Linking Big Data and Prediction Strategies: Tools, Pitfalls, and Lessons Learned.

OBJECTIVES: Modern critical care amasses unprecedented amounts of clinical data-so called "big data"-on a minute-by-minute basis. Innovative processing of these data has the potential to revolutionize clinical prognostics and decision support in the care of the critically ill but also forces clinicians to depend on new and complex tools of which they may have limited understanding and over which they have little control. This concise review aims to provide bedside clinicians with ways to think about common methods being used to extract information from clinical big datasets and to judge the quality and utility of that information. DATA SOURCES: We searched the free-access search engines PubMed and Google Scholar using the MeSH terms "big data", "prediction", and "intensive care" with iterations of a range of additional potentially associated factors, along with published bibliographies, to find papers suggesting illustration of key points in the structuring and analysis of clinical "big data," with special focus on outcomes prediction and major clinical concerns in critical care. STUDY SELECTION: Three reviewers independently screened preliminary citation lists. DATA EXTRACTION: Summary data were tabulated for review. DATA SYNTHESIS: To date, most relevant big data research has focused on development of and attempts to validate patient outcome scoring systems and has yet to fully make use of the potential for automation and novel uses of continuous data streams such as those available from clinical care monitoring devices. CONCLUSIONS: Realizing the potential for big data to improve critical care patient outcomes will require unprecedented team building across disparate competencies. It will also require clinicians to develop statistical awareness and thinking as yet another critical judgment skill they bring to their patients' bedsides and to the array of evidence presented to them about their patients over the course of care.

Repeat lactate level predicts mortality better than rate of clearance.

BACKGROUND: Lactate clearance has been developed into a marker of resuscitation in trauma, but no study has compared the predictive power of the various clearance calculations. Our objective was to determine which method of calculating lactate clearance best predicted 24-hour and in-hospital mortality after injury. STUDY DESIGN: Retrospective chart review of patients admitted to a Level-1 trauma center directly from the scene of injury from 2010 to 2013 who survived >15min, had an elevated lactate at admission (≥3mmol/L), followed by another measurement within 24h of admission. Lactate clearance was calculated using five models: actual value of the repeat level, absolute clearance, relative clearance, absolute rate, and relative rate. Models were compared using the areas under the respective receiver operating curves (AUCs), with an endpoint of death at 24h and in-hospital mortality. RESULTS: 3910 patients had an elevated admission lactate concentration on admission (mean=5.6±3.0mmol/L) followed by a second measurement (2.7±1.8mmol/L). Repeat absolute measurement best predicted 24-hour (AUC=0.85, 95% CI: 0.84-0.86) and in-hospital death (AUC=0.77; 95% CI, 0.76-0.78). Relative clearance was the best model of lactate clearance (AUC=0.77, 95% CI: 0.75-0.78 and AUC=0.705, 95% CI: 0.69-72, respectively) (p<0.0001 for each). A sensitivity analysis using a range of initial lactate measures yielded similar results. CONCLUSIONS: The absolute value of the repeat lactate measurement had the greatest ability to predict mortality in injured patients undergoing resuscitation.

Reliable Collection of Real-Time Patient Physiologic Data from less Reliable Networks: a "Monitor of Monitors" System (MoMs).

Research and practice based on automated electronic patient monitoring and data collection systems is significantly limited by system down time. We asked whether a triple-redundant Monitor of Monitors System (MoMs) to collect and summarize key information from system-wide data sources could achieve high fault tolerance, early diagnosis of system failure, and improve data collection rates. In our Level I trauma center, patient vital signs(VS) monitors were networked to collect real time patient physiologic data streams from 94 bed units in our various resuscitation, operating, and critical care units. To minimize the impact of server collection failure, three BedMaster® VS servers were used in parallel to collect data from all bed units. To locate and diagnose system failures, we summarized critical information from high throughput datastreams in real-time in a dashboard viewer and compared the before and post MoMs phases to evaluate data collection performance as availability time, active collection rates, and gap duration, occurrence, and categories. Single-server collection rates in the 3-month period before MoMs deployment ranged from 27.8 % to 40.5 % with combined 79.1 % collection rate. Reasons for gaps included collection server failure, software instability, individual bed setting inconsistency, and monitor servicing. In the 6-month post MoMs deployment period, average collection rates were 99.9 %. A triple redundant patient data collection system with real-time diagnostic information summarization and representation improved the reliability of massive clinical data collection to nearly 100 % in a Level I trauma center. Such data collection framework may also increase the automation level of hospital-wise information aggregation for optimal allocation of health care resources.

