Finding time zero: how smart diagnostics and AI are reshaping sepsis care

Introduction

Few conditions managed in the emergency department are as consequential, or as conceptually unsettled, as sepsis. We define it as infection-induced, dysregulated, and self-injurious host response; however, bedside clinicians intimately understand the diagnostic uncertainty surrounding both how it manifests and when it starts. Its opaque starting point and constellation of protean clinical findings profoundly differentiate sepsis from other time-critical diagnoses around which we have organized emergency care. Trauma, stroke, and ST-elevation myocardial infarction each rests on a discrete lesion and an objective confirmatory test. Meanwhile, sepsis identification is usually fraught with both negative cultures and heterogeneous signs and symptoms that fluctuate according to where each unique patient's immune system sits along the illness continuum. Essentially, we must decisively treat a syndrome we cannot reliably characterize.

The inherent challenges

Four compounding difficulties conspire against accurate and timely sepsis diagnosis.

First, partially because over 80% of sepsis cases arise in the community, its true “time zero,” or onset, remains unknown (1). Patients’ presenting severity of illness can, therefore, vary dramatically according to when care is eventually sought, which can vary according to many factors, including a person's awareness of the syndrome and its consequences and access to medical care. Regardless, study designs, bundled therapies, and quality metric assessments largely presume a clean inception point that biology rarely provides.

Second, sepsis is a syndrome, not a disease. The patients we lump together are biologically disparate – e.g., different pathogens invading the body via different routes, different immune states and trajectories, different physiological phenotypes. Further aggravating this problem is the siloed nature of patients’ many medical records, which limit the ability to contextualize a patient's acute problem according to their past medical history. This abundant diversity, which no human could reasonably fully integrate into a diagnostic evaluation, contributes to missed or delayed diagnosis. Moreover, the inability to rapidly and accurately characterize each patient's unique immunological state while septic limits the ability to personalize immunological therapies.

Third, identifying the sepsis-triggering pathogen often remains elusive. Blood cultures have served as the reference standard for over a century, yet the invading organism emerges from culture bottles in only 14% of patients admitted to the hospital for community-acquired sepsis (2). This woefully inadequate performance may result from a lack of live circulating pathogens from either antibiotic administration prior to culture acquisition, the intrinsic immune system clearing pathogens, insufficient blood sample volumes, localized infections lacking bacteremia, or the presence of fastidious or viral organisms, and many other causes (3). Though molecular diagnostic tools, such as PCR and MALDI-TOF, can quickly characterize a pathogen, they can only be deployed in that minority of blood samples yielding positive cultures, which usually require nearly two days to become positive.

Fourth, because real-time high-quality clinical care demands a high-functioning health system, challenges inherent in any such system, including human, environmental, and process factors, can limit optimal performance. Clinicians, each with unique training, experience, knowledge, and physical and mental states, approach problems in a somewhat unique manner, often arriving at similar diagnoses at different times. Similarly, patients carry their own unique medical, social, educational, language, and mental backgrounds, all contributing to clinicians’ ability to secure a timely and accurate syndromic diagnosis like sepsis. For example, consider a novice emergency medicine physician new to a very busy urban hospital. That clinician may struggle to optimally manage an immunocompromised, nonverbal, and combative patient with limited IV access sites. Conducting an adequate physical examination becomes difficult, and arriving at a timely sepsis diagnosis more difficult still. Environmental factors similarly contribute to the diagnostic capabilities of an ED staff increasingly burdened with routine environmental distractions (e.g., monitors beeping, urgent patient needs, serial interruptions from “best practice alerts” with poor positive predictive value or little applicability to the case at hand), inadequate equipment, or limited staffing that delays task accomplishments (4, 5). Additionally, process inefficiencies embedded within care pathways or local hospital culture can contribute to overcrowding and other problems that restrict diagnostic effectiveness (4, 6).

