跳到正文
原文
Nature 健康护理主题·· 29 天前精选AI 评分70

脓毒症AI Clinician建议与临床实际处置的偏离在3万例ICU患者中常见,并与90天死亡率相关

Discordance between actual care and recommendations by a reinforcement learning model for sepsis

AI 导读

在美国14家医院的30,678例脓毒症ICU患者中,临床团队处置与强化学习模型AI Clinician的建议在69%的4小时时段内存在偏离,仅31%完全一致。偏离评分与乳酸、SOFA评分等病情严重程度指标正相关;在SOFA>6的最重患者中,偏离评分升高与90天死亡率升高独立相关(OR=1.11,95%CI 1.05–1.18),该关联主要来自静脉液体决策而非升压药。

推荐理由

在14家医院的3万余名脓毒症ICU患者中量化临床处置与强化学习模型建议的偏离程度,为模型安全评估提供参照。

正文

Introduction

Sepsis is common and deadly, with more than 50 million cases annually worldwide and in-hospital mortality near 15%1,2. Given the substantial heterogeneity in septic shock biology and clinical presentation, hemodynamic support with vasopressors and intravenous fluids requires an individual approach, for which optimal tools are lacking3,4,5.

A new approach to individualize vasopressor and intravenous fluids in sepsis is guided by computational reinforcement learning (RL)6,7,8. RL algorithms offer unique advantages, but may provide implausible care recommendations and can lead to potential harm9,10,11,12,13,14,15,16. In one example, the “AI Clinician” model was associated with improved estimated mortality among sepsis patients receiving intravenous fluids and vasopressors during intensive care compared to clinician-directed care8. Although offline evaluation revealed discordance was present between actual care and AI Clinician recommendations, it is less clear how often this occurs and the independent association of discordance with outcome. These data are critical for model refinement, safety evaluation, and implementation at the bedside.

To address this knowledge gap, we studied discordance between the clinical team-administered care and AI Clinician recommendations among intensive care unit (ICU) patients with sepsis in 14 hospitals within a US healthcare system. We examined discordance in recommendations for vasopressors and intravenous fluid administration, and the association between discordance and outcome among strata of illness severity.

Results

Patients

A total of 30,678 ICU patients who met Sepsis-3 criteria (mean age 64 years (SD, 16 years), 51% male) were included (Fig. 1, Table 1). At ICU admission, the mean Elixhauser comorbidity score was 5.2 (SD, 2.3) and the sequential organ failure assessment (SOFA) score was 6.7 (SD, 3.3). More patients received mechanical ventilation (45%) than were administered vasopressors (33%). A total of 5359 (18%) of patients were admitted for a surgical intervention, and 4704 (15%) died during admission.

Fig. 1: Flow diagram of patient accrual.

UPMC University of Pittsburgh Medical Center, SOFA Sequential Organ Failure Assessment, ICU Intensive Care Unit.

Table 1 Clinical characteristics

Full size table

Epidemiology of discordance between clinical team and AI Clinician recommendations

Among the 30,678 patients, there were a total of 309,415 epochs during the first 48 hours of intensive care. The median epoch-level discordance score was 2 [IQR, 0–3.2], range 0–5.66 (Supplementary Fig. 2). The clinical team and AI clinician agreed in 31% of epochs (N = 94,667), and any discordance was present in 69% (n = 214,748). Discordance between the clinical team and the AI Clinician was most frequently observed in epochs when the clinical team opted for both vasopressors and intravenous fluids (N = 12,964, 99%). (Fig. 2).

Fig. 2: Distribution of epoch-level discordance scores in common clinical scenarios.

A Depicts the distribution of discordance scores when the clinical team administers neither intravenous fluid nor vasopressors. B Depicts the distribution of discordance scores when the clinical team only administers intravenous fluid. C Depicts the distribution of discordance scores when the clinical team only administers vasopressors. D Depicts the distribution of discordance scores when the clinical team administers both intravenous fluids and vasopressors.