Trends of Hemoglobin Oximetry: Do They Help Predict Blood Transfusion During Trauma Patient Resuscitation?

BACKGROUND: A noninvasive decision support tool for emergency transfusion would benefit triage and resuscitation. We tested whether 15 minutes of continuous pulse oximetry-derived hemoglobin measurements (SpHb) predict emergency blood transfusion better than conventional oximetry, vital signs, and invasive point-of-admission (POA) laboratory testing. We hypothesized that the trends in noninvasive SpHb features monitored for 15 minutes predict emergency transfusion better than pulse oximetry, shock index (SI = heart rate/systolic blood pressure), or routine POA laboratory measures. METHODS: We enrolled direct trauma patient admissions ≥18 years with prehospital SI ≥0.62, collected vital signs (continuous SpHb and conventional pulse oximetry, heart rate, and blood pressure) for 15 minutes after admission, and recorded transfusion (packed red blood cells [pRBCs]) within 1 to 3, 1 to 6, and 1 to 12 hours of admission. One blood sample was drawn during the first 15 minutes. The laboratory Hb was compared with its corresponding SpHb reading for numerical, clinical, and prediction difference. Ten prediction models for transfusion, including combinations of prehospital vital signs, SpHb, conventional oximetry, and routine POA, were selected by stepwise logistic regression. Predictions were compared via area under the receiver operating characteristic curve by the DeLong method. RESULTS: A total of 677 trauma patients were enrolled in the study. The prediction performance of the models, including POA laboratory values and SI (and the need for blood pressure), was better than those without POA values or SI. In predicting pRBC 1- to 3-hour transfusion, adding SpHb features (receiver operating characteristic curve [ROC] = 0.65; 95% confidence interval [CI], 0.53-0.77) does not improve ROC from the base model (ROC = 0.64; 95% CI, 0.52-0.76) with P = 0.48. Adding POA laboratory Hb features (ROC = 0.72; 95% CI, 0.60-0.84) also does not improve prediction performance (P = 0.18). Other POA laboratory testing predicted emergency blood use with ROC of 0.88 (95% CI, 0.81-0.96), significantly better than the use of SpHb (P = 0.00084) and laboratory Hb (P = 0.0068). CONCLUSIONS: SpHb added no benefit over conventional oximetry to predict urgent pRBC transfusion for trauma patients. Both models containing POA laboratory test features performed better at predicting pRBC use than prehospital SI, the current best noninvasive vital signs transfusion predictor.

Computer Modelling Using Prehospital Vitals Predicts Transfusion and Mortality.

OBJECTIVE: Test computer-assisted modeling techniques using prehospital vital signs of injured patients to predict emergency transfusion requirements, number of intensive care days, and mortality, compared to vital signs alone. METHODS: This single-center retrospective analysis of 17,988 trauma patients used vital signs data collected between 2006 and 2012 to predict which patients would receive transfusion, require 3 or more days of intensive care, or die. Standard transmitted prehospital vital signs (heart rate, blood pressure, shock index, and respiratory rate) were used to create a regression model (PH-VS) that was internally validated and evaluated using area under the receiver operating curve (AUROC). Transfusion records were matched with blood bank records. Documentation of death and duration of intensive care were obtained from the trauma registry. RESULTS: During the course of their hospital stay, 720 of the 17,988 patients in the study population died (4%), 2,266 (12.6%) required at least a 3-day stay in the intensive care unit (ICU), 1,171 (6.5%) required transfusions, and 210 (1.2%) received massive transfusions. The PH-VS model significantly outperformed any individual vital sign across all outcomes (average AUROC = 0.82), The PH-VS model correctly predicted that 512 of 777 (65.9%) and 580 of 931 (62.3%) patients in the study population would receive transfusions within the first 2 and 6 hours of admission, respectively. CONCLUSIONS: The predictive ability of individual vital signs to predict outcomes is significantly enhanced with the model. This could support prehospital triage by enhancing decision makers' ability to match critically injured patients with appropriate resources with minimal delays.

