Diagnostic suspicion bias and machine learning: Breaking the awareness deadlock for sepsis detection.

Many early warning algorithms are downstream of clinical evaluation and diagnostic testing, which means that they may not be useful when clinicians fail to suspect illness and fail to order appropriate tests. Depending on how such algorithms handle missing data, they could even indicate "low risk" simply because the testing data were never ordered. We considered predictive methodologies to identify sepsis at triage, before diagnostic tests are ordered, in a busy Emergency Department (ED). One algorithm used "bland clinical data" (data available at triage for nearly every patient). The second algorithm added three yes/no questions to be answered after the triage interview. Retrospectively, we studied adult patients from a single ED between 2014-16, separated into training (70%) and testing (30%) cohorts, and a final validation cohort of patients from four EDs between 2016-2018. Sepsis was defined per the Rhee criteria. Investigational predictors were demographics and triage vital signs (downloaded from the hospital EMR); past medical history; and the auxiliary queries (answered by chart reviewers who were blinded to all data except the triage note and initial HPI). We developed L2-regularized logistic regression models using a greedy forward feature selection. There were 1164, 499, and 784 patients in the training, testing, and validation cohorts, respectively. The bland clinical data model yielded ROC AUC's 0.78 (0.76-0.81) and 0.77 (0.73-0.81), for training and testing, respectively, and ranged from 0.74-0.79 in four hospital validation. The second model which included auxiliary queries yielded 0.84 (0.82-0.87) and 0.83 (0.79-0.86), and ranged from 0.78-0.83 in four hospital validation. The first algorithm did not require clinician input but yielded middling performance. The second showed a trend towards superior performance, though required additional user effort. These methods are alternatives to predictive algorithms downstream of clinical evaluation and diagnostic testing. For hospital early warning algorithms, consideration should be given to bias and usability of various methods.

Modeling of Usual Care: Vasopressor Initiation for Sepsis With Hypotension.

Usual care regarding vasopressor initiation is ill-defined. We aimed to develop a quantitative "dynamic practice" model for usual care in the emergency department (ED) regarding the timing of vasopressor initiation in sepsis. In a retrospective study of 589 septic patients with hypotension in an urban tertiary care center ED, we developed a multi-variable model that distinguishes between patients who did and did not subsequently receive sustained (>24 h) vasopressor therapy. Candidate predictors were vital signs, intravenous fluid (IVF) volumes, laboratory measurements, and elapsed time from triage computed at timepoints leading up to the final decision timepoint of either vasopressor initiation or ED hypotension resolution without vasopressors. A model with six independently significant covariates (respiratory rate, Glasgow Coma Scale score, SBP, SpO2, administered IVF, and elapsed time) achieved a C-statistic of 0.78 in a held-out test set at the final decision timepoint, demonstrating the ability to reliably model usual care for vasopressor initiation for hypotensive septic patients. The included variables measured depth of hypotension, extent of disease severity and organ dysfunction. At an operating point of 90% specificity, the model identified a minority of patients (39%) more than an hour before actual vasopressor initiation, during which time a median of 2,250 (IQR 1,200-3,300) mL of IVF was administered. This single-center analysis shows the feasibility of a quantitative, objective tool for describing usual care. Dynamic practice models may help assess when management was atypical; such tools may also be useful for designing and interpreting clinical trials.

Validation of a 'Usual Care' Model for Vasopressor Initiation in a Cohort of Emergency Department Patients with Sepsis.

Usual care regarding vasopressor (VP) initiation is ill-defined. We aimed to further validate a quantitative model for usual care in the Emergency Department (ED) regarding the timing of VP initiation in sepsis. We retrospectively studied a cohort of adult critically-ill ED patients who also received antibiotics in the ED. We applied a multivariable model previously developed from another patient cohort which distinguishes between time points at which patients were or were not subsequently started on a continuous VP infusion. The model has six independently significant predictors (respiratory rate, Glasgow Coma Scale score, systolic blood pressure, SpO2, administered intravenous fluids, and elapsed time). The outcome was initiation of VP infusion, either within the ED or within 6 hours after leaving the ED. We applied the model to all time points, beginning when all model input parameters were first available for a given patient, and ending when either VP were first started, or the patient left the ED. Out of 55,963 adult ED patients during the two-year study interval, we identified 1,629 who met our inclusion criteria. The area under the receiver operating characteristic curve was 0.81 for all patients, and 0.72 for the subset with at least one hypotensive blood pressure measurement. At a model threshold with sensitivity and specificity 0.74 and 0.74, respectively, the median advance detection time was 170.5 minutes (IQR 53 - 363).