Examination of discordance for vasopressors and intravenous fluids

Discordance was most common for decisions regarding IV fluids. For example, the clinical team administered less fluids than recommended by the AI Clinician in 79,458 epochs (26% of total), while clinicians administered more fluids than recommended by the AI Clinician in 121,879 epochs (39% of total) (Supplementary Fig. 3). The clinical team care and AI Clinician recommendations agreed on the dose of vasopressors in over 80% of time epochs (254,958 of 309,415, 82%) (Supplementary Fig. 3).

Association between discordance and markers of illness severity

Discordance scores correlated positively with markers of illness severity, including epoch-level serum lactate and SOFA score (Fig. 3), but not the time of the epoch during admission. In the falsification analysis, an increasing discordance was not correlated with age, baseline Elixhauser score, or epoch-specific sodium level (Supplementary Fig. 4).

Fig. 3: Change in discordance score across the distribution of serum lactate, SOFA score, and time.

A Depicts range of discordance scores over initial serum lactate levels. B Depicts the range of discordance scores over initial SOFA scores on admission. C Depicts range of discordance scores over time-epochs. SOFA Sequential Organ Failure Assessment

Association between level of discordance and 90-day mortality by stratum of illness severity

Among patients in the lowest (SOFA \(\le\) 4) or middle (4 \( < \) SOFA \(\le\) 6) SOFA strata, there was no association between discordance and mortality (Supplementary Fig. 5, Supplementary Table 2). There was an independent association between an increase in the discordance score and greater mortality among patients in the highest SOFA stratum (SOFA \( > \) 6, OR = 1.11, 95%CI: 1.05–1.18) (Fig. 4, Supplementary Table 2c). When restricting analysis to the vasopressor axis, there was no association between discordance and 90-day mortality (Fig. 4, Supplementary Fig. 5, Supplementary Table 3a–c). For intravenous fluids, though, a greater discordance score was associated with greater 90-day mortality in patients from the highest SOFA strata (SOFA \( > \) 6) (Fig. 4, Supplementary Table 4a–c). The models were similar in sensitivity analyses (Supplementary Fig. 6, Supplementary Table 5a–c). Observed mortality increased as averaged discordance score per patient increased (Supplementary Fig. 7).

Fig. 4: Adjusted associations between discordance score rodinals and 90-day mortality.

(A) and by vasopressor axis (B) and intravenous fluid axis (C) stratified by highest SOFA tertile (SOFA > 6). Odds ratios greater than one indicate a positive relationship between discordance score and mortality. Odds ratios less than one indicate a negative relationship between discordance score and mortality. SOFA Sequential organ Failure Assessment

Discussion

In a multicenter cohort of adult patients with sepsis, we observed that discordance between clinical team care and AI Clinician recommendations was common, and greater discordance was associated with 90-day mortality. These findings were most common in patients with greater illness severity and clinician decisions for intravenous fluid administration.

Reinforcement learning models are proposed for individualized treatment during intensive care. For instance, the AI Clinician was associated with reduced mortality, on average, when recommending lower intravenous fluid volume and higher vasopressor doses compared to the clinical team. The current study extends this work by examining the epidemiology of discordance between clinician actions and AI recommendations. Our data suggest that discordance is frequent, particularly among the most ill patient groups. Compared to patients with less discordance, greater discordance is associated with risk-adjusted odds of 90-day mortality. An important caveat is that this analysis did not specify the “correct” course of action in that epoch but only quantified disagreement.

The observations have several possible explanations. First, clinical practice is heterogeneous, and clinical practice guidelines provide uncertain guidance for management of shock. Most randomized trials for shock resuscitation are neutral, and a wide distribution of clinical decisions is not unexpected17,18. Second, the study used four-hour epochs for the AI Clinician, which may not capture the full resolution at which clinical decisions are reached in practice. The most severe patients in shock may decompensate within minutes, leading to observed discordance with RL models. Third, it is possible that the AI Clinician model may yield less certain estimates when derived from data in which treatment extremes and missing data are more common.

