Background: Patients with sepsis-induced hypotension are generally treated with a combination of intravenous fluids and vasopressors. The attributes of patients receiving a liberal compared to a restrictive fluid strategy have not been fully characterized. We use machine learning (ML) techniques to identify key predictors of restrictive versus liberal fluids strategy, and the likelihood of receiving each strategy in distinct patient phenotypes. Methods: We performed a retrospective observational study of patients at Emory University Hospital from 2014 to 2021 that were hypotensive, met Sepsis-3 criteria, and received at least 1 L of intravenous crystalloid fluids. We excluded patients with nonseptic etiologies of hypotension. Supervised ML techniques were used to identify key predictors for the two strategies. Additionally, subset analyses were performed on patients with pneumonia, congestive heart failure (CHF), or chronic kidney disease (CKD). Using unsupervised ML techniques, we also identified three distinct sepsis-induced hypotension phenotypes and evaluated their likelihood of receiving either strategy. Results: We identified N = 15,292 patients and randomly split them into training (n = 12,233) and validation (n = 3,059) datasets. XGBoost was the most accurate model (AUC: 0.84) for predicting the strategies. While worse oxygenation was the strongest predictor of utilizing a restrictive fluid strategy, top predictors of a liberal fluid strategy included higher pulse and blood urea nitrogen. In subset analyses, CHF, CKD, and pneumonia were predictive of restrictive fluid strategy. We identified three distinct sepsis-induced hypotension phenotypes: 1) mild organ injury, 2) severe hypoxemia, and 3) renal dysfunction. Conclusions: We identified key predictors of restrictive versus liberal fluids strategy and distinct patient phenotypes for sepsis-induced hypotension.

Injury Severity Score (ISS) has limited utility as a prospective predictor of trauma outcomes as it is currently scored by abstractors post-discharge. This study aimed to determine accuracy of ISS estimation at time of admission. Attending trauma surgeons assessed the Abbreviated Injury Scale of each body region for patients admitted during their call, from which estimated ISS (eISS) was calculated. The eISS was considered concordant to abstracted ISS (aISS) if both were in the same category: mild (<9), moderate (9-15), severe (16-25), or critical (>25). Ten surgeons completed 132 surveys. Overall ISS concordance was 52.2%; 87.5%, 30.8%, 34.8%, and 61.7% for patients with mild, moderate, severe, and critical aISS, respectively; unweighted k = .36, weighted k = .69. This preliminarily supports attending trauma surgeons’ ability to predict severity of injury in real time, which has important clinical and research implications.

Upadhyaya, Pulakesh1; Wang, Jeffrey2; Souza De Vale, Joao3; Lisboa, Felipe4; Schobel, Seth4; Elster, Eric4; Dente, Christopher5; Buchman, Timothy6; Kamaleswaran, Rishikesan7 Author Information

Background:Septic patients who develop acute respiratory failure (ARF) requiring mechanical ventilation represent a heterogenous subgroup of critically ill patients with widely variable clinical characteristics. Identifying distinct phenotypes of these patients may reveal insights about the broader heterogeneity in the clinical course of sepsis. We aimed to derive novel phenotypes of sepsis-induced ARF using observational clinical data and investigate their generalizability across multi-ICU specialties, considering multi-organ dynamics. Methods:We performed a multi-center retrospective study of ICU patients with sepsis who required mechanical ventilation for ≥24 hours. Data from two different high-volume academic hospital systems were used as a derivation set with N=3,225 medical ICU (MICU) patients and a validation set with N=848 MICU patients. For the multi-ICU validation, we utilized retrospective data from two surgical ICUs at the same hospitals (N=1,577). Clinical data from 24 hours preceding intubation was used to derive distinct phenotypes using an explainable machine learning-based clustering model interpreted by clinical experts. Results:Four distinct ARF phenotypes were identified: A (severe multi-organ dysfunction (MOD) with a high likelihood of kidney injury and heart failure), B (severe hypoxemic respiratory failure [median P/F=123]), C (mild hypoxia [median P/F=240]), and D (severe MOD with a high likelihood of hepatic injury, coagulopathy, and lactic acidosis). Patients in each phenotype showed differences in clinical course and mortality rates despite similarities in demographics and admission co-morbidities. The phenotypes were reproduced in external validation utilizing an external MICU from second hospital and SICUs from both centers. Kaplan-Meier analysis showed significant difference in 28-day mortality across the phenotypes (p<0.01) and consistent across both centers. The phenotypes demonstrated differences in treatment effects associated with high positive end-expiratory pressure (PEEP) strategy. Conclusion:The phenotypes demonstrated unique patterns of organ injury and differences in clinical outcomes, which may help inform future research and clinical trial design for tailored management strategies.

