This is the most active line of research, aiming to issue alerts hours before a definitive diagnosis of sepsis. Models are typically deployed in emergency departments (ED), general wards, or ICUs, with the goal of identifying infected or suspected infected patients at risk of progressing to sepsis (2). The setting of the lead-time prediction window is crucial, usually between 4 and 24 h (4, 9). An effective warning system must strike a balance between providing sufficient intervention time and avoiding excessively high false positive rates (9).

Predicting short-term or long-term adverse outcomes for patients diagnosed with sepsis can aid in risk stratification, resource allocation, and treatment decision-making. Common outcomes include hospitalization/ICU mortality rate (5, 10), 28-day or 90-day mortality rate (11), as well as long-term functional outcomes such as post–intensive care syndrome (PICS).

The harm of sepsis is mainly achieved by inducing organ dysfunction. Therefore, predicting the risk of specific organ complications, such as septic shock, has important clinical value (12), acute respiratory distress syndrome (ARDS), acute kidney injury (AKI) (13), disseminated intravascular coagulation (DIC) and sepsis associated encephalopathy (SAD) (14), etc.

This type of task focuses on the consumption of medical resources, such as ICU or length of hospital stay (LOS) (15), the risk of readmission and the need for mechanical ventilation or hemodynamic support.

This is a cutting-edge application of predictive models aimed at guiding treatment choices. For example, predicting patients’ responsiveness to fluid resuscitation and their demand for specific vasoactive drugs (16), or predict its response to immunomodulatory therapy based on immune phenotype.

When setting these outcomes, it is necessary to precisely define the time zero, observation window, and prediction window at which the event occurs (17). For binary classification tasks, determining the event time is crucial. In survival analysis, competitive risks (such as discharge) and censoring need to be considered (18). The setting of lead time for prediction needs to take into account clinical operability, that is, the reasonable time from the issuance of the alarm to the effective intervention of the clinical team. In addition, the definitions of all outcome measures should be standardized and operationalized as much as possible, for example, by uniformly adopting the Sepsis-3 standard, clarifying the hemodynamic definition of shock, and dynamically evaluating changes in organ function scores, which is crucial for ensuring the comparability and reproducibility of the model (19).

This is the most commonly used type of data, sourced from EHR and CIS. It includes: Vital signs: such as heart rate, blood pressure, respiratory rate, body temperature, oxygen saturation, etc. (4). Laboratory examination: Complete blood count (CBC) (2, 20), blood gas analysis, liver and kidney function, coagulation indicators, etc. Medication records: antibiotics, vasoactive drugs, sedatives, etc. Fluid intake and output: Accurately recording the fluid balance is an important indicator for evaluating the circulation state. Clinical scoring: SOFA, APACHE II, SAPS II and their dynamic changes can serve as strong predictive factors (7).

The high-frequency data streams generated by bedside monitors, such as electrocardiogram (ECG) waveforms, invasive arterial pressure waveforms, ventilator parameter curves, etc., contain richer physiological information than sparsely recorded vital signs. Signal processing and feature extraction of these waveforms can capture heart rate variability (HRV) (21), the variability of blood pressure and subtle changes in cardiopulmonary interaction(s) may be early signals of organ dysfunction.

Clinical text: Through natural language processing (NLP) technology, key information such as evidence of “suspected infection,” symptom descriptions, physical examination findings, etc. can be extracted from medical orders, course records, nursing records, and discharge summaries to compensate for the lack of structured data (22). Medical imaging: Deep learning, especially convolutional neural networks (CNN), can extract information from chest X-rays (CXR) (23), automatically extract imaging features from bedside ultrasound (POCUS) or CT scans for identifying the source of infection (such as pneumonia) (24) or evaluate organ damage (such as pulmonary edema, heart function) (25).

Traditional biomarker: procalcitonin (PCT) (20), C-reactive protein (CRP), interleukin-6 (IL-6), lactate, lactate and its clearance are important auxiliary tools for the diagnosis and prognosis of sepsis. Some studies have also explored novel biomarkers such as soluble urokinase type plasminogen activator receptor (suPAR) and lipids (such as ceramides) (26) predictive value. Multi omics data: High throughput techniques such as transcriptomics, proteomics, and metabolomics can reveal the complex host response network of sepsis. By identifying specific gene expression profiles or metabolite characteristics, patients can be stratified into different immune endotypes, providing the possibility for precise treatment and personalized prediction. Microbiological data: including pathogen identification, microbiome composition, and antibiotic resistance spectrum, are crucial for guiding anti infective treatment and predicting treatment response.