There are implications for the refinement and implementation of RL models. First, the AI Clinician was reported to have greater benefit when patients received both vasopressors and intravenous fluids administration; however, these are also moments when discordance was most common8. Discordance may offer a dynamic metric to evaluate on-demand RL model performance, a feature seen as increasingly important11,12,14,19. Second, the greater discordance in IV fluid decisions is noteworthy. Intravenous fluids are one of the most common drugs administered to patients in the ICU20, but data are unclear on how much volume, which target, and at what rate20,21. Future RL models may need to offer treatment recommendations in a more rapid cadence with specific fluid volumes so as to inform these dynamic decisions. Third, discordance was more frequent amongst patients with higher illness severity, and this may represent a target population for clinical implementation of RL models. Fourth, our discordance score is generalizable to any future RL-based algorithm with any N x N action space. Future studies may help expand the clinical impact of directionality on discordance or focus on action spaces that have more clinical relevance.

This study has many limitations. First, the method of calculating discordance between RL models and clinical care is one approach of many and may not translate to clinically meaningful differences in intravenous fluid or vasopressor administration. Other approaches may incorporate the directionality of the relationship or use non-linear models. The directionality of the discordance score has clinical plausibility for significance, and we could not ascertain whether discordance impacts mortality equally bidirectionality. Second, we enrolled patients at a single health care system, albeit multiple heterogeneous hospitals, which may limit the external validity of the findings. Third, this observational study is subject to unmeasured confounding in the relationship between discordance and outcome. Fourth, we only evaluated one RL use case. Other RL models may have different epidemiology and impact of discordance.

In adult patients with sepsis, discordance between clinical team care and reinforcement learning model recommendations was common and associated with 90-day mortality.

Methods

This study was approved by the University of Pittsburgh Institutional Review Board (IRB 20010238) with a waiver of informed consent.

Overview

The AI Clinician is a computational clinical decision support instrument that recommends a precision treatment policy for intravenous fluid and vasopressor support in septic patients. On average, selection of a treatment policy congruent with the AI Clinician recommendations is associated with improved clinical outcome8. The current investigation extends prior work by ascertaining clinical scenarios in which the clinical team and AI Clinician recommendations were discordant, quantifying the amount of discordance, and understanding the independent association of discordance with outcome. Since resuscitation decisions may be more difficult in sicker patients, the analysis was a priori stratified by those with or without severe organ dysfunction as measured by the sequential organ failure assessment (SOFA) score.

Patients

The electronic health records (EHR) of adults from a healthcare system of 14 academic and community hospitals in western Pennsylvania from January 1, 2013 to December 31, 2017 were studied. Patients were included if they met sepsis criteria within six hours of hospital arrival and were admitted to the ICU (i.e., community-acquired sepsis). A six-hour timeframe was selected to capture a homogeneous cohort of exclusively community-acquired sepsis patients. Sepsis-3 criteria included: (i) suspected infection as defined by the combination of antibiotic administration and body fluid culture sampling and (ii) the presence of organ dysfunction, defined by a SOFA score of two or more22. Sepsis onset was identified by the timing of fluid culture sampling or antibiotic administration, whichever occurred first. Patients were excluded if they met any of the following criteria: (i) unavailable intravenous fluid intake/output or medication administration data, (ii) a do not resuscitate (DNR) or do not intubate (DNI) status at any point of the index hospitalization, or (iii) a hospital stay of less than eight hours. A random sample of 80% of the cohort was used for model development, and the remaining 20% was used for model validation.

AI clinician model

The AI Clinician model was implemented in UPMC data using the originally published 46 out of the 48 variables including static measurements such as demographics (e.g, age, sex, weight) and baseline Elixhauser comorbidity score as well as dynamic measurements such as vital signs (e.g. heart rate, systolic blood pressure, temperature), laboratory measurements (e.g. creatinine, lactate, bicarbonate), intravenous fluid and vasopressor administration (e.g. norepinephrine equivalents, intravenous fluid intake and fluid balance)23,24. Extracted data were organized in non-overlapping, four-hour time blocks relative to sepsis onset. Data were extracted from sepsis onset to 48 hours following initial presentation. We used the worst available measurement (minimum or maximum, as clinically relevant) for variables with multiple samplings. A time-limited, parameter-specific, sample-and-hold approach or iterative imputation by random forests with predictive mean matching were used to impute missing data points (Supplementary Table 6)8,25,26. The primary outcome was 90-day mortality, ascertained from discharge disposition as well as Social Security Death Index 27,28,29.