Performance assessment in airway management training for nonanesthesiology trainees: an analysis of 4,282 airway procedures performed at a level-1 trauma center.

BACKGROUND: Although the use of an anesthesiology "airway" rotation to train the nonanesthesiologist is commonly employed, little data exist on the utility, clinical exposure, and outcomes of these programs. METHODS: A prospectively collected observational dataset of airway procedures completed by trainees in a 4-week, anesthesiology-based, airway rotation at an academic, level-1 trauma center from July 2010 to September 2012 was reviewed. Prospectively defined data points were collected through an online data tool and included patient demographics, location, date, best laryngoscopic view, and attempt details. At the authors' institution, an attending trauma anesthesiologist is present for all intubation attempts. The primary outcome was first-attempt success. RESULTS: A total of 4,282 self-reported, airway procedures were identified. The median number of procedures performed was 50.4 ± 13.2 (range, 20 to 93; 25th quartile = 41; 75th quartile = 57). Multivariate logistic regression analysis modeling of first-attempt success rate identified two independent predictors of success: rotation week (odds ratio, 1.42; 95% CI, 1.32 to 1.61; P < 0.0001) and number of previous intubation attempts before rotation (odds ratio, 1.23; 95% CI, 1.03 to 1.46; P = 0.02. In addition, the percentage of cases with a self-reported laryngoscopic grade 1 view increased significantly from 61 to 74% (P = 0.015) from week 1 to week 4 of the rotation. CONCLUSIONS: An anesthesiology-based program for airway training of nonanesthesiologists demonstrates improved self-reported, perceived first-attempt success over the course of training with improved ability to visualize glottic structures.

Timing of intracranial hypertension following severe traumatic brain injury.

BACKGROUND: We asked whether continuous intracranial pressure (ICP) monitoring data could provide objective measures of the degree and timing of intracranial hypertension (ICH) in the first week of neurotrauma critical care and whether such data could be linked to outcome. METHODS: We enrolled adult (>17 years old) patients admitted to our Level I trauma center within 6 h of severe TBI. ICP data were automatically captured and ICP 5-minute means were grouped into 12-hour time periods from admission (hour 0) to >7 days (hour 180). Means, maximum, percent time (% time), and pressure-times-time dose (PTD, mmHg h) of ICP >20 mmHg and >30 mmHg were calculated for each time period. RESULTS: From 2008 to 2010, we enrolled 191 patients. Only 2.1% had no episodes of ICH. The timing of maximum PTD20 was relatively equally distributed across the 15 time periods. Median ICP, PTD20, %time20, and %time30 were all significantly higher in the 84-180 h time period than the 0-84 h time period. Stratified by functional outcome, those with poor functional outcome had significantly more ICH in hours 84-180. Multivariate analysis revealed that, after 84 h of monitoring, every 5% increase in PTD20 was independently associated with 21% higher odds of having a poor functional outcome (adjusted odds ratio = 1.21, 95% CI 1.02-1.42, p = 0.03). CONCLUSIONS: Although early elevations in ICP occur, ICPs are the highest later in the hospital course than previously understood, and temporal patterns of ICP elevation are associated with functional outcome. Understanding this temporal nature of secondary insults has significant implications for management.