Cardiac surgical outcome prediction by blood pressure variability indices Poincaré plot and coefficient of variation: a retrospective study.

BACKGROUND: Recent literature suggests a significant association between blood pressure variability (BPV) and postoperative outcomes after cardiac surgery. However, its outcome prediction ability remains unclear. Current prediction models use static preoperative patient factors. We explored the ability of Poincaré plots and coefficient of variation (CV) by measuring intraoperative BPV in predicting adverse outcomes. METHODS: In this retrospective, observational, cohort study, 3687 adult patients (> 18 years) undergoing cardiac surgery requiring cardio-pulmonary bypass from 2008 to 2014 were included. Blood pressure variability was computed by Poincare plots and CV. Standard descriptors (SD) SD1, SD2 were measured with Poincare plots by ellipse fitting technique. The outcomes analyzed were the 30-day mortality and postoperative renal failure. Logistic regression models adjusted for preoperative and surgical factors were constructed to evaluate the association between BPV parameters and outcomes. C-statistics were used to analyse the predictive ability. RESULTS: Analysis found that, 99 (2.7%) patients died within 30 days and 105 (2.8%) patients suffered from in-hospital renal failure. Logistic regression models including BPV parameters (standard descriptors from Poincare plots and CV) performed poorly in predicting postoperative 30-day mortality and renal failure [Concordance(C)-Statistic around 0.5]. They did not add any significant value to the standard STS risk score [C-statistic: STS alone 0.7, STS + BPV parmeters 0.7]. CONCLUSIONS: In conclusion, BP variability computed from Poincare plots and CV were not predictive of mortality and renal failure in cardiac surgical patients. Patient comorbid conditions and other preoperative factors are still the gold standard for outcome prediction. Future directions include analysis of dynamic parameters such as complexity of physiological signals in identifying high risk patients and tailoring management accordingly.

Clustering Blood Pressure Trajectories in Septic Shock in the Emergency Department.

Hemodynamic management of sepsis in the emergency department relies on fluid resuscitation and vasoactive therapy to maintain adequate blood pressure and end-organ perfusion. While typical practice targets certain thresholds of blood pressure (such as 65 mmHg mean arterial or 90 mmHg systolic blood pressure [SBP]), little consideration is given to temporal dynamics of blood pressure. In this work, we use unsupervised learning methods to reveal characteristic SBP trajectories in the two hours either surrounding the start of hypotensive episodes (SBP <; 90 mmHg) or immediately preceding the initiation of vasopressor therapy. Our results show that hypotensive episodes tended to either resolve very quickly (within 40 minutes) or extend for prolonged periods (at least 1 hour). Those with prolonged hypotension constituted 74% of all patients with at least one measurement of SBP <; 90 mmHg. Of them, patients who entered hypotension by a large, acute drop from a normal SBP over the preceding hour had a greater incidence of subsequent vasopressor administration than those with a more gradual decline into hypotension. Overall, our results suggest that a significant subset of patients, especially those with stable but low SBP, should have received vasopressors when they did not, or should have received them sooner. Dynamic trajectories appear to be important factors in clinical practice of hemodynamic management of sepsis.

Characterizing Fluid Response and Sepsis Progression in Emergency Department Patients.

Sepsis and septic shock are major global public health concerns. The main therapies for sepsis-related hypotension are fluid resuscitation and vasopressor therapy, though it can be challenging to determine the amount of fluid that should be given or the optimal timing to transition to vasopressor administration. To characterize patients' response to fluid bolus therapy (FBT) and analyze the sepsis progress using multiple vital signs, we mined a database containing 761 patients presenting to the Emergency Department (ED) with vital signs and laboratory values indicating high risk of septic shock. By clustering the patients' mean arterial pressure (MAP) time series during a time window around FBT, we found that clusters showing fluid responsiveness during the two hours after FBT only included about 25% of studied boluses. In addition, MAP responses tended to vary based on the initial MAP level. We also found that the trajectories of heart rate and MAP in a 2-D plane demonstrated general trends related to the hemodynamic progression of sepsis and previously described phases of septic shock. Potentially compensatory and decompensatory responses of the cardiovascular system to the insults of sepsis were reflected in the clusters representative of different phases.

