Identifying Trigger Cues for Hospital Blood Transfusions Based on Ensemble Learning Methods

Background Traumatic shock is the leading cause of preventable death with most patients dying within the first 6 hours. This underscores the importance of prehospital interventions, and growing evidence suggests prehospital transfusion improves survival. Optimizing transfusion triggers in the prehospital setting is key to improving outcomes for patients in hemorrhagic shock. Our objective was to identify factors associated with early in-hospital transfusion requirements available to prehospital clinicians in the field to develop a simple algorithm for prehospital transfusion, particularly for patients with occult shock. Methods We included trauma patients transported by a single critical care transport service to a level I trauma center between 2012 and 2019. We used logistic regression, Fast and Frugal Trees (FFTs), and Bayesian analysis to identify factors associated with early in-hospital blood transfusion as a potential trigger for prehospital transfusion. Results We included 2,157 patients transported from the scene or emergency department (ED) of whom 207 (9.60%) required blood transfusion within 4 hours of admission. The mean age was 47 (IQR = 28–62) and 1,480 (68.6%) patients were male. From 13 clinically relevant factors for early hospital transfusions, four were incorporated into the FFT in following order: 1) SBP, 2) prehospital lactate concentration, 3) Shock Index, 4) AIS of chest (sensitivity = 0.81, specificity = 0.71). The chosen thresholds were similar to conventional ones. Using conventional thresholds resulted in lower model sensitivity. Consistently, prehospital lactate was among most decisive factors of hospital transfusions identified by Bayesian analysis (OR = 2.31; 95% CI 1.55–3.37). Conclusions Using an ensemble of frequentist statistics, Bayesian analysis and machine learning, we developed a simple, clinically relevant, prehospital algorithm to help identify patients requiring transfusion within 4 hours of hospital arrival.


Introduction
Hemorrhagic shock is the leading cause of preventable death among injured patients. 1Shock occurs in a continuum with progressive end-organ damage and leads to death if inadequately treated.Aggressive resuscitation according to damage control principles reduces the risk of death from hemorrhagic shock. 2 Damage control resuscitation with prehospital blood products lowers the risk of death, although the role for prehospital blood remains unclear. 3,4Early resuscitation prevents the consequences of hemorrhagic shock and poor outcomes but is di cult to achieve in the prehospital environment with constrained diagnostic and therapeutic capabilities.Current eld triage guidelines use vital signs and level of consciousness to determine the need for expedient transport to a trauma center, but these guidelines may overlook many patients with unrecognized or compensated shock who may bene t from early blood administration. 5Indications for prehospital blood transfusion after injury vary considerably and rely on arbitrary vital sign thresholds and obvious symptoms of hemorrhagic shock. 6ior work shows that elevated serum lactate levels in trauma patients may indicate sepsis and multiorgan dysfunction, increasing the chance of mortality. 7,8Prehospital clinicians can measure serum lactate levels using rapid, relatively inexpensive point of care tests to guide current triage decisions in the case of serious injury.In our previous work, we found that increased prehospital lactate levels were associated with higher odds of 24-hour hospital transfusion, even among patients without hypotension. 5ehospital lactate may be a useful prompt for prehospital transfusion.To mitigate signi cant physiologic derangement, prehospital professionals need a reliable but simple approach to rapidly and accurately identify patients who are most likely to bene t from prehospital blood.Our objective was to develop a parsimonious clinically relevant algorithm to identify patients requiring early hospital transfusion using data available in the prehospital setting.This algorithm may be a guide for prehospital blood product administration.
We hypothesized that using state of the art statistical techniques to control for known confounders, we would identify a subset of factors highly predictive of transfusion need after injury, thereby creating a simple in-eld operational model for identifying patients who need blood during trauma resuscitation.We aimed to compare the accuracy of data-driven methods with conventional triage criteria thresholds to determine variables with the optimal sensitivity and speci city for identifying trauma patients who require a blood transfusion.We also aimed to develop proof of concept decision models with components that could be adapted to different prehospital services such as rural versus urban settings.