Full of sound and fury, signifying nothing: burden of transient noncritical monitor alarms in a trauma resuscitation unit.

We examined the types of patient monitor alarms encountered in the trauma resuscitation unit of a major level 1 trauma center. Over a 1-year period, 316688 alarms were recorded for 6701 trauma patients (47 alarms/patient). Alarms were more frequent among patients with a Glasgow Coma Scale of 8 or less. Only 2.4% of all alarms were classified as "patient crisis," with the rest in the presumably less critical categories "patient advisory," "patient warning," and "system warning." Nearly half of alarms were ≤5 seconds in duration. In this patient population, a 2-second delay would reduce alarms by 25%, and a delay of 5 seconds would reduce all alarms by 49%.

Brief episodes of intracranial hypertension and cerebral hypoperfusion are associated with poor functional outcome after severe traumatic brain injury.

BACKGROUND: Management strategies after severe traumatic brain injury (TBI) target prevention and treatment of intracranial hypertension (ICH) and cerebral hypoperfusion (CH). We have previously established that continuous automated recordings of vital signs (VS) are more highly correlated with outcome than manual end-hour recordings. One potential benefit of automated vital sign data capture is the ability to detect brief episodes of ICH and CH. The purpose of this study was to establish whether a relationship exists between brief episodes of ICH and CH and outcome after severe TBI. MATERIALS: Patients at the R Adams Cowley Shock Trauma Center were prospectively enrolled over a 2-year period. Inclusion criteria were as follows: age >14 years, admission within the first 6 hours after injury, Glasgow Coma Scale score <9 on admission, and placement of a clinically indicated ICP monitor. From high-resolution automated VS data recording system, we calculated the 5-minute means of intracranial pressure (ICP), cerebral perfusion pressure (CPP), and Brain Trauma Index (BTI = CPP/ICP). Patients were stratified by mortality and 6-month Extended Glasgow Outcome Score (GOSE). RESULTS: Sixty subjects were enrolled with a mean admission Glasgow Coma Scale score of 6.4 ± 3.1, a mean Head Abbreviated Injury Severity Scale score of 4.2 ± 0.7, and a mean Marshall CT score of 2.5 ± 0.9. Significant differences in the mean number of brief episodes of CPP <50 and BTI <2 in patients with a GOSE 1-4 versus GOSE 5-8 (9.4 vs. 4.7, p = 0.02 and 9.3 vs. 4.9, p = 0.03) were found. There were significantly more mean brief episodes per day of ICP >30 (0.52 vs. 0.29, p = 0.02), CPP <50 (0.65 vs. 0.28, p < 0.001), CPP <60 (1.09 vs. 0.7, p = 0.03), BTI <2 (0.66 vs. 0.31, p = 0.002), and BTI <3 (1.1 vs. 0.64, p = 0.01) in those patients with GOSE 1-4. Number of brief episodes of CPP <50, CPP <60, BTI <2, and BTI <3 all demonstrated high predictive power for unfavorable functional outcome (area under the curve = 0.65-0.75, p < 0.05). CONCLUSIONS: This study demonstrates that the number of brief 5-minute episodes of ICH and CH is predictive of poor outcome after severe TBI. This finding has important implications for management paradigms which are currently targeted to treatment rather than prevention of ICH and CH. This study demonstrates that these brief episodes may play a significant role in outcome after severe TBI.

Comparison of massive and emergency transfusion prediction scoring systems after trauma with a new Bleeding Risk Index score applied in-flight.