A Markov decision process was used to model patient experience and included a progression of states in response to actions. A state is defined as the clinical status of a patient based on available static and dynamic measurements in each 4-hour epoch. Patient states unique to our cohort were created using K-means clustering of time-series data. Optimization of the number of states was determined using several analyses, including: Akaike and Bayesian information criteria, within-state sum of squared errors, proportion of variance explained, and state size (at least 50 observations per state were required). Actions are defined as treatment options, in which the AI Clinician used a five-by-five matrix of varying vasopressor and intravenous fluid choices (Supplementary Table 1). No restriction on vasopressor and intravenous fluid choices was placed on the AI Clinician. Different action space cutoffs were selected compared to the original model to ensure a more equal distribution of patients between actions. A policy value is the sum of rewards (+100) and penalties (−100) given the value of a specific action that minimizes 90-day mortality. An exploration actor (epsilon) of 10% was used to permit dynamic learning in which a random action was chosen to facilitate model exploration.

Discordance score

To define discordance between the clinical team and AI Clinician actions, we determined a discordance score in each 4-hour epoch (Supplementary Fig. 1). The action space was a five-by-five grid, where each cell represented a unique combination of intravenous fluid (row) and vasopressor (column) dose levels. Each treatment action was mapped to a coordinate on this symmetric grid, and discordance was calculated as the Euclidean distance between the positions of the clinical team and the AI Clinician. Discordance scores ranged from 0 (complete agreement) to a maximum of 5.66.

Statistical analysis

We conducted analyses among patients as well as descriptive analyses at the level of the 4-hr epoch. Patient characteristics were summarized using descriptive statistics. Categorical variables were reported as frequencies (N) and percentages (%), while continuous variables were presented as means with standard deviations (SD) or medians with interquartile ranges (IQR), as appropriate.

To understand the epidemiology of discordant recommendations between the clinical team and AI Clinician, we used five steps: (i) descriptive epidemiology of discordance between the clinical team and AI Clinician, (ii) examination of discordance for vasopressors and intravenous fluids, (iii) exploration of association between individual markers of illness severity (e.g., lactate and SOFA score) and discordance, (iv) multivariable modeling at the patient-level of the association between discordance and 90-day mortality by pre-specified stratum of illness severity (SOFA \(\le\) 4, 4 \( < \) SOFA \(\le\) 6, and SOFA \( > \) 6) and v.) sensitivity analyses.

First, we qualitatively describe the distribution of discordance scores. We studied common clinical scenarios: (a) clinical team takes no action, (b) clinical team administers intravenous (IV) fluids only, (c) clinical team administers vasopressors only, (d) clinical team administers both IV fluids and vasopressors. We studied the univariate association between discordance score and epoch-level SOFA score or serum lactate. We conducted a falsification analysis examining the relationship between epoch-level sodium level, baseline Elixhauser score, age, and discordance with similar models.

We modeled the independent association between discordance score and 90-day mortality. The proportional odds assumption was violated, and we used a mixed multinomial logistic model (Supplementary Eq. 1). Although 90-day mortality was the primary outcome, discordance score was modeled as the outcome variable for convergence. Reversal of the outcome and predictor variables does not alter the underlying nature of the association or interpretation of the odds ratio. We calculated an individual ordinal discordance score per epoch and accounted for multiple epochs per patient using a logistic regression model with a random intercept for patient identifier. The discordance score category was modeled as the ordinal outcome. An odds ratio greater than one indicates a higher adjusted odds of being in a higher discordance score category. Consequently, for the predictor of the dichotomous outcome variable of 90-day mortality, the reference group was patients who survived to 90 days; therefore, the odds ratio compares patients who died within 90 days versus those who survived. Adjustment variables included time from admission (days), age, sex, baseline Elixhauser score = 0, and initial serum lactate. We a priori stratified regression analyses by SOFA score, using separate models for each tertile (SOFA \(\le\) 4, 4 \( < \) SOFA \(\le\) 6, and SOFA \( > \) 6). We repeated the models to evaluate discordance for individual treatments (vasopressor or intravenous fluids only). We performed sensitivity analyses restricting to patients who received at least one treatment by the clinical team, excluding those who never received fluids or vasopressors.