Injury Severity Score (ISS) as a prospective predictive variable is limited, as it is scored post-discharge by registrars. We followed a phase 1 pilot investigation of the feasibility of prospective ISS estimation (eISS) by trauma surgeons within 1 day of admission with an investigation of the impact of a simple educational aid on the accuracy of these estimations. Eleven surgeons evaluated 178 patients in phase 2. With the educational aid, ISS concordance improved from 74.6% to 85.1% for non-severe (abstracted ISS, aISS <16) injuries and from 78.8% to 83.1% for severe (aISS ≥16) injuries; weighted k improved from 0.53 to 0.72. Abbreviated Injury Scale (AIS) concordance similarly improved in five of seven body regions. The ability to prospectively document ISS has important clinical and research implications. There remains opportunity to refine educational aides and harness the EHR to further improve prediction accuracy and facilitate adoption in standard clinical workflows.

Background: The Georgia Quality Improvement Program (GQIP) surgical collaborative participating hospitals have shown consistently poor performance in the post-operative sepsis category of National Surgical Quality Improvement Program data as compared with national benchmarks. We aimed to compare crude versus risk-adjusted post-operative sepsis rankings to determine high and low performers amongst GQIP hospitals. Patients and Methods: The cohort included intra-abdominal general surgery patients across 10 collaborative hospitals from 2015 to 2020. The American College of Surgeons National Surgical Quality Improvement Program (ACS-NSQIP) sepsis definition was used among all hospitals for case abstraction and NSQIP data were utilized to train and validate a multivariable risk-adjustment model with post-operative sepsis as the outcome. This model was used to rank GQIP hospitals by risk-adjusted post-operative sepsis rates. Rankings between crude and risk-adjusted post-operative sepsis rankings were compared ordinally and for changes in tertile. Results: The study included 20,314 patients with 595 cases of post-operative sepsis. Crude 30-day post-operative sepsis risk among hospitals ranged from 0.81 to 5.11. When applying the risk-adjustment model which included: age, American Society of Anesthesiology class, case complexity, pre-operative pneumonia/urinary tract infection/surgical site infection, admission status, and wound class, nine of 10 hospitals were re-ranked and four hospitals changed performance tertiles. Conclusions: Inter-collaborative risk-adjusted post-operative sepsis rankings are important to present. These metrics benchmark collaborating hospitals, which facilitates best practice exchange from high to low performers.

Introduction Accurate identification of venous thromboembolism (VTE) is critical to develop replicable epidemiological studies and rigorous predictions models. Traditionally, VTE studies have relied on international classification of diseases (ICD) codes which are inaccurate – leading to misclassification bias. Here, we developed ClotCatcher, a novel deep learning model that uses natural language processing to detect VTE from radiology reports. Methods Radiology reports to detect VTE were obtained from patients admitted to Emory University Hospital (EUH) and Grady Memorial Hospital (GMH). Data augmentation was performed using the Google PEGASUS paraphraser. This data was then used to fine-tune ClotCatcher, a novel deep learning model. ClotCatcher was validated on both the EUH dataset alone and GMH dataset alone. Results The dataset contained 1358 studies from EUH and 915 studies from GMH ( n = 2273). The dataset contained 1506 ultrasound studies with 528 (35.1%) studies positive for VTE, and 767 CT studies with 91 (11.9%) positive for VTE. When validated on the EUH dataset, ClotCatcher performed best (AUC = 0.980) when trained on both EUH and GMH dataset without paraphrasing. When validated on the GMH dataset, ClotCatcher performed best (AUC = 0.995) when trained on both EUH and GMH dataset with paraphrasing. Conclusion ClotCatcher, a novel deep learning model with data augmentation rapidly and accurately adjudicated the presence of VTE from radiology reports. Applying ClotCatcher to large databases would allow for rapid and accurate adjudication of incident VTE. This would reduce misclassification bias and form the foundation for future studies to estimate individual risk for patient to develop incident VTE.