BACKGROUND: Assessment of blood consumption (ABC), shock index (SI), and Revised Trauma Score (RTS) are used to estimate the need for blood transfusion and triage. We compared Bleeding Risk Index (BRI) score calculated with trauma patient noninvasive vital signs and hypothesized that prehospital BRI has better performance compared with ABC, RTS, and SI for predicting the need for emergent and massive transfusion (MT). METHODS: We analyzed 2-year in-flight data from adult trauma patients transported directly to a Level I trauma center via helicopter. The BRI scores 0 to 1 were derived from continuous features of photoplethymographic and electrocardiographic waveforms, oximetry values, blood pressure trends. The ABC, RTS, and SI were calculated using admission data. The area under the receiver operating characteristic curve (AUROC) with 95% confidence interval (CI) was calculated for predictions of critical administration threshold (CAT, ≥3 units of blood in the first hour) or MT (≥10 units of blood in the first 24 hours). DeLong's method was used to compare AUROCs for different scoring systems. p < 0.05 was considered statistically significant. RESULTS: Among 1,396 patients, age was 46.5 ± 20.1 years (SD), 67.1% were male. The MT rate was 3.2% and CAT was 7.6%, most (92.8%) were blunt injury. Mortality was 6.6%. Scene arrival to hospital time was 35.3 ± (10.5) minutes. The BRI prediction of MT with AUROC 0.92 (95% CI, 0.89-0.95) was significantly better than ABC, SI, or RTS (AUROCs = 0.80, 0.83, 0.78, respectively; 95% CIs 0.73-0.87, 0.76-0.90, 0.71-0.85, respectively). The BRI prediction of CAT had an AUROC of 0.91 (95% CI, 0.86-0.94), which was significantly better than ABC (AUROC, 077; 95% CI, 0.73-0.82) or RTS (AUROC, 0.79; 95% CI, 0.74-0.83) and better than SI (AUROC, 0.85; 95% CI, 0.80-0.89). The BRI score threshold for optimal prediction of CAT was 0.25 and for MT was 0.28. CONCLUSION: The autonomous continuous noninvasive patient vital signs-based BRI score performs better than ABC, RTS, and SI predictions of MT and CAT. Bleeding Risk Index does not require additional data entry or expert interpretation. LEVEL OF EVIDENCE: Prognostic test, level III.

Blood Transfusion Indicators Following Trauma in the Non-Massively Bleeding Patient.

BACKGROUND: Establishing transfusion guidelines during trauma resuscitation is challenging. Our objective was to evaluate indications for transfusion in trauma patients who emergently received ≤2 units of red blood cells (RBC) during the first hour of resuscitation. METHODS: A single center retrospective study included non-massively bleeding trauma patients stratified into 2 groups: 1) with a clinical indication for transfusion and 2) with no indication for transfusion. Admission vital signs (VS), injury severity score (ISS), shock index, and laboratory values were compared between the two groups using the Wilcoxon rank-sum test. RESULTS: Among 111 non-massively bleeding trauma patients, 40 presented no indication for transfusion. All patients presented similar ISS and VS. The 71 patients presenting with an indication for transfusion had higher bicarbonate (22.6 vs 20.8) and lower lactate levels (4.7 v 6.6) (p<0.05). CONCLUSION: Lactate and bicarbonate blood levels may be potential indicators for RBC transfusion need during trauma resuscitation in non-massively bleeding patients.

The role of cardiac troponin I in prognostication of patients with isolated severe traumatic brain injury.

BACKGROUND: Cardiac dysfunction is frequently observed after severe traumatic brain injury (sTBI); however, its significance is poorly understood. Our study sought to elucidate the association of cardiac troponin I (cTnI) elevation with all-cause in-hospital mortality following isolated sTBI (brain Abbreviated Injury Scale score ≥3 and admission Glasgow Coma Scale score ≤8, no Abbreviated Injury Scale score ≥3 to any other bodily regions). METHODS: We retrospectively reviewed all adult patients (aged ≥18 years) with isolated sTBI admitted to a Level I trauma center between June 2007 and January 2014. Patients must have cTnI values within 24 hours of admission. Mortality risks were examined by Cox proportional hazard model. RESULTS: Of 580 patients identified, 30.9% had detectable cTnI in 24 hours of admission. The median survival time was 4.19 days (interquartile range, 1.27-11.69). When adjusted for potential confounders, patients in the highest cTnI category (≥0.21 ng/mL) had a significantly higher risk of in-hospital mortality (hazard ratio, 1.39; 95% confidence interval, 1.04-1.88) compared with patients with undetectable cTnI. Mortality risk increased with higher troponin levels (p < 0.0001). This association was more pronounced in patients aged 65 years or younger (hazard ratio, 2.28; 95% confidence interval, 1.53-3.40; p < 0.0001) while, interestingly, insignificant in those older than 65 years (p = 0.0826). CONCLUSION: Among patients with sTBI, cTnI elevation is associated with all-cause in-hospital mortality via a nonlinear positive trend. Age modified the effect of cTnI on mortality. LEVEL OF EVIDENCE: Prognostic and epidemiologic study, level III.

