad PA, Fang MC, Abe-Jones Y, Calfee CS, Matthay MA, Kangelaris KN (2020) Time to recognition of sepsis in the emergency department using electronic health record data: a comparative analysis of systemic inflammatory response syndrome, sequential organ failure assessment, and quick sequential organ failure assessment. Critical Care Medicine 48(2):200–9. 10.1097/CCM.0000000000004132 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Jiang S, Dai S, Li Y, Zhou X, Jiang C, Tian C, et al (2025) Development and validation of a screening tool for sepsis without laboratory results in the emergency department: a machine learning study. EClinicalMedicine 80:103048. 10.1016/j.eclinm.2024.103048 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Suttapanit K, Satiracharoenkul S, Sanguanwit P, Prachanukool T (2022) The Accuracy of Sepsis Screening Score for Mortality Prediction at Emergency Department Triage. Western Journal of Emergency Medicine 23(5):698–705. 10.5811/westjem.2022.6.56754 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Subbe CP, Kruger M, Rutherford P, Gemmel L (2001) Validation of a Modified Early Warning Score in medical admissions. QJM: An International Journal of Medicine 94(10):521–6. 10.1093/qjmed/94.10.521 [DOI] [PubMed] [Google Scholar]
- 9.Ko R-E, Kwon O, Cho K-J, Lee YJ, Kwon J-M, Park J, et al (2022) Quick Sequential Organ Failure Assessment score and the Modified Early Warning Score for predicting clinical deterioration in general ward patients regardless of suspected infection. Journal of Korean Medical Science 37(16):e122. 10.3346/jkms.2022.37.e122 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Dorsett M, Kroll M, Smith CS, Asaro P, Liang SY, Moy HP (2017) qSOFA has poor sensitivity for prehospital identification of severe sepsis and septic shock. Prehospital Emergency Care 21(4):489–97. 10.1080/10903127.2016.1274348 [DOI] [PubMed] [Google Scholar]
- 11.Roney JK, Whitley BE, Maples JC, Futrell LS, Stunkard KA, Long JD (2015) Modified Early Warning Scoring (MEWS): evaluating the evidence for tool inclusion of sepsis screening criteria and impact on mortality and failure to rescue. Journal of Clinical Nursing 24:3343–54. 10.1111/jocn.12952 [DOI] [PubMed] [Google Scholar]
- 12.Oduncu AF, Kıyan GS, Yalçınlı S (2021) Comparison of qSOFA, SIRS, and NEWS scoring systems for diagnosis, mortality, and morbidity of sepsis in emergency department. American Journal of Emergency Medicine 48:54–9. 10.1016/j.ajem.2021.04.006 [DOI] [PubMed] [Google Scholar]
- 13.Brunetti E, Presta R, Rinaldi G, Ronco G, De Vito D, Brambati T, et al (2023) Predictors of In-Hospital Mortality in Older Inpatients with Suspected Infection. Journal of the American Medical Directors Association 24(12):1868–73. 10.1016/j.jamda.2023.06.012 [DOI] [PubMed] [Google Scholar]
- 14.Knack SK, Scott N, Driver BE, Prekker ME, Black LP, Hopson C, et al (2024) Early physician gestalt versus usual screening tools for the prediction of sepsis in critically ill emergency patients. Ann Emerg Med 84(3):246–258. 10.1016/j.annemergmed.2024.02.009 [DOI] [PubMed] [Google Scholar]
- 15.Kijpaisalratana N, Sanglertsinlapachai D, Techaratsami S, Musikatavorn K, Saoraya J (2022) Machine learning algorithms for early sepsis detection in the emergency department: a retrospective study. Int J Med Inform 160:104689. 10.1016/j.ijmedinf.2022.104689 [DOI] [PubMed] [Google Scholar]
- 16.Brann F, Sterling NW, Frisch SO, Schrager JD (2024) Sepsis prediction at emergency department triage using natural language processing: retrospective cohort study. JMIR AI 3:e49784. 10.2196/49784 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Wang HE, Shapiro NI, Griffin R, Safford MM, Judd S, Howard G (2012) Chronic medical conditions and risk of sepsis. PLoS One 7(10):e48307. 10.1371/journal.pone.0048307 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Yende S, Iwashyna TJ, Angus DC (2014) Interplay between sepsis and chronic health. Trends Mol Med 20(4):234–8. 