All statistical tests were two-sided, and a p-value < 0.05 was considered statistically significant. Analyses were performed using Stata version 18.0 (StataCorp LLC), and figures were created using GraphPad Prism version 10 (GraphPad Software).

Data availability

The dataset is available upon request from the corresponding authors and subsequent institutional board approval.

Code availability

The code is available upon request from the corresponding authors.

References

  1. Rudd, K. E. et al. Global, regional, and national sepsis incidence and mortality, 1990-2017: analysis for the Global Burden of Disease Study. Lancet 395, 200–211 https://doi.org/10.1016/S0140-6736(19)32989-7 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

  2. Rhee, C. et al. Incidence and trends of Sepsis in US hospitals using clinical vs claims data, 2009-2014. JAMA 318, 1241–1249 https://doi.org/10.1001/jama.2017.13836 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  3. Evans, L. et al. Executive summary: Surviving Sepsis Campaign: International Guidelines for the Management of Sepsis and Septic Shock 2021. Crit. Care Med. 49, 1974–1982 https://doi.org/10.1097/CCM.0000000000005357 (2021).

    Article  PubMed  Google Scholar 

  4. Faust, J. S. Moving beyond the Centers for Medicare & Medicaid Services’ “severe sepsis and septic shock early management bundle” core quality measure. Ann. Emerg. Med. 78, 20–26 https://doi.org/10.1016/j.annemergmed.2021.03.003 (2021).

    Article  PubMed  Google Scholar 

  5. Barbash, I. J. et al. Treatment patterns and clinical outcomes after the introduction of the medicare Sepsis Performance Measure (SEP-1). Ann. Intern. Med https://doi.org/10.7326/M20-5043

  6. Shimabukuro, D. W., Barton, C. W., Feldman, M. D., Mataraso, S. J. & Das, R. Effect of a machine learning-based severe sepsis prediction algorithm on patient survival and hospital length of stay: a randomised clinical trial. BMJ Open Respir. Res. 4, e000234 https://doi.org/10.1136/bmjresp-2017-000234 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

  7. Wu, X., Li, R., He, Z., Yu, T. & Cheng, C. A value-based deep reinforcement learning model with human expertise in optimal treatment of sepsis. npj Digit. Med 6, 15 https://doi.org/10.1038/s41746-023-00755-5 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  8. Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C. & Faisal, A. A. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat. Med. 24, 1716–1720 https://doi.org/10.1038/s41591-018-0213-5 (2018).

    Article  CAS  PubMed  Google Scholar 

  9. Crowe, B. & Rodriguez, J. A. Identifying and addressing bias in artificial intelligence. JAMA Netw. Open 7, e2425955 https://doi.org/10.1001/jamanetworkopen.2024.25955 (2024).

    Article  PubMed  Google Scholar 

  10. Ratwani, R. M., Sutton, K. & Galarraga, J. E. Addressing AI algorithmic bias in health care. JAMA 332, 1051 https://doi.org/10.1001/jama.2024.13486 (2024).

    Article  PubMed  Google Scholar 

  11. Celi, L. A. et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities: A global review. PLOS Digit Health 1, e0000022 https://doi.org/10.1371/journal.pdig.0000022 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

  12. Ranard, B. L. et al. Minimizing bias when using artificial intelligence in critical care medicine. J. Crit. Care 82, 154796 https://doi.org/10.1016/j.jcrc.2024.154796 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  13. Zack, T. et al. Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study. Lancet Digit Health 6, e12–e22 https://doi.org/10.1016/S2589-7500(23)00225-X (2024).