Automated continuous vital signs predict use of uncrossed matched blood and massive transfusion following trauma.

BACKGROUND: Recognizing the use of uncross-matched packed red blood cells (UnXRBCs) or predicting the need for massive transfusion (MT) in injured patients with hemorrhagic shock can be challenging.A validated predictive model could accelerate decision making regarding transfusion. METHODS: Three transfusion outcomes were evaluated in adult trauma patients admitted to a Level I trauma center during a 4-year period (2009-2012): use of UnXRBC, use of greater than 4 U of packed red blood cells within 4 hours (MT1), and use of equal to or greater than 10 U of packed red blood cells within 24 hours (MT2). Vital sign (VS) features including heart rate, systolic blood pressure, and shock index (heart rate / systolic blood pressure) were calculated for 5, 10, and 15 minutes after admission. Five models were then constructed. Model 1 used preadmission VS, Model 2 used admission VS, and Models 3, 4, and 5 used continuous VS features after admission over 5, 10, and 15 minutes, respectively, to predict the use of UnXRBC, MT1, and MT2. Models were evaluated for their predictive performance via area under the receiver operating characteristic (ROC) curve, positive predictive value, and negative predictive value. RESULTS: Ten thousand six hundred thirty-six patients with more than 5 million continuous VS data points during the first 15 minutes after admission were analyzed. Model using preadmission and admission VS had similar ability to predict UnXRBC, MT1, or MT2. Compared with these two models, predictive ability was significantly improved as duration of VS monitoring increased. Continuous VS for 5 minutes had ROCs of 0.83 (confidence interval [CI], 0.83-0.84), 0.85 (CI, 0.84-0.86), and 0.86 (CI, 0.85-0.88) to predict UnXRBC, MT1, and MT2, respectively. Similarly, continuous VS for 10 minutes had a ROCs of 0.86 (CI, 0.85--0.86), 0.87 (CI, 0.86-0.88), and 0.88 (CI, 0.87-0.90) to predict UnXRBC, MT1, and MT2, respectively. Continuous VS for 15 minutes achieved the highest ROCs of 0.87 (CI, 0.87-0.88), 0.89 (CI, 0.88-0.90), and 0.91 (CI, 0.91-0.92) to predict UnXRBC, MT1, and MT2, respectively. CONCLUSION: Models using continuous VS collected after admission improve prediction for the use of UnXRBC or MT in patients with hemorrhagic shock. Decision models derived from automated continuous VS in comparison with single prehospital and admission VS identify the use of emergency blood use and can direct earlier blood product administration, potentially saving lives. LEVEL OF EVIDENCE: Therapeutic study, level III.

Permutation entropy analysis of vital signs data for outcome prediction of patients with severe traumatic brain injury.