Sepsis outcomes are time-dependent, and the longer effective source control either from antimicrobial use or mechanical removal of an infection and support care are delayed, the worse patients do. This therapeutic urgency generates a justified pressure to act, and payers have hard-wired that pressure into practice through the CMS SEP-1 measure. A mandate rewarding speed drives reflexive broad-spectrum antibiotic initiation rather than targeting therapy according to the pathogen.

The clinical message is uncomfortable but clear: we commit to empiric, often broad-spectrum antibiotic therapy while genuinely uncertain whether infection is present, and, if present, which organism drives the infection. Additionally, due to challenges in diagnosing sepsis, we often detect it well after its onset and lack tools to consistently and meaningfully personalize immunologics or other therapies to attenuate the adverse effects of the syndrome. Nonspecific, excessive antibiotic use carries greater consequences for patients (e.g., drug side effects, microbiome alterations, superinfections), hospitals (e.g., costs, resource utilization), and society (e.g., antimicrobial resistance). In other words, empiric antibiotic-induced problems result from inadequate pathogen diagnostics, and generic bundled care mandates result from lack of real-time immunological characterization.

Why this moment is different

Despite inherent diagnostic challenges to optimally manage sepsis, emerging tools provide reason for measured optimism. At-home and wearable “smart” devices, which generate both continuous physiologic and biochemical data, move detection upstream, improving prediction of hard endpoints such as unplanned readmission. The feasibility of utilizing outpatient wearable data to enhance early outpatient sepsis detection has been demonstrated in low- and middle-income countries, and the United Kingdom's National Health Service has recently adopted plans to deploy this technology nationally for similar purposes (7–9). These same signals could surface the earliest physiologic drift of an evolving infection towards sepsis, providing the closest we are likely to come to a legitimate “time zero.” This at-home identification could be accomplished using acute data augmented by the background knowledge of a patient's medical history rather than by a busy clinician in the trenches deciphering vague findings without access to comprehensive longitudinal data or by a chart abstractor weeks later.

With simultaneous interrogation, novel pathogen-identification and sample processing methods and host-response diagnostics engage sepsis's defining duality directly. For example, direct-from-blood pathogen diagnostic tools, though currently expensive and logistically cumbersome to access from most regions, are valuable adjuncts to overcome existing limitations of blood culture (10, 11). As their associated costs inevitably decrease and access becomes more widespread, the use of such platforms is likely to expand. Additionally, recent advances in bodily fluid sample processing methods are enhancing the speed and accuracy of pathogen identification (12, 13). From the perspective of rapid host response characterization, various research and commercial applications have been deployed to designate the likelihood of infection and, if present, the likelihood of a viral or bacterial origin (14). One platform in particular utilizes a compilation of transcriptomic (mRNA) data to generate a five-point graded scale to report the likelihood of infection and infection type, as well as to prognosticate illness severity (i.e., seven-day need for ICU-level care) (15). These various tools may help distinguish true infection from its many mimics and offer a path to genuine stewardship at a time when most septic patients never yield a positive culture.

Artificial intelligence (AI) has the opportunity to enhance every aspect of the sepsis diagnostic process. Whether from infection detection (e.g., database-assessed epidemiological surveillance, radiographic image analysis), pathogen detection (e.g., automated agar plate interpretation), pathogen identification (e.g., digital microscopy to recognize malaria, mass spectrometric pattern recognition following a positive blood culture), antimicrobial susceptibility and antimicrobial selection optimization, host response characterization integrated into a digital twin testing environment, sepsis therapeutic guidance, or sepsis detection & prognostication platforms, AI is already being deployed to enhance accuracy and speed along different aspects of the sepsis management process (16–19).