    Article  CAS  PubMed  Google Scholar 

  14. Iqbal, U., Hsu, Y.-H. E., Celi, L. A., Li, Y.-C. J. Artificial intelligence in healthcare: Opportunities come with landmines. BMJ Health Care Inform. 31 https://doi.org/10.1136/bmjhci-2024-101086.

  15. Wong, A. et al. External validation of a widely implemented proprietary sepsis prediction model in hospitalized patients. JAMA Intern. Med. 181, 1065–1070 https://doi.org/10.1001/jamainternmed.2021.2626 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

  16. Garnacho-Montero, J. & Martín-Loeches, I. Clinical management of sepsis can be improved by artificial intelligence: no. Intensive Care Med. 46, 378–380 https://doi.org/10.1007/s00134-020-05947-1 (2020).

    Article  PubMed  Google Scholar 

  17. National Heart, Lung, and Blood Institute Prevention and Early Treatment of Acute Lung Injury Clinical Trials Network Early restrictive or liberal fluid management for sepsis-induced hypotension. N. Engl. J. Med. 388, 499–510 https://doi.org/10.1056/NEJMoa2212663 (2023).

    Article  Google Scholar 

  18. Lasik, E., Tardy, C. & Jacobs, F. M. Restriction of intravenous fluid in ICU patients with septic shock. N. Engl. J. Med. 387, 857 https://doi.org/10.1056/NEJMc2210366 (2022).

    Article  PubMed  Google Scholar 

  19. Classen, D. C., Longhurst, C. & Thomas, E. J. Bending the patient safety curve: how much can AI help? npj Digit. Med. 6, 2 https://doi.org/10.1038/s41746-022-00731-5 (2023).

    Article  PubMed  PubMed Central  Google Scholar 

  20. Finfer, S., Myburgh, J. & Bellomo, R. Intravenous fluid therapy in critically ill adults. Nat. Rev. Nephrol. 14, 541–557 https://doi.org/10.1038/s41581-018-0044-0 (2018).

    Article  PubMed  Google Scholar 

  21. Finfer, S. et al. Resuscitation fluid use in critically ill adults: an International Cross-sectional Study in 391 Intensive Care Units. Crit. Care 14, R185 https://doi.org/10.1186/cc9293 (2010).

    Article  PubMed  PubMed Central  Google Scholar 

  22. Seymour, C. W. et al. Assessment of clinical criteria for sepsis: for the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA 315, 762–774 https://doi.org/10.1001/jama.2016.0288 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  23. Elixhauser, A., Steiner, C., Harris, D. R. & Coffey, R. M. Comorbidity measures for use with administrative data. Med Care 36, 8–27 https://doi.org/10.1097/00005650-199801000-00004 (1998).

    Article  CAS  PubMed  Google Scholar 

  24. Brown, S. M. et al. Survival after shock requiring high-dose vasopressor therapy. Chest 143, 664–671 https://doi.org/10.1378/chest.12-1106 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  25. M M. Fast Imputation of Missing Values. 2019. Accessed August 2, 2024 https://CRAN.R-project.org/package=missRanger

  26. Hug, C.W. Detecting Hazardous Intensive Care Patient Episodes Using Real-Time Mortaluty Models. Dissertation. Massachussets Institute of Technology. Accessed June 1, 2025. https://dspace.mit.edu/entities/publication/3c905f71-1e3b-4b2e-996c-23fd2c227ca9

  27. Kellum, J. A. et al. Understanding the inflammatory cytokine response in pneumonia and sepsis: results of the Genetic and Inflammatory Markers of Sepsis (GenIMS) Study. Arch. Intern Med. 167, 1655–1663 https://doi.org/10.1001/archinte.167.15.1655 (2007).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  28. Hill, M. E. & Rosenwaike, I. The Social Security Administration’s Death Master File: the completeness of death reporting at older ages. Soc. Secur Bull. 64, 45–51 (2001).