Data quality and governance are the foundation of successful modeling. When dealing with these multi-source heterogeneous data, many challenges must be addressed, including data loss (and its underlying mechanisms such as information loss), imprecise timestamps and alignment issues, delays in label acquisition, and multi center data heterogeneity caused by differences in practice among different medical institutions. Adopting standardized data models (such as OMOP-CDM) (27) and strict data governance processes are key to improving research quality and model portability.

To facilitate reproducibility and benchmarking, several large, publicly accessible datasets and benchmark challenges have shaped sepsis prediction research (Table 1).

Common algorithms: compared with traditional statistical models, machine learning algorithms can better capture nonlinear relationships and high-order interactions in data. Tree-based models, such as Random Forest (RF) (5), Extreme Gradient Boosting (XGBoost) (7, 31), and LightGBM (32), due to their high performance and robustness, are widely used in sepsis prediction. Other algorithms such as Support Vector Machine (SVM), K-Nearest Neighbor (kNN), and Naive Bayes are also commonly used as comparison baselines (10). Feature engineering: The success of machine learning largely relies on feature engineering, which involves extracting meaningful variables from raw data. This includes creating lag features, differential features, and statistical features that reflect physiological variability (such as standard deviation, coefficient of variation) to capture the dynamic evolution of diseases (33).

Temporal model: For time series data collected intensively in ICU, deep learning models can automatically learn complex temporal dependencies (34). Recurrent Neural Networks (RNNs) and their variants—Long Short-Term Memory Networks (LSTM) (35) and Gated Recurrent Unit (GRU)—are commonly used architectures for processing such data. In recent years, Time Convolutional Networks (TCNs) and Transformer models based on self attention mechanisms have also received attention due to their advantages in processing long sequence data. Representation learning: Another major advantage of deep learning is the ability to perform self supervised learning from unlabeled data, resulting in rich data representations. For example, models can be trained from massive ECGs (36) or learn universal physiological state representations from arterial pressure waveforms, and then use these representations for downstream sepsis prediction tasks (37). Multimodal Fusion: Deep learning frameworks provide powerful capabilities for fusing heterogeneous data such as structured data, text, images, and waveforms. By designing specific network structures (such as attention mechanisms) to integrate information from different modalities, it is expected to achieve more accurate predictions than a single data source (38).

Causal inference: Traditional prediction models mainly focus on correlation, while clinical decision-making requires more causal insights. The causal inference method aims to estimate the individualized treatment effects (ITE) of treatment interventions from observational data to avoid confounding bias caused by treatment choices themselves (39). Reinforcement Learning (RL): RL provides a theoretical framework for sequential decision problems. By modeling patient status, clinical interventions, and long-term outcomes as a Markov decision process, RL can learn optimal dynamic treatment strategies, such as fluid management or titration of vasoactive drugs, achieving the integration of prediction and decision-making (40).

In multi center collaboration, data privacy and security are the main obstacles. Federated Learning allows centers to collaboratively train a global model by exchanging model parameters without sharing raw patient data, providing a feasible solution for building large-scale and diverse prediction models (41).

To facilitate comparison across paradigms, Table 2 summarizes key characteristics, typical use cases, and advantages and limitations of traditional scoring systems, conventional machine learning models, and deep learning approaches in sepsis prediction.

Figure 2 provides an overview of major methodological paradigms in sepsis prediction, spanning rule-based clinical scores, conventional machine learning, deep learning with temporal modeling, and emerging multimodal approaches.

Clear definition: It is necessary to clarify the inclusion and exclusion criteria of the study, the precise operational definition of events (such as sepsis onset), and the time zero (such as ICU admission time or the time when the suspected infection criteria are first met). Sample size: Sufficient sample size and number of events should be ensured to avoid overfitting of the model. Data partitioning: The dataset should be strictly divided into training set, validation set, and testing set. The most ideal validation method is to conduct external validation in time and space, that is, to test the performance of the model on data from different time periods (temporal external validation) or different medical centers (external validation across centers), in order to evaluate its generalization ability (7, 20).

Missing data: The proportion and pattern of missing data should be reported and processed using appropriate methods such as multiple imputation (42). Some models, such as gradient boosting trees, can inherently handle missing values. Sometimes, the absence itself is a form of information (informative missingness) that can be used as a feature input model. Category imbalance: Adverse events such as sepsis are usually rare events in the population, which can lead to imbalanced data categories. The processing methods include resampling (oversampling or undersampling) or introducing category weights in the loss function. When evaluating model performance on imbalanced datasets, the area under the precision recall curve (AUPRC) is typically more informative than the area under the receiver operating characteristic (ROC) curve (43).