BACKGROUND:Ultramassive transfusion (UMT) is a resource-demanding intervention for trauma patients in hemorrhagic shock, and associated mortality rates remains high. Current research has been unable to identify a transfusion ceiling or point where UMT transitions from lifesaving to futility. Furthermore, little consideration has been given to how time-specific patient data points impact decisions with ongoing high-volume resuscitation. Therefore, this study sought to use time-specific machine learning modeling to predict mortality and identify parameters associated with survivability in trauma patients undergoing UMT. METHODS:A retrospective review was conducted at a Level I trauma (2018-2021) and included trauma patients meeting criteria for UMT, defined as ≥20 red blood cell products within 24 hours of admission. Cross-sectional data were obtained from the blood bank and trauma registries, and time-specific data were obtained from the electronic medical record. Time-specific decision-tree models predicating mortality were generated and evaluated using area under the curve. RESULTS:In the 180 patients included, mortality rate was 40.5% at 48 hours and 52.2% overall. The deceased received significantly more blood products with a median of 71.5 total units compared with 55.5 in the survivors ( p < 0.001) and significantly greater rates of packed red blood cells and fresh frozen plasma at each time interval. Time-specific decision-tree models predicted mortality with an accuracy as high as 81%. In the early time intervals, hemodynamic stability, undergoing an emergency department thoracotomy, and injury severity were most predictive of survival, while, in the later intervals, markers of adequate resuscitation such as arterial pH and lactate level became more prominent. CONCLUSION:This study supports that the decision of "when to stop" in UMT resuscitation is not based exclusively on the number of units transfused but rather the complex integration of patient and time-specific data. Machine learning is an effective tool to investigate this concept, and further research is needed to refine and validate these time-specific decision-tree models. LEVEL OF EVIDENCE:Prognostic and Epidemiological; Level IV.

BACKGROUND: Thoracic injury can cause impairment of lung function leading to respiratory complications such as pneumonia (PNA). There is increasing evidence that centralmemory T cells of the adaptive immune systemplay a key role in pulmonary immunity. We sought to explorewhether assessment of cell phenotypes using flow cytometry (FCM) could be used to identify pulmonary infection after thoracic trauma. METHODS: We prospectively studied trauma patients with thoracic injuries who survived >48 hours at a Level 1 trauma center from 2014 to 2020. Clinical and FCMdata from serum samples collected within 24 hours of admission were considered as potential variables. Random forest and logistic regression models were developed to estimate the risk of hospital-acquired and ventilator-associated PNA. Variables were selected using backwards elimination, and models were internally validated with leave-one-out. RESULTS: Seventy patients with thoracic injurieswere included (median age, 35 years [interquartile range (IQR), 25.25-51 years]; 62.9% [44 of 70] male, 61.4% [42 of 70] blunt trauma). The most common injuries included rib fractures (52 of 70 [74.3%]) and pulmonary contusions (26 of 70 [37%]). The incidence of PNAwas 14 of 70 (20%). Median Injury Severity Scorewas similar for patientswith and without PNA (30.5 [IQR, 22.6-39.3] vs. 26.5 [IQR, 21.6-33.3]). The final random forest model selected three variables (Acute Physiology and Chronic Health Evaluation score, highest pulse rate in first 24 hours, and frequency of CD4(+) central memory cells) that identified PNAwith an area under the curve of 0.93, sensitivity of 0.91, and specificity of 0.88. A logistic regression with the same features had an area under the curve of 0.86, sensitivity of 0.76, and specificity of 0.85. CONCLUSION: Clinical and FCM data have diagnostic utility in the early identification of patients at risk of nosocomial PNA following thoracic injury. Signs of physiologic stress and lower frequency of central memory cells appear to be associated with higher rates of PNA after thoracic trauma.