    PubMed  Google Scholar 

  29. Seymour, C. W. et al. Derivation, validation, and potential treatment implications of novel clinical phenotypes for sepsis. JAMA 321, 2003–2017 https://doi.org/10.1001/jama.2019.5791 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  30. Mehta, H. B. et al. Development and validation of the summary elixhauser comorbidity score for use with ICD-10-CM-coded data among older adults. Ann. Intern Med. 175, 1423–1430 https://doi.org/10.7326/M21-4204 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

We would like to acknowledge the significant contribution of the patients, families, researchers, clinical staff, Biostatistical and Data Management Core (BDMC), and the Clinical Research Investigation and Systems Modeling of Acute Illness (CRISMA) Center of the University of Pittsburgh Department of Critical Care Medicine.

Funding

This work received funding support from the National Institutes of Health for Dr. Seymour (R35GM119519).

Author information

Authors and Affiliations

  1. Division of Pulmonary, Allergy, Critical Care, and Sleep Medicine, Department of Medicine, University of Pittsburgh Medical Center, Pittsburgh, PA, USA

    Peter C. Nauka

  2. Department of Critical Care Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA

    Jason N. Kennedy, Bryan J. McVerry, Derek C. Angus & Christopher W. Seymour

  3. Clinical Research, Investigation and Systems Modeling of Acute Illness (CRISMA) Center, Pittsburgh, PA, USA

    Jason N. Kennedy, Derek C. Angus & Christopher W. Seymour

  4. Department of Biostatistics, University of Pittsburgh, Pittsburgh, PA, USA

    Xinlei Chen, Chung-Chou H. Chang & Lu Tang

  5. Department of Surgery and Cancer, Imperial College London, London, UK

    Matthieu Komorowski

  6. Department of Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA

    Chung-Chou H. Chang

  7. Department of Anesthesia and Perioperative Care, Zuckerberg San Francisco General Hospital and Trauma Center, University of California San Francisco, San Francisco, CA, USA

    Romain Pirracchio

  8. Department of Critical Care Medicine, VA Pittsburgh Healthcare System, Pittsburgh, PA, USA

    Bryan J. McVerry

  9. Department of Computer Science, University of Pittsburgh, Pittsburgh, PA, USA

    Milos Hauskrecht

  10. Department of Emergency Medicine, University of Pittsburgh School of Medicine, Pittsburgh, PA, USA

    Christopher W. Seymour

Authors

  1. Peter C. Nauka
  2. Jason N. Kennedy
  3. Xinlei Chen
  4. Matthieu Komorowski
  5. Chung-Chou H. Chang
  6. Romain Pirracchio
  7. Lu Tang
  8. Bryan J. McVerry
  9. Derek C. Angus
  10. Milos Hauskrecht
  11. Christopher W. Seymour

Contributions

Dr Nauka conceived the project, performed the data analysis, drafted figures, and drafted and edited the manuscript. Dr Kennedy performed data analysis and edited portions of the manuscript. Ms Chen and Dr Chang conceived the project, assisted and supervised data analysis, and edited the manuscript. Drs. Komorowski, Pirrachio, Tang, McVerry, Angus and Hauskrecht conceived the project and edited the manuscript. Dr Seymour conceived the project, drafted figures, supervised data analysis, and drafted and edited the manuscript.

Corresponding author

Correspondence to Christopher W. Seymour.

Ethics declarations

Competing interests

Dr. Seymour reports personal fees from Octapharma, Deepull, Beckman Coulter, Sobi, and Edwards Inc., unrelated to the current manuscript. He reports funding for travel to conferences from the International Sepsis Forum, Monash University, and the ISICEM. He is Associate Editor at JAMA.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Nauka, P.C., Kennedy, J.N., Chen, X. et al. Discordance between actual care and recommendations by a reinforcement learning model for sepsis. npj Health Syst. 3, 101 (2026). https://doi.org/10.1038/s44401-026-00156-2

Download citation

  • Received: 09 June 2026

  • Accepted: 01 September 2026

  • Published: 22 October 2026

  • Version of record: 22 October 2026

  • DOI: https://doi.org/10.1038/s44401-026-00156-2

来源:Nature 健康护理主题 · nature.com