Discrimination: measures the ability of a model to distinguish between patients with and without events, commonly measured by AUROC and AUPRC (44). Calibration: Evaluating the consistency between the predicted probability of the model and the actual observed frequency. A well calibrated model that predicts a 20% risk should correspond to approximately 20% of patients in that risk group actually experiencing events (45). The evaluation tools include calibration curve, Brier score, and calibration slope. Clinical utility: Decision curve analysis (DCA) is an effective tool for evaluating the clinical net benefits of a model. It links the predictive performance of the model to specific decision thresholds, helping clinicians determine at which risk threshold to use the model more beneficial than the default strategy of “treat all” or “not treat all” (46).

Data leakage: in the process of model development, it is necessary to be vigilant about the possibility of future information leakage to past predictions (47). For example, using data after an event to train a model that predicts events before they occur. When processing time series data, the division of cross validation should maintain temporal continuity. Common bias: it is necessary to pay attention to and try to control the backdoor bias and immortal time bias caused by intervention measures.

Reporting guidelines: Research reports should follow internationally recognized guidelines, such as TRIPOD (Transparent Reporting of Multivariate Prediction Models) (48)And its extension TRIPOD-AI for AI models, as well as PROBAST (a bias risk assessment tool for predictive model research). Reproducibility: In order to promote scientific validation and collaboration, researchers should make their code, model weights, and provide detailed data dictionaries and model cards as much as possible to enhance the reproducibility of their research.

The “Artificial Intelligence Sepsis Expert” (AISE) algorithm developed by Nemati et al. (4) utilizes 65 clinical variables to predict sepsis attacks 4 to 12 h in advance in the ICU, demonstrating good performance on an external validation set. SepsisAI developed by Gupta et al. (35), a deep learning model based on LSTM, aims to predict in-hospital sepsis in real-time and has specially designed alarm suppression logic to reduce alarm fatigue. Some studies focus on general ward or emergency department scenarios, using more limited datasets for early warning, which is crucial for preventing patients from deteriorating and being transferred to the ICU. These models typically require a trade-off between early warning and positive predictive value (PPV) to adapt to the resources and workflows of different clinical environments.

Many studies have shown that machine learning models outperform traditional scoring in predicting mortality rates in sepsis patients. For example, Taylor et al. (5) used a random forest model to predict in-hospital mortality in emergency sepsis patients, which performed better than traditional methods. The XGBoost based model developed by Wang et al. (7) showed excellent mortality prediction ability in both MIMIC-IV database and external validation of Chinese teaching hospitals (AUROC were 0.873 and 0.844, respectively), and provided interpretability analysis based on SHAP. Li et al.’s (10) study also confirmed that gradient boosting decision trees (GBDT) perform well in predicting mortality rates in ICU sepsis patients.

Progress has also been made in predictive models for specific organ dysfunction. For example, Zhang et al. developed an integrated machine learning model for early prediction of sepsis related AKI (13), and another study focused on predicting sepsis associated delirium (14). These models help clinical doctors take preventive measures in advance.

Research has shown that combining routine laboratory tests such as CBC with machine learning can effectively predict sepsis without relying on expensive or delayed biomarkers (2, 20). The model that integrates multiple omics data is still in the exploratory stage, but has shown great potential. By identifying transcriptome features associated with specific immune response patterns, patients can be stratified for risk and may guide future immunomodulatory therapies.

Existing research has shown that machine learning and deep learning models typically achieve high levels of discrimination in sepsis prediction tasks (with AUROC mostly between 0.80–0.95). Compared to traditional scoring, these models can better utilize massive and dynamic data. However, many models still lack rigorous external validation, and their transferability in different populations and medical environments remains a key issue. In addition, the complexity of the model also requires interpretability so that clinical doctors can understand and trust its predictive results.

The main reason for the decline in model performance is the change in data distribution, known as “domain shift.” This can be a spatial shift between different hospitals or regions, or a temporal drift caused by changes in clinical practice, testing equipment, patient populations, or disease epidemiology (such as the COVID-19 pandemic) over time (6, 49). Therefore, strict external validation and temporal external validation are the gold standards for evaluating model generalization.

When the model is deployed to a new environment, even if its discrimination remains good, its calibration may deteriorate. Therefore, it is necessary to recalibrate the model, such as adjusting its prediction probability through methods such as Platt scaling or isotonic regression. More advanced strategies include transfer learning and domain adaptation, which can utilize a small amount of labeled or unlabeled data from new environments to fine tune the model to adapt to new data distributions.