Modern clinical models used for sepsis detection integrate multimodal, high-dimensional data impossible for any clinician to hold in working memory and resolve it into actionable subgroups. To date, multiple sepsis-detecting platforms have been deployed, each with a unique AI model, study design to assess predictive capabilities, and resultant diverse performance characteristics, all of which restrict the ability to directly compare one to another (Table 1). Epic's Sepsis Model (ESM) is the most widely utilized and the most independently scrutinized, with a 2021 external validation analysis of its first version (ESM v1) finding poor real-world performance (33% sensitivity, 83% specificity, 12% positive predictive value [PPV], area under the receiver operating characteristic curve [AUC] 0.63) and a similar 2026 multicenter prospective analysis of ESM v2 also finding modest performance (PPV 13%–26%, AUC 0.82–0.92 across 227,091 encounters at four US health systems, with high institutional variability); the v2 evaluation documented low positive predictive value and high alert burden, prompting its authors to recommend alert-silencing strategies (20, 21). A retrospective study of over 600 patients presenting to the emergency department with sepsis at a different hospital found that the ESM v1 and v2 models yielded AUCs of 0.70 and 0.85, respectively, but, more importantly each model alerted the clinical team of the possibility of sepsis in two-thirds of cases after those patients were already being managed for the syndrome (22). The ESM has not been FDA-authorized.

Table 1. Key characteristics of current leading AI sepsis prediction tools.

In a multicenter study of ∼3,500 adults with suspected infection (∼2,400 derivation; ∼400 internal validation; ∼700 external validation), Prenosis’ Sepsis ImmunoScore used up to 22 data inputs to assess sepsis risk within 24 h (23). The model secured FDA de novo authorization by achieving AUCs of 0.85, 0.80, and 0.81 across these cohorts, utilizing a validation that explicitly separated predictive cases (sepsis onset after test output) from diagnostic cases (sepsis onset before output) (24). Rather than a single binary alert, the ImmunoScore FDA documentation details four discrete risk tiers (Low, Medium, High, Very High) where the predictive probability of sepsis and critical illness appropriately scales. In the pure predictive analysis (i.e., sepsis onset after test output), the ImmunoScore maintained a statistically significant step-wise increase in predictive value across risk categories (Low: 2.17%, Medium: 5.52%, High: 14.22%, Very High: 50.0%). A subsequent manufacturer-involved evaluation of over 6,000 adults demonstrated that the ImmunoScore outperformed common clinical scores (SOFA, NEWS, qSOFA, and SIRS) and individual biomarkers (25).

Bayesian Health's Targeted Real-time Early Warning System (TREWS) originated in 2015 retrospective derivation work using cases from the MIMIC-II ICU database, in which the Modified Early Warning Score (MEWS) served as the comparator for progression to septic shock; that work established operating characteristics but presented no evidence that TREWS could alter the outcome it predicted (26). The system was later evaluated prospectively across five hospitals, where among 6,877 sepsis patients identified by the alert before antibiotic initiation, provider confirmation within three hours was associated with an adjusted absolute mortality reduction of 3.3% and a relative reduction of 18.7% (27). Independent commentary cautioned that this observational design cannot establish causal benefit, noting that the investigators themselves acknowledged the need for a randomized controlled trial but found it difficult to operationalize and that no RCT-level evidence yet demonstrates clinical benefit from AI-based sepsis alerts (28). A published but non-peer-reviewed exchange sharpened the point – a competing developer group argued the arms were assigned so as to credit the system with early treatments while attributing delays to clinicians, without adjusting for process-of-care confounders such as time to serum lactate result; the original authors replied that this rested on unrealistic assumptions about provider behavior and showed that adding such variables did not substantively alter their results (29). Separately, when Bayesian's related Sepsis Flagging Device obtained FDA 510(k) clearance using the ImmunoScore as a predicate, validation on an all-comer population yielded a flag-level positive predictive value of 11.7% at roughly 3% prevalence, and the device missed its pre-specified encounter-level positive percent agreement endpoint (lower confidence bound 74.2% against a required 80%), clearing on totality of evidence (30). These clearances assess safety, substantial equivalence, and diagnostic agreement; they are not causal designs capable of showing that a model's predictive output alters clinical trajectories or improves outcomes.