Permutation entropy is computationally efficient, robust to outliers, and effective to measure complexity of time series. We used this technique to quantify the complexity of continuous vital signs recorded from patients with traumatic brain injury (TBI). Using permutation entropy calculated from early vital signs (initial 10-20% of patient hospital stay time), we built classifiers to predict in-hospital mortality and mobility, measured by 3-month Extended Glasgow Outcome Score (GOSE). Sixty patients with severe TBI produced a skewed dataset that we evaluated for accuracy, sensitivity and specificity. The overall prediction accuracy achieved 91.67% for mortality, and 76.67% for 3-month GOSE in testing datasets, using the leave-one-out cross validation. We also applied Receiver Operating Characteristic analysis to compare classifiers built from different learning methods. Those results support the applicability of permutation entropy in analyzing the dynamic behavior of TBI vital signs for early prediction of mortality and long-term patient outcomes.

Predicting secondary insults after severe traumatic brain injury.

BACKGROUND: Secondary insults such as hypotension, hypoxia, cerebral hypoperfusion, and intracranial hypertension are associated with poor outcome following severe traumatic brain injury (TBI). Preventing and minimizing the effect of secondary insults are essential in the management of severe TBI. At present, clinicians have no way to predict the development of these events, limiting their ability to plan appropriate timing of interventions. We hypothesized that processing continuous vital signs (VS) data using machine learning methods could predict the development of future intracranial hypertension. METHODS: Continuous VS including intracranial pressure (ICP), heart rate, systolic blood pressure, and mean arterial pressure data were collected from adult patients admitted to a single Level I trauma center requiring an ICP monitor. We tested the ability of Nearest Neighbor Regression (NNR) to predict changes in ICP by algorithmically learning from the patients' past physiology. RESULTS: Continuous VS were collected on 132 adult patients over a minimum of 3 hours per patient (5,466 hours total; 65,600 data points). Bland-Altman plots show that NNR provides good agreement in predicting actual ICP with a bias of 0.02 (±2 SD = 4 mm Hg) for the subsequent 5 minutes and -0.02 (±2 SD = 10 mm Hg) for the subsequent 2 hours. CONCLUSION: We have demonstrated that with the use of physiologic data, it is possible to predict with reasonable accuracy future ICP levels following severe TBI. NNR predicts ICP changes in clinically useful time frames. This ability to predict events may allow clinicians to make better decisions about the timing of necessary interventions, and this method could support the future development of minimally invasive ICP monitoring systems, which may lead to better overall clinical outcomes after severe TBI. LEVEL OF EVIDENCE: Prognostic study, level III.

Comparison of Decision-Assist and Clinical Judgment of Experts for Prediction of Lifesaving Interventions.

Early recognition of hemorrhage during the initial resuscitation of injured patients is associated with improved survival in both civilian and military casualties. We tested a transfusion and lifesaving intervention (LSI) prediction algorithm in comparison with clinical judgment of expert trauma care providers. We collected 15 min of pulse oximeter photopletysmograph waveforms and extracted features to predict LSIs. We compared this with clinical judgment of LSIs by individual categories of prehospital providers, nurses, and physicians and a combined judgment of all three providers using the Area Under Receiver Operating Curve (AUROC). We obtained clinical judgment of need for LSI from 405 expert clinicians in135 trauma patients. The pulse oximeter algorithm predicted transfusion within 6 h (AUROC, 0.92; P < 0.003) more accurately than either physicians or prehospital providers and as accurately as nurses (AUROC, 0.76; P = 0.07). For prediction of surgical procedures, the algorithm was as accurate as the three categories of clinicians. For prediction of fluid bolus, the diagnostic algorithm (AUROC, 0.9) was significantly more accurate than prehospital providers (AUROC, 0.62; P = 0.02) and nurses (AUROC, 0.57; P = 0.04) and as accurate as physicians (AUROC, 0.71; P = 0.06). Prediction of intubation by the algorithm (AUROC, 0.92) was as accurate as each of the three categories of clinicians. The algorithm was more accurate (P < 0.03) for blood and fluid prediction than the combined clinical judgment of all three providers but no different from the clinicians in the prediction of surgery (P = 0.7) or intubation (P = 0.8). Automated analysis of 15 min of pulse oximeter waveforms predicts the need for LSIs during initial trauma resuscitation as accurately as judgment of expert trauma clinicians. For prediction of emergency transfusion and fluid bolus, pulse oximetry features were more accurate than these experts. Such automated decision support could assist resuscitation decisions, trauma team, and operating room and blood bank preparations.

