1Department of Emergency Medicine, Korea University Ansan Hospital, Ansan, Korea
2Core Research and Development Center, Korea University Ansan Hospital, Ansan, Korea
Copyright © 2026 The Korean Society of Emergency Medicine
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/).
Conflicts of interest
The authors have no conflicts of interest to declare.
Funding
The authors received no financial support for this study.
Data availability
Data sharing is not applicable as no new data were created or analyzed in this study.
Many cautions listed for screening, prognosis, treatment, and phenotyping reflect established methodological concerns in sepsis research that may be inherited, amplified, or obscured by AI. Hallucinations are associated with generative AI, whereas retrieval failure and weak source fidelity are particularly relevant to RAG-based systems.
AI, artificial intelligence; ECG, electrocardiography; EHR, electronic health record; ICU, intensive care unit; LLM, large language model; RAG, retrieval-augmented generation.
Tabular EHR data refer to structured, row-by-column clinical data such as demographics, vital signs, laboratory results, medications, organ support, timestamps, and outcomes. Cox models include conventional Cox proportional hazards models with fixed covariates and time-dependent Cox models with covariates updated over time.
AI, artificial intelligence; GAM, generalized additive model; EBM, explainable boosting machine; EHR, electronic health record; XGBoost, Extreme Gradient Boosting; LightGBM, Light Gradient Boosting Machine; CatBoost, Categorical Boosting; SHAP, Shapley Additive Explanations; LIME, Local Interpretable Model-agnostic Explanations; ICE, individual conditional expectation; ICU, intensive care unit; RNN, recurrent neural network; GRU, gated recurrent unit; LSTM, long short-term memory; TCN, temporal convolutional network; CNN, convolutional neural network; ECG, electrocardiography; RAG, retrieval-augmented generation; LLM, large language model;
| Clinical or research task | Typical data and timing | Target output or AI role in sepsis | Candidate model families | Key methodological and AI-related cautions |
|---|---|---|---|---|
| Screening and diagnosis | Demographics, comorbidities, vital signs, laboratory results, triage variables, clinical text, microbiology results, imaging, or ECG; single time point or short observation window | Estimated probability of sepsis or infection, clinical alert, and supporting evidence | Conventional statistical and interpretable models; tree-based structured-data machine learning models; modality-specific encoders; multimodal fusion models | Clinical label and time-zero definition, temporal leakage, class imbalance, false-positive burden, and alert fatigue |
| Prognostic prediction | Serial EHR variables, organ support, treatment response, progress or nursing notes, imaging, ECG, and physiological waveforms; longitudinal or real-time data | Horizon-specific or time-to-event risk of septic shock, new or worsening organ failure, ICU transfer, short- or long-term mortality, or recovery | Time-to-event and dynamic prediction models; tree-based models using trajectory features; temporal deep learning models; multimodal fusion models | Irregular sampling, missingness, censoring, competing events, treatment-dependent variables, time-varying decisions, and calibration over time |
| Treatment and decision support | Longitudinal hemodynamics, laboratory results, organ support, treatment history, clinical notes, microbiology or radiology findings, and local protocols; updated over time according to treatment exposure and response | Risk-based prompts, estimated treatment effects, prioritized options, or source-linked recommendations for fluids, vasopressors, corticosteroids, antibiotics, source control, organ support, or ICU transfer | Causal inference and treatment effect models for intervention-related questions; prediction models for risk estimation only; LLM/RAG-based systems for source-linked presentation | Risk prediction does not establish treatment benefit; confounding by indication, immortal time bias, protocol alignment, uncertainty, and clinician verification |
| Phenotyping and heterogeneity | EHR variables, laboratory results, biomarkers, treatment exposure, outcomes, omics-derived high-dimensional data, and optional multimodal data; baseline profiles or longitudinal cohort data | Phenotype, cluster, latent representation, potential responder subgroup, or trial-eligibility assessment | Phenotyping and representation learning models; multimodal fusion models; causal inference and treatment effect models for treatment effect heterogeneity | Sensitivity to preprocessing and model selection, cluster instability, limited reproducibility and transportability, biological plausibility, and clinical actionability |
| Documentation and communication | Structured EHR data combined with clinical notes, microbiology or radiology reports, retrieved guidelines, and local protocols; encounter-level, handoff-level, hospital-course, or discharge-level information | Source-linked summary, handoff draft, explanation, guideline summary, or draft recommendation for clinician review | Generative AI and LLM/RAG-based systems; structured-to-text systems; modality-specific encoders and multimodal fusion models when source data are diverse | Hallucination, outdated or irrelevant retrieval, weak source fidelity, overconfident language, automation bias, and required clinician verification |
| Model family | Model or approach example | Suitable data and task | Main strength | Key consideration in sepsis | Explainability and verification approach |
|---|---|---|---|---|---|
| Conventional statistical and interpretable models | Linear regression, logistic regression, penalized regression, GAM, EBM | Tabular EHR data; screening, diagnosis, and initial risk prediction | Transparent and clinically familiar; relatively easy to implement; low computational burden | Limited ability to capture complex temporal, high-dimensional, or multimodal relationships; model assumptions require checking | Coefficients, odds ratios, score components, nomograms, shape functions, and risk curves |