Methods
We performed a retrospective analysis of prehospital factors that predict the need for emergent blood administration (within 4 hours) in adult (age > 16 years) trauma patients.The hours were calculated as number of minutes between ED arrival and discharge dates divided by 60.These dates are electronic timestamps.We included trauma patients with recorded venous lactate who were transported by a regional critical care transport service between 2012 and 2019.We excluded subjects with isolated traumatic brain injury (TBI) (18.6%), those that died in the emergency department (0.4%), and those with missing data (< 4%).Isolated TBI was de ned as head abbreviated injury scale (AIS > 2) and no other severe injuries (AIS face, neck, chest, spine, arms, abdomen, legs, external > 2) as these patients are not likely to require transfusion.The University Human Research Protections O ce approved this study.
The data was from a regional critical care transport service that has 18 helicopter and 2 ground bases across four states.Blood is available at all bases; 2 units of PRBCs is taken by helicopters on all missions.Crews complete 13,000 missions annually and include a minimum of a critical care nurse and paramedic.They are trained to perform point of care testing for blood gases and lactate concentration (iSTAT One, CG4+, Abbott Laboratories Princeton, NJ).They use these data to inform resuscitation and titrate mechanical ventilation.
To build an operational in-eld model to identify the need for blood use, we used an ensemble of methodologic approaches.Our rst approach was to construct Fast and Frugal Trees (FFTs) using prehospital factors associated with hospital blood administration, previously identi ed using logistic regression as in uencing hospital blood decisions (Table 1). 5Factors associated with hospital blood administration were used to nd data-driven thresholds.The algorithm that builds FFTs compares FFT receiver operating characteristics to those of other common model-building approaches: CART, logistic regression.Random Forest (RF) and Support Vector Machine (SVM) methods (see Appendix). 9 implemented FFT and Bayesian approaches as independent yet complimentary methods that validate each other's ndings.A heuristic (rule of thumb) FFT approach minimizes variance but is more prone to bias, 9 whereas a Bayesian approach is less biased and more prone to higher variance. 10Using both FFT and Bayesian approaches minimizes the overall error from both bias and variance.
FFTs are decision trees that differ from conventional decision trees in three ways: 1) they contain a minimal number of variables/cues needed to decide, 2) they make a decision after every node, and 3) they can only have two branches per node. 9These trees are salient (we know how the machine arrived at the decision), robust against over tting and good at identifying new cases of the outcome variable.This makes FFTs ideal to guide fast decisions in dynamic and dangerous environments. 9We split the analysis data set 50/50 into training and testing datasets (a common starting point for evaluating machine learning algorithms) 11 and applied the FTT algorithm.For more information about the FFT algorithm, please see the Appendix/Supplemental Methods section.
Our second approach was a Bayesian analysis of factors predicting in-hospital transfusion to con rm or supplement our prior approaches.Our goal was to identify a parsimonious model to predict transfusion within 4 hours of hospital admission.A Bayesian approach was employed for several reasons.First, prior information from our group and others may be used to provide updated knowledge about variables most strongly associated with the probability that a trauma patient requires a blood transfusion.Second, a hazard with frequentist statistics is that P values and con dence intervals may be di cult to interpret; highly signi cant P values may not be clinically meaningful or intuitively comprehensible.Third, Bayesian methods yield the probability of a speci c outcome given the data. 10inally, we synthesized the results of our approaches to create a proposed clinical algorithm of indications for prehospital blood transfusion.

