Search

Sepsis is a leading cause of ICU mortality, and early detection is critical for improving patient outcomes. However, existing machine learning models often rely on hourly aggregated data, limiting their ability to capture rapid physiological changes, and frequently lack interpretability, reducing clinical trust and usability. This paper proposes a conceptual framework that integrates Temporal Convolutional Networks (TCNs) with an attention mechanism to analyze high-frequency, minute-level vital sign data for early sepsis prediction. The architecture includes a data input layer, a TCN-based feature extractor with causal dilated convolutions and residual connections, an attention module for identifying clinically relevant time points and variables, and a prediction head that estimates the risk of sepsis within a 6-hour horizon. The proposed approach enables efficient parallel processing, improved temporal sensitivity, and enhanced interpretability compared to recurrent models. While offering advantages in real-time prediction and explainability, challenges remain in handling missing data, ensuring generalizability across ICUs, and minimizing false alarms for clinical deployment.

Over the past five years, sepsis prediction models have reported strong retrospective performance, often exceeding AUROC 0.85–0.90 by leveraging vital signs, laboratory data, and machine learning to predict sepsis earlier than clinical recognition. However, despite these results, bedside adoption remains minimal, and external or prospective validations frequently show substantial performance decline, with clinicians still relying on traditional criteria such as qSOFA and SIRS. This position paper argues that AUROC is an insufficient and potentially misleading metric for clinical deployment, as it reflects retrospective rank discrimination rather than real-world utility, calibration, or actionable impact. High AUROC scores often conceal poor threshold selection, excessive alert burden, and clinically unacceptable alarm fatigue, while retrospective evaluations create an overly optimistic view that fails in real-time settings. We propose shifting evaluation toward clinically meaningful metrics such as net benefit, alert burden per patient-day, and number needed to alert at clinician-defined thresholds, alongside earlier incorporation of workflow requirements. Ultimately, the continued dominance of AUROC-centric evaluation represents a systemic mismatch between model development and clinical reality, limiting sepsis prediction tools from achieving meaningful impact at the bedside.

Sepsis remains a major cause of mortality in intensive care units, largely due to delayed recognition and the limitations of current machine learning models that rely on retrospective, static electronic health record data. Although these models often show strong offline performance, their clinical translation is constrained by mismatches between training conditions and real-time bedside environments. Most existing systems depend on hourly aggregates or batch processing, introducing delays that reduce their usefulness within the narrow therapeutic window for intervention. In contrast, continuous vital sign streams generated by modern bedside monitors represent an underused source of real-time physiological information. This perspective argues that effective sepsis prediction requires a shift toward edge AI architectures that enable low-latency, privacy-preserving inference directly at the point of care. By treating physiological signals as continuous data streams rather than static records, and by deploying computation at the bedside instead of centralized cloud systems, models can better align with clinical realities. Such an approach could improve early detection, reduce alert fatigue through more context-aware predictions, and mitigate privacy, latency, and bandwidth challenges associated with cloud-based solutions. Ultimately, transitioning from retrospective modeling to real-time, edge-enabled decision support represents a necessary evolution in clinical AI, requiring close collaboration between clinicians, engineers, and data scientists to enable deployable, trustworthy, and timely sepsis prediction systems.

Sepsis remains a major cause of mortality in intensive care units worldwide, with an estimated 49 million cases and over 11 million deaths annually, highlighting the need for earlier detection to improve outcomes. This systematic review synthesizes evidence on machine learning models for early sepsis prediction in adult ICU patients from 2017 to 2021, focusing on prediction horizons, data modalities, and validation approaches. A comprehensive search of PubMed, Embase, IEEE Xplore, ACM Digital Library, and arXiv identified studies meeting criteria for ICU-based sepsis prediction with at least a 4-hour forecast window, following PRISMA guidelines. Of 1,478 records screened, 35 studies were included, with prediction horizons ranging from 4 to 24 hours and most relying on hourly vital sign data and internal validation. Reported performance varied widely depending on horizon length, data sampling, and validation rigor, with external validation generally producing lower but more realistic results. Overall, while machine learning models show promising predictive ability, limitations in generalizability and standardization remain, emphasizing the need for stronger validation frameworks and reporting practices to support clinical translation.

Sepsis prediction models perform poorly when transferred between ICUs due to demographic and practice variation, leading to substantial performance drops caused by differences in patient populations, admission criteria, and data recording standards, which limits reliable deployment across healthcare systems. Retraining models from scratch requires large labeled datasets that many ICUs lack due to cost, time, and resource limitations, making it difficult for low-resource settings to develop or adopt effective predictive tools. We propose a meta-learning approach that enables rapid adaptation of sepsis prediction models using few-shot local data, leveraging pre-training across multiple ICUs to support fast personalization in new environments. The framework includes meta-training across diverse source ICUs to learn a generalizable initialization and meta-adaptation at the target ICU using only a few gradient updates on limited data, enabling efficient few-shot learning. This approach improves sepsis prediction in low-resource and heterogeneous ICU settings by reducing data requirements and increasing robustness to demographic shifts, supporting more equitable access to AI tools in critical care. The proposed framework enables efficient and fair deployment of sepsis prediction models across diverse ICUs, bridging resource gaps and improving scalability and adaptability of clinical AI systems globally.

Sepsis prediction models in intensive care units often degrade over time due to changes in clinical practice, patient populations, and data recording processes, a phenomenon known as model drift that can compromise patient safety. Traditional federated learning approaches are not well-suited to these evolving conditions, as they assume static data distributions and typically require costly retraining that risks forgetting previously learned knowledge, while also being constrained by privacy limitations that prevent central data pooling. To address these challenges, this paper proposes a federated continual learning framework that enables ongoing, privacy-preserving model adaptation across multiple hospitals without catastrophic forgetting. The framework integrates local continual learning methods (such as elastic weight consolidation or memory replay) with federated aggregation and importance-weighted parameter updates to support continuous learning from new clinical data while preserving prior knowledge. This design allows each institution to adapt models to local data shifts while collaboratively improving a shared global model without sharing patient-level data. Overall, the proposed approach offers a scalable solution for maintaining robust, adaptive sepsis prediction systems in dynamic healthcare environments, reducing the need for repeated full retraining and supporting long-term clinical deployment.

Sepsis continues to be a major contributor to morbidity and mortality among hospitalized patients globally, especially within intensive care and emergency departments, where rapid recognition is essential for improving survival through timely treatment. In recent years, machine learning approaches have gained attention for their ability to predict sepsis onset using routinely collected electronic health record data. This systematic review, conducted in accordance with PRISMA 2020 guidelines, synthesizes evidence from studies published between 2017 and 2025, focusing on model architectures, feature selection and engineering strategies, prediction time horizons, and validation methodologies. Searches across major biomedical and informatics databases identified 67 eligible studies. The included literature shows that logistic regression, ensemble tree-based algorithms, and deep learning models are most frequently applied for sepsis prediction tasks. However, the majority of studies rely on retrospective datasets with internal validation, while only a limited number incorporate prospective or real-world validation frameworks. Overall, although reported model performance is often strong in retrospective analyses, a consistent decline in accuracy is observed when models are evaluated in real clinical environments. These findings highlight that prospective validation and improved generalizability are still underdeveloped areas, underscoring the need for future research to emphasize real-time deployment and robust external validation before clinical integration.