| Tree-based structured-data machine learning models | Random forest, XGBoost, LightGBM, CatBoost | Tabular EHR data; single-time-point prediction or summarized longitudinal data | Captures nonlinear relationships and interactions; often provides a strong baseline for structured clinical data | Less transparent; may learn institution-specific missingness or practice patterns; feature attribution should not be interpreted as evidence of causality | Permutation importance, SHAP, LIME, partial dependence plots, and ICE plots |
| Time-to-event and dynamic prediction models | Cox models, landmarking models, joint models, competing-risk models | Longitudinal structured data; mortality, organ failure, ICU transfer, recovery, and other time-to-event outcomes | Accounts for event timing, censoring, and updated patient information; provides horizon-specific risk estimates | Requires a clear time origin, outcome definition, censoring strategy, and approach to competing events and treatment-dependent variables | Survival curves, dynamic risk curves, landmark-specific estimates, and time-dependent effects |
| Temporal deep learning models | RNN, GRU, LSTM, TCN, temporal transformer | Serial EHR data and physiological waveforms; dynamic prognosis or real-time monitoring | Captures complex trajectories and temporal interactions | Requires relatively large datasets and computational resources; sensitive to irregular sampling, missingness, domain shift, and latency; model compression may be needed for real-time deployment | Trajectory visualization, temporal attribution, occlusion analysis, saliency maps, and cautious use of attention visualization |
| Modality-specific encoders | Clinical language models, CNN, vision transformer, pretrained or foundation-model-based encoders, one-dimensional CNN, signal transformer | Clinical text, imaging, ECG, and physiological waveforms | Extracts information from high-dimensional modality-specific data; pretrained or self-supervised models may reduce annotation requirements | Sensitive to documentation bias, image or signal quality, artifacts, synchronization, domain shift, and computational burden | Text or token attribution, class activation maps, image occlusion, waveform-segment attribution, and robustness checks |
| Multimodal fusion models | Early fusion, late fusion, cross-attention models, multimodal transformers | Combinations of EHR data, text, imaging, physiological signals, biomarkers, or omics data | Integrates complementary information from multiple data sources | Missing modalities, temporal alignment, heterogeneous data quality, modality dominance, model opacity, and computational burden | Modality ablation, modality-specific attribution, cross-modal contribution displays, and representative case comparisons |
| Generative AI and RAG-based systems | RAG-based LLMs, guideline-constrained LLMs, structured-to-text LLMs | Structured, unstructured, or multimodal clinical data; extraction, summarization, explanation, documentation support, and draft recommendations for clinician review | Supports information extraction, summarization, documentation, and source-linked communication | Hallucination, retrieval failure or outdated retrieval, weak source fidelity, overconfident language, cost, and inference latency | Retrieved-source display, citation and version tracking, audit trails, uncertainty display, and clinician verification |
| Causal inference and treatment effect models | Target trial emulation, propensity score methods, marginal structural models, causal forests, uplift models, dynamic treatment regimes | Longitudinal treatment and outcome data; average or individualized treatment effect estimation | More appropriate than risk prediction models alone for questions about intervention benefit | Confounding by indication, limited overlap, immortal time bias, unmeasured confounding, and difficulty evaluating learned treatment policies | Causal diagrams, balance and overlap checks, subgroup treatment-effect plots, and sensitivity analyses |
| Phenotyping and representation learning models | Clustering, latent class analysis, autoencoders, contrastive learning, multimodal representation learning | Structured EHR, biomarkers, omics, or multimodal data; phenotype, cluster, latent representation, or responder subgroup identification | Useful for studying sepsis heterogeneity, trial enrichment, and hypothesis generation | Sensitive to preprocessing and model choice; clusters may be unstable, poorly reproducible, or difficult to interpret clinically | Phenotype profiles, cluster heatmaps, latent-space visualization, representative cases, and stability analyses |
| Clinical implementation area in sepsis | AI role | Human-in-the-loop design | Potential risk without human oversight |
|---|---|---|---|
| Screening and diagnosis | Flags possible sepsis, estimates infection probability, highlights concerning patterns, and summarizes supporting evidence | The clinician reviews the relevant input data, supporting evidence, and uncertainty; confirms whether the alert is clinically meaningful; and decides whether additional evaluation or treatment is needed | False-positive alerts may lead to unnecessary evaluation or treatment, whereas poorly contextualized alerts can increase alert fatigue |
| Prognostic prediction | Tracks the risk of shock, organ failure, ICU transfer, death, or recovery and identifies worsening or improving risk patterns | The clinician interprets the estimated risk in the context of bedside assessment, treatment response, comorbidities, goals of care, and information not captured by the model | Risk estimates may be treated as deterministic or may influence subsequent care, potentially reinforcing self-fulfilling predictions |