A model that performs well in the overall population may perform poorly in specific subgroups (such as patients of different ages, genders, races, or with specific underlying diseases). This may be due to systematic bias present in the data. Therefore, it is necessary to evaluate the fairness of the model’s performance in key subgroups to ensure that it does not exacerbate healthcare inequality. For example, in scenarios of immune suppression, elderly vulnerability, or resource constraints, the performance of the model may require special attention.

Model deployment is not a one-time solution. A machine learning operations (MLOps) system needs to be established to continuously monitor the online performance of the model and detect drift in its discrimination and calibration. Once performance degradation is detected beyond an acceptable range, an alert needs to be triggered and the model’s rollback, retraining, or update process needs to be initiated to form a closed-loop model governance and lifecycle management.

The model must be seamlessly integrated into existing clinical information systems (such as EHR) and workflows. This involves how to present predictive results (such as risk scores, warning information), triggering logic, frequency, and suppression strategies of alerts to clinical doctors. A core challenge is to avoid ‘alarm fatigue’, where too many false positive alarms lead to clinical doctors losing trust and response to them (35).

Any warning system needs to set a risk threshold for triggering actions. The selection of this threshold not only depends on the performance of the model (such as sensitivity and specificity), but should also be coupled with clinical resources (such as ICU beds, manpower). Through sensitivity analysis, it is possible to evaluate how the optimal threshold changes under different resource constraints to achieve the maximization of clinical net benefits.

Figure 3 schematically links model-predicted risk thresholds with resource constraints and potential clinical actions, highlighting how decision-analytic thinking can guide threshold selection.

Although workflow integration and threshold selection are essential, the real-world safety of sepsis prediction models ultimately relies on how clinicians interpret and act on model outputs. The following section summarizes clinician-oriented interpretive prerequisites for safe deployment.

The ultimate evidence to prove the clinical utility of the model comes from prospective impact studies. The research design can include: pre - and post control study: comparing the changes in clinical process indicators (such as antibiotic initiation time, fluid resuscitation volume) and clinical outcomes (such as mortality rate, ICU length of stay) before and after model deployment. Cluster randomized controlled trials (cRCTs): Randomly allocate medical units (such as wards or hospitals) to intervention groups (using models) and control groups (conventional treatments), which is the gold standard design for evaluating the effectiveness of clinical decision support systems (CDSS). A/B testing or silent mode deployment: Running the model in the real world without displaying the results to clinical doctors (silent mode) as a control to evaluate its potential impact. Although such research evidence is still scarce, some studies have begun prospective validation (50).

The experience of a few sepsis warning systems already deployed in clinical settings, such as Epic’s Sepsis Prediction Model, suggests that their real-world performance may be lower than reported during the development phase (28), this highlights the importance of continuous performance monitoring and external recalibration. In addition, the design of human-computer interaction, interpretability of models, and how to avoid automation bias (i.e., excessive reliance on models while ignoring clinical judgments) are key factors determining the success or failure of the system.

In addition to clinical outcomes, the economic impact of model deployment should also be evaluated, including changes in healthcare costs, resource utilization efficiency, and bed turnover, and a cost–benefit analysis should be conducted.

The physiological parameters and response patterns to infection in children and newborns are significantly different from those in adults. Data sparsity is the main challenge in modeling in this field. Meeus et al. developed a machine learning model for premature infants to predict late onset sepsis (LOS) and necrotizing enterocolitis (NEC), demonstrating promising potential (31). In pediatrics, age-dependent vital sign norms and different immune-response trajectories often require age-stratified modeling, pediatric-specific feature normalization, or explicit incorporation of age as a non-linear effect modifier. In neonates, models commonly integrate gestational age, birth weight, and ventilator parameters and are designed to function under sparse sampling and frequent missingness typical of NICU workflows.

The risk and manifestations of sepsis vary among elderly and vulnerable populations, pregnant and postpartum women, as well as immunocompromised patients (such as tumors and organ transplants). For example, Ke et al. (51) developed an in-hospital mortality prediction model specifically for elderly sepsis patients. For immunocompromised patients, baseline inflammatory markers and vital signs may be blunted, making fixed thresholds less informative. Practical adaptations include emphasizing trajectory features (e.g., within-patient change), incorporating immunosuppression-related variables (e.g., neutropenia status, recent chemotherapy/transplant), and using subgroup-specific recalibration to reduce systematic underestimation or overestimation of risk.