Blood Pressure Coefficient of Variation and Its Association With Cardiac Surgical Outcomes.

BACKGROUND: Multiple studies completed in the ambulatory nonsurgical setting show a significant association between short- and long-term blood pressure variability and poor outcomes. However, perioperative blood pressure variability outcomes have not been well studied, especially in the cardiac surgical setting. In this study, we sought to assess whether systolic and mean arterial blood pressure variability were associated with 30-day mortality and in-hospital renal failure in patients undergoing cardiac surgery requiring cardiopulmonary bypass. Furthermore, blood pressure variability has not been evaluated specifically during each phase of surgery, namely in the pre-, intra- and postbypass phases; thus, we aimed also to assess whether outcomes were associated with phase-specific systolic and mean arterial blood pressure variability. METHODS: All patients undergoing cardiac surgery from January 2008 to June 2014 were enrolled in this retrospective, single-center study. Demographic, intraoperative, and postoperative outcome data were obtained from the institution's Society of Thoracic Surgery database and Anesthesia Information Management System. Systolic and mean arterial blood pressure variability were assessed using the coefficient of variation (CV). The primary outcomes were 30-day mortality and in-hospital renal failure in relation to the entire duration of a case, while the secondary outcomes assessed phase-specific surgical periods. In an effort to control the family-wise error rate, P values <.0125 were considered significant for the primary outcomes. RESULTS: Of the 3687 patients analyzed, 2.7% of patients died within 30 days of surgery and 2.8% experienced in-hospital renal failure. After adjusting for significant covariates, we found a statistically significant association between increasing CV for systolic blood pressure (CVSBP) and 30-day mortality and in-hospital renal failure. For every 0.10 increase in CVSBP, there was a 150% increase in the odds of death (odds ratio, 2.50; 95% confidence interval, 1.60-3.92; P < .0001) and there was a 104% increase in odds of experiencing renal failure (odds ratio, 2.04; 95% confidence interval, 1.33-3.14; P = .001). The association with mortality was driven primarily by the prebypass period, because the association between CVSBP and mortality during the prebypass phase was significant (P = .01), and not during the postbypass phase (P = .08). There was no significant association between CV for mean arterial blood pressure and either death or renal failure during any period of surgery, including the bypass phase. CONCLUSIONS: Increasing systolic blood pressure variability was associated with 30-day mortality and development of renal failure, with surgery phase-specific relationships observed. Further research is required to determine how to prospectively detect blood pressure variability and elucidate opportunities for intervention.

Prediction of postoperative outcomes using intraoperative hemodynamic monitoring data.

Major surgeries can result in high rates of adverse postoperative events. Reliable prediction of which patient might be at risk for such events may help guide peri- and postoperative care. We show how archiving and mining of intraoperative hemodynamic data in orthotopic liver transplantation (OLT) can aid in the prediction of postoperative 180-day mortality and acute renal failure (ARF), improving upon predictions that rely on preoperative information only. From 101 patient records, we extracted 15 preoperative features from clinical records and 41 features from intraoperative hemodynamic signals. We used logistic regression with leave-one-out cross-validation to predict outcomes, and incorporated methods to limit potential model instabilities from feature multicollinearity. Using only preoperative features, mortality prediction achieved an area under the receiver operating characteristic curve (AUC) of 0.53 (95% CI: 0.44-0.78). By using intraoperative features, performance improved significantly to 0.82 (95% CI: 0.56-0.91, P = 0.001). Similarly, including intraoperative features (AUC = 0.82; 95% CI: 0.66-0.94) in ARF prediction improved performance over preoperative features (AUC = 0.72; 95% CI: 0.50-0.85), though not significantly (P = 0.32). We conclude that inclusion of intraoperative hemodynamic features significantly improves prediction of postoperative events in OLT. Features strongly associated with occurrence of both outcomes included greater intraoperative central venous pressure and greater transfusion volumes.

Classification models to predict vasopressor administration for septic shock in the emergency department.