| Treatment and decision support | Suggests or prioritizes possible actions, such as fluids, vasopressor escalation, corticosteroid consideration, antibiotic optimization, source control, or ICU transfer, without autonomously determining treatment | The clinician verifies the indication, contraindications, prior treatment response, bedside findings, and goals of care, and then approves, modifies, or overrides the recommendation | Recommendations may overlook contraindications, prior treatment response, bedside findings, or goals of care and may be followed despite uncertain treatment benefit |
| LLM-assisted documentation and communication | Generates source-linked chart summaries, handoff drafts, recommendation rationales, trial-screening summaries, or patient-specific guideline summaries | The clinician checks the underlying source, source version, patient-specific accuracy, and final wording and remains responsible for the final documentation or communication | Hallucinated, outdated, or weakly sourced information may enter clinical documentation or communication without verification |
| Phenotyping for clinical stratification | Suggests clinically meaningful subgroups, such as high-risk organ dysfunction patterns, host response patterns, or potential treatment response profiles | Clinicians review whether the suggested phenotype is consistent with bedside findings, infection source, organ dysfunction, and available biomarkers before using it to guide monitoring intensity, consultation, or trial referral | Unstable or poorly interpretable phenotypes may be treated as clinically meaningful subgroups without sufficient bedside or biological support |
| Research stage | Human role and key decisions | Why oversight matters |
|---|---|---|
| Clinical or research question and intended use | Define the clinical or research question, target population, care setting, prediction timing, target outcome, and intended action before model development | This prevents AI-first framing before a clinically meaningful question has been defined; unclear time zero, target population, or outcome definitions reduce interpretability |
| Ethical, privacy, and accountability considerations | Confirm IRB approval, privacy protection, deidentification, secure data storage, data access control, accountability, and audit trails | Privacy breaches, re-identification risk, inappropriate data access, unclear responsibility, and lack of auditability can limit safe and trustworthy AI research |
| Data source, quality, and timing | Define the data source, collection period, variable definitions, missing data handling, measurement units, timestamps, and data availability at the intended prediction time | Clinical workflows determine what is measured and recorded; workflow-driven missingness, inconsistent timing, or postprediction information may produce misleading associations or apparent performance |
| Model development and performance evaluation | Select the model family, preprocessing strategy, train/validation/test split, tuning approach, internal validation, and performance metrics appropriate to the intended use | AUROC or AUPRC alone is insufficient; poor calibration, class imbalance, overfitting, and lack of subgroup evaluation can limit clinical usefulness; flexible preprocessing, feature construction, and model tuning may obscure bias or overfitting if their clinical and temporal implications are not reviewed |
| External and temporal validation | Evaluate model performance across different time periods, institutions, patient groups, care settings, and practice environments when feasible | Good internal performance does not guarantee transportability across institutions, populations, or practice patterns |
| Clinical interpretation and transparent reporting | Clinicians and AI researchers jointly assess clinical plausibility, explainability, actionability, and reporting completeness | Plausible explanations do not establish causality or clinical correctness; novel predictors and phenotypes require clinical interpretation and reproducibility assessment |
| Clinical implementation and postdeployment monitoring | Define workflow integration, user interface, human review process, prospective evaluation, performance drift monitoring, update policy, and model-withdrawal criteria | A technically accurate model may fail if it is not actionable, poorly integrated into workflow, or not monitored after deployment; deployment may change clinical behavior and the data-generating process; performance drift, subgroup errors, excessive alert burden, unintended workflow effects, and safety signals may therefore emerge over time |
Many cautions listed for screening, prognosis, treatment, and phenotyping reflect established methodological concerns in sepsis research that may be inherited, amplified, or obscured by AI. Hallucinations are associated with generative AI, whereas retrieval failure and weak source fidelity are particularly relevant to RAG-based systems.
AI, artificial intelligence; ECG, electrocardiography; EHR, electronic health record; ICU, intensive care unit; LLM, large language model; RAG, retrieval-augmented generation.
Tabular EHR data refer to structured, row-by-column clinical data such as demographics, vital signs, laboratory results, medications, organ support, timestamps, and outcomes. Cox models include conventional Cox proportional hazards models with fixed covariates and time-dependent Cox models with covariates updated over time.
AI, artificial intelligence; GAM, generalized additive model; EBM, explainable boosting machine; EHR, electronic health record; XGBoost, Extreme Gradient Boosting; LightGBM, Light Gradient Boosting Machine; CatBoost, Categorical Boosting; SHAP, Shapley Additive Explanations; LIME, Local Interpretable Model-agnostic Explanations; ICE, individual conditional expectation; ICU, intensive care unit; RNN, recurrent neural network; GRU, gated recurrent unit; LSTM, long short-term memory; TCN, temporal convolutional network; CNN, convolutional neural network; ECG, electrocardiography; RAG, retrieval-augmented generation; LLM, large language model;
AI, artificial intelligence; ICU, intensive care unit; LLM, large language model.
AI, artificial intelligence; IRB, institutional review board; AUROC, area under the receiver operating characteristic curve; AUPRC, area under the precision-recall curve.