Assessing trauma care provider judgement in the prediction of need for life-saving interventions.

INTRODUCTION: Human judgement on the need for life-saving interventions (LSI) in trauma is poorly studied, especially during initial casualty management. We prospectively examined early clinical judgement and compared clinical experts' predictions of LSI to their later occurrence. PATIENTS AND METHODS: Within 10-15 min of direct trauma admission, we surveyed the predictions of pre-hospital care providers (PHP, 92% paramedics), trauma centre nurses (RN), and attending or fellow trauma physicians (MD) on the need for LSI. The actual outcomes including fluid bolus, intubation, transfusion (<1h and 1-6h), and emergent surgical interventions were observed. Cohen's kappa statistic (K) and percentage agreement were used to measure agreement among provider responses. Sensitivity, specificity, negative predictive value (NPV) and positive predictive value (PPV) were calculated to compare clinical judgement to actual patient interventions. RESULTS: Among 325 eligible trauma patient admissions, 209 clinical judgement of LSIs were obtained from all three providers. Cohen's kappa statistic for agreement between pairs of provider groups demonstrated no "disagreement" (K<0) between groups, "fair" agreement for fluid bolus (K=0.12-0.19) and blood transfusion 0-6h (K=0.22-0.39), and "moderate" (K=0.45-0.49) agreement between PHP and RN regarding intubation and surgical interventions, but no "excellent" (K ≥ 0.81) agreement between any pair of provider groups for any intervention. The percentage agreement across the different clinician groups ranged from 50% to 83%. NPV was 90-99% across providers for all interventions except fluid bolus. CONCLUSIONS: Expert clinical judgement provides a benchmark for the prediction of major LSI use in unstable trauma patients. No excellent agreement exists across providers on LSI predictions. It is possible that quality improvement measures and computer modelling-based decision-support could reduce errors of LSI commission and omission found in resuscitation at major trauma centres and enhance decision-making in austere trauma settings by less well-trained providers than those surveyed.

Shortening time to stroke treatment using ambulance telemedicine: TeleBAT.

OBJECTIVE: Patients with ischemic stroke treated with tissue plasminogen activator (rt-PA) have better outcomes when treated closer to the time of symptom onset and within the 3-hour window. We previously demonstrated the clinical use of TeleBAT, a mobile telemedicine system for stroke. We tested the impact of that system on time to treatment for patients with acute stroke. METHODS: Validity and reliability were tested by comparing neurologic examination scores obtained using our wireless system, which transmits video of a patient from a moving ambulance to desktop computers, with those obtained using the National Institute of Neurological Disorders and Stroke training videotape. TeleBAT validity and good interrater reliability were defined a priori as a kappa statistic of r > 0.5. We compared the average time to treatment for our TeleBAT-evaluated intervention group with that for our control group. The intervention group consisted of two actor patients with stroke mimicking 12 stroke scenarios and evaluated using TeleBAT. The control group consisted of patients with stroke evaluated and treated with rt-PA on arrival to the emergency department. Data were analyzed using standard t test. RESULTS: National Institutes of Health Stroke Scale items calculated by the neurologists suggest TeleBAT is valid for assessing patients with stroke remotely. Interrater reliability was high: the neurologists gleaned the same information from TeleBAT transmissions. Kappa values for both validity and reliability exceeded 0.5. The mean time to treatment for patients assessed by TeleBAT was 17 +/- 4 minutes compared with 33 +/- 17 minutes for our control group (P = .0033). CONCLUSION: TeleBAT seems to be a valid and reliable means of evaluating stroke neurologic deficits. Time to treatment was shortened using ambulance transport time to evaluate patients as candidates for thrombolytic therapy. Future studies should use a randomized design with patients with acute stroke.