INTRODUCTION: The pathophysiology of the inflammatory response after major trauma is complex, and the magnitude correlates with severity of tissue injury and outcomes. Study of infection-mediated immune pathways has demonstrated that cellular microRNAs may modulate the inflammatory response. The authors hypothesize that the expression of microRNAs would correlate to complicated recoveries in polytrauma patients (PtPs). METHODS: Polytrauma patients enrolled in the prospective observational Tissue and Data Acquisition Protocol with Injury Severity Score of >15 were selected for this study. Polytrauma patients were divided into complicated recoveries and uncomplicated recovery groups. Polytrauma patients' blood samples were obtained at the time of admission (T0). Established biomarkers of systemic inflammation, including cytokines and chemokines, were measured using multiplexed Luminex-based methods, and novel microRNAs were measured in plasma samples using multiplex RNA hybridization. RESULTS: Polytrauma patients (n = 180) had high Injury Severity Score (26 [20-34]) and complicated recovery rate of 33%. MicroRNAs were lower in PtPs at T0 compared with healthy controls, and bivariate analysis demonstrated that variations of microRNAs correlated with age, race, comorbidities, venous thromboembolism, pulmonary complications, complicated recovery, and mortality. Positive correlations were noted between microRNAs and interleukin 10, vascular endothelial growth factor, Acute Physiology and Chronic Health Evaluation, and Sequential Organ Failure Assessment scores. Multivariable Lasso regression analysis of predictors of complicated recovery based on microRNAs, cytokines, and chemokines revealed that miR-21-3p and monocyte chemoattractant protein-1 were predictive of complicated recovery with an area under the curve of 0.78. CONCLUSION: Systemic microRNAs were associated with poor outcomes in PtPs, and results are consistent with previously described trends in critically ill patients. These early biomarkers of inflammation might provide predictive utility in early complicated recovery diagnosis and prognosis. Because of their potential to regulate immune responses, microRNAs may provide therapeutic targets for immunomodulation.

Traumatic injuries to the mesenteric vessels are rare and often lethal. Visceral vessels, such as the superior mesenteric artery (SMA) and vein (SMV), supply blood to the small and large bowel by a rich system of collaterals. Because fewer than 100 such injuries have been described in the literature, they pose challenges in both diagnosis and management and can unfortunately result in high mortality rates. Prompt diagnosis, surgical intervention, and resuscitation can lead to improved outcomes. Here, we review the literature surrounding traumatic injuries of the SMA/SMV and discuss management strategies.

Background: An emerging body of literature supports the role of individualized prognostic tools to guide the management of patients after trauma. The aim of this study was to develop advanced modeling tools from multidimensional data sources, including immunological analytes and clinical and administrative data, to predict outcomes in trauma patients.Methods: This was a prospective study of trauma patients at Level 1 centers from 2015 to 2019. Clinical, flow cytometry, and serum cytokine data were collected within 48 hours of admission. Sparse logistic regression models were developed, jointly selecting predictors and estimating the risk of ventilator -associated pneumonia, acute kidney injury, complicated disposition (death, rehabilitation, or nursing facility), and return to the operating room. Model parameters (regularization controlling model sparsity) and performance estimation were obtained via nested leave-one-out cross-validation.Results: A total of 179 patients were included. The incidences of ventilator-associated pneumonia, acute kidney injury, complicated disposition, and return to the operating room were 17.7%, 28.8%, 22.5%, and 12.3%, respec-tively. Regarding extensive resource use, 30.7% of patients had prolonged intensive care unit stay, 73.2% had prolonged length of stay, and 23.5% had need for prolonged ventilatory support. The models were developed and cross-validated for ventilator-associated pneumonia, acute kidney injury, complicated dispositions, and return to the operating room, yielding predictive areas under the curve from 0.70 to 0.91. Each model derived its optimal predictive value by combining clinical, administrative, and immunological analyte data.Conclusion: Clinical, immunological, and administrative data can be combined to predict post-traumatic outcomes and resource use. Multidimensional machine learning modeling can identify trauma patients with complicated clinical trajectories and high resource needs.(c) 2022 Elsevier Inc. All rights reserved.