Clairyon's COMPOSER (COnformal Multidimensional Prediction Of SEpsis Risk) platform, a deep-learning sepsis model, yielded a 17% relative reduction in hospital mortality when deployed in two EDs with over 6,300 patients analyzed using a quasi-experimental study design (31). Model enhancements with a large language model (COMPOSER-LLM) that ingests unstructured clinical data (e.g., clinical notes) and generates a differential diagnosis tool (COMPOSER-LLM_DDx) to account for sepsis mimics was subsequently integrated into the platform and assessed in 2,500 adult ED patients (1,746 [with sepsis] in the retrospective cohort; 754 [with sepsis] in the prospective cohort). COMPOSER-LLM_DDx demonstrated a PPV of 52.9% using retrospective data but the PPV increased to ∼58% when applied prospectively in real-world ED settings, all while also cutting false alarms (32). Despite these promising findings, the COMPOSER-LLM_DDx model has not yet been submitted for FDA approval.

Real-world prospective, externally validated evaluations of AI technology are certainly warranted to determine its impact on meaningful clinical outcomes (33). Regardless of the model being used, the richer, cleaner, and more robust the data input, including the free text that structured models discard, the more systems can distinguish signal from noise, as high false-positive rates will only add to burdensome alarm fatigue that only degrades staff morale and quality medical care (34). Not only does the quality of those data matter but the frequency and clinical relevance of data collection similarly is important. For example, vital signs are often checked in the inpatient wards no more than every four hours while they are assessed at least every hour in the ICU, providing a sooner opportunity to detect a concerning deviation from an expected clinical course; interestingly, outpatient “smart” wearable collect data continuously.

The sources of those data also impact the performance and logistical practicality of the AI model; for example, if a model relies on the integration of a laboratory value (e.g., lactate) already being ordered for the suspicion of sepsis or clinical deterioration, the model could expect to report greater diagnostic accuracy but is unlikely to prompt an intervention that would not have already been taken. Moreover, if particular data needed for an AI algorithm are slow to arrive due to either lab processing delays (e.g., in the event of high patient volumes) or case complexity resulting in prolonged diagnostic suspicion, that AI model will be unable to generate its prediction in a timely fashion, potentially notifying a clinician to draw blood cultures and administer antibiotics after those interventions have already been performed. Because these real-world challenges may not manifest in retrospective analyses, rigorous prospective studies are necessary prior to confidently adopting these AI diagnostic tools. However, when hospitals evaluate the adoption of AI sepsis models, those decisions should not be based on convenience (e.g., simply “turning on another feature” of an electronic health record already in use) but instead on the underpinning scientific rigor and practical clinical impacts of that platform on patients and clinicians caring for them. Meaningful adoption (i.e., consistently reliable use by bedside clinicians and the promotion of actionable interventions) and the realization of desired clinical outcomes (e.g., faster appropriate antimicrobial administration and stewardship, more effective triage) demand strategic integration of implementation science principles into model deployment.

Toward precision and synergy

The greatest gains will not come from any one of these advances in isolation. Consider the patient whose entire medical history is queried for a personalized understanding of what findings deviate from a healthy state, whose wearables flag physiologic drift and prompt earlier assessment, whose bodily fluids provide the substrate for rapid pathogen identification, whose host-response assay characterizes the immune and physiologic phenotype, and whose data feed an AI model that contemporaneously matches their therapy to their unique biology. Each tool answers a distinct failure mode – earlier detection for time-dependence, subphenotyping for heterogeneity, host-pathogen diagnostics for characterization – but, in combination, a consistently timely, accurate, and personalized sepsis diagnosis is possible.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Lorenzo Pelagatti, Careggi University Hospital, Italy

Reviewed by: Chaitali Singhal, Translational Health Science and Technology Institute (THSTI), India

Smitesh Padte, WellSpan Health, United States