Data were extracted from the MIMIC-IV (Medical Information Mart for Intensive Care IV) database, developed by the MIT (Massachusetts Institute of Technology) Laboratory for Computational Physiology. This deidentified database includes clinical and waveform data from ICU and emergency department patients at Beth Israel Deaconess Medical Center (2008-2019) [21,22]. The physiological indicators were obtained from patient monitors. Blood oxygen saturation (SpO₂), heart rate (HR), and respiratory rate (RES) were collected 1-10 times per hour. Blood glucose (GLU) and body temperature (TEM) were measured once every 1 to 5 hours, depending on the patient’s condition. Other indicators were collected hourly.

Patients with sepsis were identified per the 2018 Chinese Guidelines for Sepsis or Septic Shock Management (positive blood culture + antibiotic use + Sequential Organ Failure Assessment score ≥2). A control cohort included ICU patients without sepsis. Among 1118 patients (550 with sepsis and 568 controls), 8 real-time physiological indicators were selected: HR, systolic blood pressure (SP), diastolic blood pressure (DP), mean arterial pressure (MP), RES, TEM, SpO₂, and GLU. Stratified sampling divided the cohort into training (n=894) and test (n=224) groups. Baseline characteristics showed no significant differences (Table 1).

First, the data underwent outlier removal. Since part of the MIMIC-IV data was manually entered by health care providers, potential input errors or anomalies were addressed. Clinical knowledge was applied to define valid physiological ranges and filter absolute outliers (for instance, TEM: 20 °C-50 °C; SpO₂: 21%-100%; HR: 0-300 bpm). Data points outside these ranges were deemed invalid and removed. Next, missing data imputation was performed to address gaps in the cleaned dataset. A hybrid approach combining multiple interpolation methods was applied:

1. Mean imputation: Replacing missing values with the overall mean of the feature.

2. Class-specific mean imputation: Using mean values from subgroups (eg, sepsis vs nonsepsis cohorts).

3. Linear interpolation: Filling gaps using linear trends between adjacent valid data points.

4. Forward-fill imputation: Propagating the last valid observation forward.

The entire preprocessing workflow is illustrated in Figure 1.

To minimize reliance on long-term temporal dependencies, the prediction model was designed to operate on a 3-hour sliding window of real-time physiological data. This allows the model to initiate sepsis risk prediction immediately after 3 hours of monitoring. A 3-hour window allows sufficient time for sign monitoring and subsequent sepsis prediction, providing doctors with ample time for intervention. From the 3-hour time series of each physiological indicator, 3 linear parameters were computed: mean value over the 3-hour window, fluctuation coefficient (SD within the window), and endpoint value (the last recorded value in the window).

This generated a 24-dimensional feature vector for each patient, structured as follows:

All features were standardized (z score normalization) to ensure consistent scaling before model training. This preprocessing pipeline ensures robust, clinically meaningful inputs for the subsequent machine learning workflow.

We developed and evaluated the sepsis prediction model for actual ICU sign monitoring to achieve a model with greater generality in a broader clinical setting. Critically ill patients in the ICU have different needs for GLU monitoring frequency. The incidence of glucose metabolism disorders in critically ill patients is high, and the incidence increases in turn in sepsis, severe sepsis and septic shock [23], which not only reflects the abnormal secretion of hormones and the severity of the disease, but is also closely related to the increase of mortality and complications [24]. High or low GLU is one of the causes of organ dysfunction. Therefore, it is necessary to increase the frequency of GLU monitoring in the face of the clinical background of insufficient energy intake, high catabolism, impaired GLU regulation mechanism, or the implementation of insulin treatment. At the same time, it is also necessary to increase the frequency of GLU monitoring in critically ill patients with diabetes or hypoglycemia. Whereas in most other ICU patients, GLU is usually measured only once or twice a day.

Therefore, to improve the generalization of sepsis prediction models, we developed and evaluated 2 sepsis prediction models, one based on high-frequency glucose monitoring and the other based on routine vital signs monitoring without glucose. By combining the 2 models, a model with greater generality in a wider range of clinical settings was achieved.

This study used 6 machine learning algorithms—support vector machine, random forest, ExtraTrees, XGBoost (Extreme Gradient Boosting), AdaBoost, and Logistic—to construct real-time sepsis prediction models. Model performance was evaluated and compared using metrics including accuracy, precision, recall, F₁-score, and area under the receiver operating characteristic curve (AUROC).