Optimal management of sepsis and septic shock in the emergency department (ED) involves timely decisions related to intravenous fluid resuscitation and initiation of vasoactive medication support. A decision-support tool trained on electronic health record data, can help improve this complex decision. We retrospectively extracted vital signs, lab measurements, and fluid administration information from 807 patient visits over a two-year period to a major ED. Patients selected for inclusion had a high likelihood of septic shock. We trained binary classifiers to discriminate between patients administered vasopressors in the ED and those not administered vasopressors at any point. Using features extracted from the entire ED visit record yielded a maximum area under the receiver-operating characteristic curve (AUC) of 0.798 (95% CI 0.725-0.849) in a hold-out test set. In a separate task, we used individual vital signs observations with lab results to predict vasopressor administration, yielding a maximum AUC of 0.762 (95% CI 0.748-0.777). Lastly, we trained separate classifiers for different subgroups of vital signs observations. These subgroups were defined by the cumulative number of fluid boluses delivered at the time of the observation. The maximum AUC achieved by any of these classifiers was 0.815 (95% CI 0.784-0.853), occurring for vital signs observations made after 2 bolus administrations. Classifiers in all tasks significantly outperformed existing clinical tools for assessing prognosis in ED sepsis. This work shows how relatively few features can provide instantaneous and accurate prediction of need for an intervention that is typically a complex clinical decision.

Intraoperative hemodynamics predict postoperative mortality in orthotopic liver transplantation.

Liver transplantation remains the only curative treatment option for a variety of end-stage liver diseases. Prediction of major adverse events following surgery has traditionally focused on static predictors that are known prior to surgery. The effects of intraoperative management can now be explored due to the archiving of high-resolution monitoring data. We extracted intraoperative hemodynamic trend data of 55 patients undergoing orthotopic liver transplantation (OLT) and computed 12 features from the systolic arterial blood pressure (ABP), cardiac index, central venous pressure (CVP), and stroke volume variation (SVV) signals. Using a logistic regression classifier with a leave-one-out cross-validation procedure, we selected subsets of these features to predict mortality up to 180 days after surgery. Best performance was achieved with a combination of 3 features - median absolute deviation (MAD) of ABP, median CVP, and time spent with SVV <; 10% - reaching an area under the receiver-operating characteristic (or c-statistic) of 0.808. Odds ratios (OR) computed from the coefficients of the multivariate logistic regression model constructed from these features showed that greater time spent with SVV <; 10% (OR = 0.981 min(-1), p = 0.001) and greater MAD of systolic ABP (OR = 0.696 mmHg(-1), p = 0.026) were significantly associated with survival. Adding preoperative measures such as age and serum concentrations of albumin, bilirubin, and creatinine failed to improve performance of the prediction model. These results show that the course of intraoperative hemodynamics can predict 180-day mortality after OLT.

Macro- and microscale variables regulate stent haemodynamics, fibrin deposition and thrombomodulin expression.

Drug eluting stents are associated with late stent thrombosis (LST), delayed healing and prolonged exposure of stent struts to blood flow. Using macroscale disturbed and undisturbed fluid flow waveforms, we numerically and experimentally determined the effects of microscale model strut geometries upon the generation of prothrombotic conditions that are mediated by flow perturbations. Rectangular cross-sectional stent strut geometries of varying heights and corresponding streamlined versions were studied in the presence of disturbed and undisturbed bulk fluid flow. Numerical simulations and particle flow visualization experiments demonstrated that the interaction of bulk fluid flow and stent struts regulated the generation, size and dynamics of the peristrut flow recirculation zones. In the absence of endothelial cells, deposition of thrombin-generated fibrin occurred primarily in the recirculation zones. When endothelium was present, peristrut expression of anticoagulant thrombomodulin (TM) was dependent on strut height and geometry. Thinner and streamlined strut geometries reduced peristrut flow recirculation zones decreasing prothrombotic fibrin deposition and increasing endothelial anticoagulant TM expression. The studies define physical and functional consequences of macro- and microscale variables that relate to thrombogenicity associated with the most current stent designs, and particularly to LST.

Microfluidic fabrication of stable nanoparticle-shelled bubbles.

We introduce a microfluidic approach to generating monodisperse, stable nanoparticle-shelled bubbles using air-in-oil-in-water (A/O/W) compound bubbles as templates. The oil phase of the A/O/W compound bubbles comprises a volatile organic solvent and a hydrophobic silica nanoparticle. Upon evaporation of the organic solvent, the nanoparticles in the oil layer form a stiff shell at the air-water interface, which drastically enhances the stability of the bubbles against dissolution and coarsening. On the basis of this approach, we demonstrate that it is also possible to generate functional bubbles stabilized by composite shells that are composed of mixtures of hydrophobic materials and nanoparticles with unique properties.