10.1016/j.molmed.2014.02.005 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.de Araújo GC, Pardini A, Lima C (2023) The impact of comorbidities and COVID-19 on the evolution of community onset sepsis. Scientific Reports 13(1):10589. 10.1038/s41598-023-37709-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Wattanapaiboon K, Banditlerdruk S, Vattanavanit V (2020) Presenting Symptoms in Sepsis: Is the Mnemonic “SEPSIS” Useful? Infection and drug resistance 13:2199–2204. 10.2147/IDR.S263964 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Doman M, Thy M, Dessajan J, Dlela M, Do Rego H, Cariou E, Ejzenberg M, Bouadma L, de Montmollin E, Timsit JF (2023) Temperature control in sepsis. Frontiers in medicine 10:1292468. 10.3389/fmed.2023.1292468 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Harding C, Pompei F, Bordonaro SF, McGillicuddy DC, Burmistrov D, Sanchez LD (2020) Fever Incidence Is Much Lower in the Morning than the Evening: Boston and US National Triage Data. The western journal of emergency medicine 21(4):909–917. 10.5811/westjem.2020.3.45215 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Wang, HE, Jones AR, Donnelly JP (2017) Revised National Estimates of Emergency Department Visits for Sepsis in the United States. Critical care medicine 45(9):1443–1449. 10.1097/CCM.0000000000002538 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Xie F, Chakraborty B, Ong MEH, Goldstein BA, Liu N (2020) Autoscore: a machine learning–based automatic clinical score generator and its application to mortality prediction using electronic health records. JMIR Med Inform 8(10):e21798. 10.2196/21798 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Xie F, Ning Y, Liu M, Li S, Saffari SE, Yuan H, et al (2023) A universal AutoScore framework to develop interpretable scoring systems for predicting common types of clinical outcomes. STAR Protocols 4(2):102302. 10.1016/j.xpro.2023.102302 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Machado FR, Assunção MSCd, Cavalcanti AB, Japiassú AM, Azevedo LCPd, Oliveira MC (2016) Getting a consensus: advantages and disadvantages of Sepsis 3 in the context of middle-income settings. Rev Bras Ter Intensiva 28(4):361–5. 10.5935/0103-507X.20160068 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Thanh Duyen LT, Manh BV, Phuong Thao TT, Khanh LV, Linh Trang BN, Giang NT, et al (2022) Prognostic significance of the angiopoietin-2 for early prediction of septic shock in severe sepsis patients. Future Sci OA 8(10):FSO825. 10.2144/fsoa-2022-0077 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Portier I, Campbell RA (2021) Role of platelets in detection and regulation of infection. Arteriosclerosis, Thrombosis, and Vascular Biology 41(1):70–8. 10.1161/ATVBAHA.120.314645 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 29.Hawwari I, Rossnagel L, Rosero N, Maasewerd S, Vasconcelos MB, Jentzsch M, et al (2024) Platelet transcription factors license the pro-inflammatory cytokine response of human monocytes. EMBO Mol Med 16(8):1901–29. 10.1038/s44321-024-00093-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Claushuis TA, van Vught LA, Scicluna BP, Wiewel MA, Klein Klouwenberg PM, Hoogendijk AJ, et al (2016) Thrombocytopenia is associated with a dysregulated host response in critically ill sepsis patients. Blood 127(24):3062–72. 10.1182/blood-2015-11-680744 [DOI] [PubMed] [Google Scholar]
- 31.Mao Q, Jay M, Hoffman J, Calvert J, Barton C, Shimabukuro D, et al (2018) Multicentre validation of a sepsis prediction algorithm using only vital sign data in the emergency department, general ward and ICU. BMJ Open 8(1):e017833. 10.1136/bmjopen-2017-017833 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Kim T, Tae Y, Yeo HJ, Jang JH, Cho K, Yoo D, et al (2023) Development and validation of deep-learning-based sepsis and septic shock early prediction system (DeepSEPS) using real-world ICU data. J Clin Med 12(22):7156. 10.3390/jcm12227156 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
De-identified datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.