The TreeSHAP (Tree-Based Shapley Additive Explanations) method was used to provide both local (individual prediction) and global explanations for the sepsis prediction model. TreeSHAP, grounded in cooperative game theory, constructs an additive explanation model that treats all features as “contributors” to the prediction outcome. For each sample, TreeSHAP generates a Shapley value for each feature, quantifying its marginal contribution to the model’s decision. For the sepsis prediction model with 24 features, let the original model be f, trained using AdaBoost. The explanation model g in SHAP (Shapley Additive Explanations) is defined as:

x=(x₁,x₂…x₂₄) represents the feature vector of a single patient. f(x) is the prediction from the original model. g(x) is the prediction from the explanation model. ø_i is the Shapley value for the ith feature.

S is a subset of {1,2,3…24}, with 2^24-1 possible combinations. |S| is the number of features in subset S. f_x (S ∪ i) and f_x(S) are model predictions with and without feature i, respectively. Global interpretation aggregates local explanations by averaging the absolute Shapley values across all samples.

Our study was conducted in accordance with the guidelines of the Helsinki Declaration. The Review Committee of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center approved access to the MIMIC-IV database. Authors fulfilled the database access request. All these data were deidentified; therefore, the study was exempt from ethical approval and informed consent requirements.

The performances of the sepsis prediction models based on high-frequency glucose monitoring are summarized in Figure 2 and Table 2. Among all models, AdaBoost achieved the highest accuracy of 0.70 (95% CI 0.68-0.71), precision of 0.69 (95% CI 0.68-0.71), F₁-score of 0.69 (95% CI 0.67-0.70), and AUROC of 0.76 (95% CI 0.74-0.77), demonstrating superior real-time prediction capability. The performances of the sepsis prediction models based on routine vital signs monitoring without glucose are summarized in Figure 3 and Table 3. Among all models, AdaBoost achieved the highest accuracy of 0.67 (95% CI 0.66-0.69), precision of 0.67 (95% CI 0.65-0.68), and AUROC of 0.75 (95% CI 0.74-0.77). The experimental results show that the sepsis prediction model based on high-frequency monitoring of GLU has better prediction performance, and the sepsis prediction model based on routine vital signs monitoring has a wider application scenario.

Our study developed a real-time sepsis prediction model that integrates high timeliness and clinical interpretability based on 3-hour dynamic temporal sequences of 8 rapidly accessible physiological indicators. The real-time sepsis prediction model demonstrated robust performance. The output interpretation of explainable artificial intelligence (XAI) enhanced model transparency through both individual prediction and global explanations, and it linked the potential physiological or pathophysiological significance, including the patient’s hemodynamics, thermoregulatory response, and the balance between oxygen delivery and oxygen consumption. Although current consensus emphasizes early intervention for sepsis management, the optimal predictors guiding intervention and markers of early sepsis severity remain unclear. In this study, we further elucidated the model’s output results, explored the potential relationships between relevant sign parameters and sepsis, and discussed their potential physiological or pathophysiological significance. This enhances the interpretability and credibility of the XAI method, supporting the model’s applicability in real-world clinical practice. Finally, the web-based platform significantly enhanced clinical utility by providing real-time risk assessment, statistical summaries, trend analysis, and actionable insights.

However, as a retrospective study, potential biases may exist in this study. Future efforts should prioritize multicenter validation and large-scale prospective studies to strengthen the robustness of these results. XAI holds immense promise in sepsis diagnosis and treatment, yet its development and clinical application face significant challenges [36]. Data quality remains a critical bottleneck. Heterogeneous hospital databases, inconsistent data collection and storage standards, and poor interoperability between health care information systems have led to fragmented “data silos,” hindering the application of large-scale clinical feature datasets. Additionally, the reliance on limited public or institution-specific databases in research settings restricts generalizability. In clinical practice, clinicians must integrate XAI predictions with laboratory findings (eg, lactate or procalcitonin), imaging results, and patient-specific symptoms to ensure accurate diagnosis and treatment planning.

In ICU settings, real-time physiological indicators—such as HR, RES, and SpO₂—are typically used to monitor symptom fluctuations rather than generate objective diagnostic reports. This limitation stems from the significant variability in individual physiological data, the complexity of multiparameter interactions, and the diverse clinical implications of these metrics. This study demonstrates that XAI can bridge this gap by synthesizing real-time physiological data into actionable insights. By analyzing dynamic trends and providing interpretable explanations, XAI uncovers the diagnostic potential of these data. Our model, focused on sepsis risk prediction, leverages real-time physiological features to generate predictions while emphasizing the interpretability of results. In the future, XAI systems could deliver intelligent diagnostic reports integrating disease prediction, anomaly detection, and causal analysis of abnormal indicators, empowering clinicians to navigate complex physiological data with precision.