Postoperative, trauma (9), and burn patients often have strong non infectious inflammatory reactions, which makes the diagnosis and prediction of sepsis more challenging, requiring models that can distinguish between infectious and non infectious inflammation.

In these environments, the available data (such as laboratory tests) is very limited. Therefore, it is necessary to develop simplified models based on a small number of easily accessible variables, such as vital signs and physical examinations, which can be implemented on low-cost hardware or even paper-based tools. Parsimonious models and early warning scores can be paired with structured triage protocols to support escalation decisions, while the development of wearable devices and remote monitoring technology provides new possibilities for scalable early warning in these scenarios.

Across these special populations, effective adaptation strategies consistently include subgroup-specific recalibration, greater reliance on within-patient trajectories rather than absolute thresholds, and explicit incorporation of context-specific constraints (e.g., data sparsity or resource limitation) into model design and evaluation.

Drawing on the success of NLP and computer vision, developing multimodal foundational models capable of processing and integrating clinical text, images, waveforms, and structured data will be an important direction for the future. By performing self supervised learning on massive unlabeled clinical data, these models can learn universal and rich physiological and pathological representations, providing a powerful foundation for various downstream prediction tasks.

Future models will not only predict ‘what will happen’, but also answer ‘why it will happen’ and ‘what if…’. By combining machine learning with causal inference and systems biology models, a “digital twin” that simulates the physiological state of patients can be constructed to conduct counterfactual inference, evaluate the potential impact of different intervention measures on individual patients, and achieve a leap from correlation prediction to actionable individualized decision support.

Integrating multiple omics data into a predictive model aimed at identifying different immune subtypes of sepsis. This is expected to achieve “diagnosis and treatment integration”, where the model can predict high risks while also recommending treatment strategies targeting specific immune phenotypes (such as immune enhancement or immune suppression), and serve as a companion diagnostic tool to guide clinical trials and personalized treatments.

In order to cope with model drift, future clinical AI systems need to have the ability of continuous learning, which means continuously learning and adapting from new data streams without completely forgetting old knowledge. This requires finding a balance between the automation of model updates and clinical validation, regulatory compliance, and may require exploring innovative models such as “regulatory sandboxes.”

With the increasing demand for data sharing and model collaboration, technologies and regulations that protect patient privacy will become increasingly important. Privacy computing technologies such as federated learning and differential privacy, as well as compliance with regulations such as the Medical Device Software (SaMD) and the EU Artificial Intelligence Act, will be necessary conditions for model development and deployment.

By sharing datasets (such as PhysioNet Challenge) (52), establishing reproducible baseline models and evaluation criteria can promote the healthy development of the field. Building an interdisciplinary team consisting of clinical experts, data scientists, and engineers is key to ensuring the clinical relevance and usability of the model.

Many models that perform well on retrospective data are difficult to generate practical value in clinical practice due to complex and diverse reasons. Firstly, the evaluation criteria of the model may be too low (such as compared to simple clinical scoring), or there may be bias in label definition (such as based on coding for sepsis diagnosis). Secondly, high AUROC does not directly translate into clinical usability; the calibration degree of the model, performance at specific decision thresholds (PPV/NPV), and compatibility with clinical workflows are more important. Finally, human factors such as the acceptance and trust of clinical doctors, as well as alarm fatigue, are the ultimate checkpoints that determine whether technology can be implemented (8).

In critical care medicine, the “black box” model often causes concerns among clinical doctors. Interpretability is not just about providing feature importance ranking. A more valuable explanation would be individualized and dynamic, able to tell clinical doctors why the model is sounding an alarm for this specific patient at the current time point. Techniques such as SHAP (7) or LIME can provide local explanations. Furthermore, providing counterfactual explanations (such as’ if a patient’s certain indicator changes, the predicted results will change ‘) may have more clinical guidance significance (53).

Although machine learning models usually outperform in performance, traditional scoring systems still have irreplaceable value in resource limited or fast evaluation scenarios due to their simplicity, transparency, and lack of complex computing environments. The future trend may be to adopt a hybrid strategy, such as using traditional scoring for initial screening, and then using computationally intensive machine learning models for fine evaluation of high-risk populations, or integrating traditional scoring as one of the features into machine learning models.

For research teams or medical institutions wishing to develop and apply sepsis prediction models, a pragmatic roadmap should start with rigorous single center retrospective model development, followed by rigorous multi center, out of time external validation to confirm its robustness. On this basis, small-scale prospective silent deployment pilots can be conducted to evaluate their performance in the real world and potential impact on clinical workflows, and ultimately consider conducting larger scale impact studies.