Results
Of the patients transported over the 7-year study period, we identi ed 2,157 trauma patients with a prehospital lactate value (Fig. 1) obtained according to the Blood Administration protocol (Supplemental Table 1 and Appendix 2).STROBE guidelines are shown in grey rectangles.
Subjects with trauma were received in a trauma or burn unit and/or had the following mechanisms of injury: assault, animal bite, burn, electrocution (non-lightning), gunshot wound, stabbing/cutting, machinery accident; pedestrian, bicycle, motor vehicle, all-terrain vehicle, motorcycle, water transport, or aircraft accident, crash or collision.
We excluded information about prehospital blood and crystalloids given by the prehospital care service and prior to arrival from the decision process because of signi cant collinearity (i.e., relationship between model predictors) related to in-hospital blood administration.We provided the FFT algorithm with 13 variables to choose from based on clinical value and availability to the prehospital clinicians. 12Among them were AIS scores provided as a surrogate for injury condition that is visible to prehospital clinician, which we also previously found to associate with hospital transfusion.While we acknowledge the AIS value would not be available in the prehospital setting; however, we use them here as a proxy for clinically recognizable anatomic injury patterns that are used in the eld by EMS clinicians for trauma triage purposes.Five of the thirteen variables were not selected by the algorithm as they were not associated with need for blood transfusion: 1) critical high heart > 120 bpm, 2) AIS abdomen > 2, 3) AIS spine > 2, 4) injury type (blunt or penetrating), and 5) shock index (SI) range (i.e., difference between highest and lowest SI).
The algorithm generated four variables highly associated with hospital blood transfusions within 4 hours of arrival (Fig. 2).The variable chosen by the algorithm were evaluated in the following sequence: 1) minimum SBP (continuous), 2) prehospital venous lactate (continuous), 3) minimal SI (continuous), and 4) AIS chest > 2 (categorical).The predictors that were not selected by the FFT algorithm were 1) age, 2) mission type (scene or interfacility transfer), 3) AIS head > 2, and 4) AIS lower extremities > 2. The sensitivity for this FFT was 0.81 and speci city 0.71 based on data-driven variable sequence and thresholds.
A pilot FFT was obtained using training and testing datasets (the testing dataset N = 1,121) and selected from a "fan" of possible trees as having the best balance between sensitivity and speci city.A default sensitivity weight of 0.5 resulted in a "zig-zag" shape with alternating decisions.The ROC panel shows a comparison of parameters for the resulting FFT and other common model-building approaches: CART (C, red), Logistic Regression (LR, blue), Random Forest (RF, purple) and Support Vector Machine (SVM, yellow).
We applied the FFT de nitions from the pilot experiment with rounded thresholds to the entire study population and got similar performance (Supplemental Fig. 1A, sensitivity = 0.84, speci city = 0.70).Next, we maximized the sensitivity parameter with an aim to administer hospital blood to the greatest number of eligible patients while minimizing erroneous administrations.Setting the weighting parameter to any value in 0.7-1 range resulted in a "positive-rake" FFT that made positive blood decisions after every node (Supplemental Fig. 1B, sensitivity = 0.93, speci city = 0.39).Also, from Supplemental Fig. 1B The resulting FFT out-performed other model-building approaches (e.g., CART and logistic regression (LR)) by creating a decision support model for early hospital blood administration with higher sensitivity and speci city (Supplemental Fig. 1B).Finally, we altered the tree de nitions with conventional thresholds used in current eld triage guidelines and the literature to simplify for potential use in the prehospital environment. 13e FFT algorithm found variable thresholds that were different from conventional ones (Fig. 2).We explored thresholds already in common use (i.e., SBP threshold of 90 mmHg and prehospital lactate of 4 mmol/L) or based on ease of calculation for the prehospital provider (SI > 1 = HR > BP). 14 Applying conventional thresholds (Supplemental Fig. 1C) instead data-driven ones (Supplemental Fig. 1B) to the dataset greatly reduces the sensitivity but increases the speci city parameter.We tested (a) how altering the FFT de nition with conventional thresholds would in uence the sensitivity and speci city parameters (Supplemental Fig. 1C, Table 2, rst blue row) and (b) if a balance between speci city and sensitivity can be reached by using a combination of conventional and newly found thresholds (Table 2, yellow rows).The trees were created the same way as in Supplemental Fig. 1B (Table 2, rst row) differing only by the threshold values (thresholds and parameters of FFT from Supplemental Fig. 1B are highlighted orange in Table 2).Table 2 illustrates how varying the threshold for SBP, lactate, and shock index alters the sensitivity, speci city, and overall performance based on Youden's As expected, using a higher SBP, lower lactate, or lower SI threshold increases sensitivity but decreases speci city.Youden's J statistic = sensitivity + speci city -1 summarizes the performance of each model.
Tree from row 1.1 is depicted in Supplemental Figure 1B.
Tree from row 1.8 is depicted in Supplemental Figure 1C.We performed sensitivity analyses by removing the lactate term from the models and using FFT-derived vs. conventional thresholds for SBP and SI (Table 2, rows 2.1-2.4),recognizing that prehospital lactate may not be widely available.The sensitivity was often higher for the models containing the lactate term (compare rows 2.1 and 1.1/3, 2.2 and 1.2/4, 2.3 and 1.5/7, 2.4 and 1.6/8), but the speci city and Youden's J index were lower.
We also assessed current practice of prehospital blood transfusion by the critical care service and the need for early in-hospital transfusion.Table 3 shows a cross-tabulation of actual prehospital blood administration by early hospital transfusions.Of 207 subjects who required early hospital transfusions, 79 (38.2%) subjects also received blood before arriving to the hospital (Table 3, upper left quadrant).The majority (73) of these 79 subjects had SBP < 90 mmHg and received prehospital blood according to the prehospital care service protocol for blood transfusions.Among 60 patients who received prehospital transfusions but did not require hospital blood (Table 3, upper right quadrant), 33 (55.0%) patients had SBP < 90 mmHg.Patients who received blood with systolic blood pressures > 90 mmHg, either received the product on the order of the physician or in deviation from the protocol.Synthesizing and operationalizing the results from our approaches for potential eld use, we developed an algorithm for prehospital blood transfusion that incorporates prehospital SBP, prehospital lactate, shock index, and severe abdominal injuries (Fig. 3).This algorithm allows for different threshold values that may be tailored according to system resources and time considerations.
We also applied the FFT de nitions from Supplemental Fig. 1B but excluding the node for severe chest injuries (Supplemental Fig. 2).The resulting sensitivity and speci city parameters were slightly lower than those of the four-factor model (Supplemental Fig. 1B, Supplemental Fig. 2).

Discussion
Using advanced statistical methods to control for confounders and to maximize the information provided by a large cohort of adult trauma patients with granular prehospital data, we identi ed four variables that predict early in-hospital transfusions.These variables, which are accessible by prehospital clinicians, were selected by an FFT algorithm to facilitate the decision to administer prehospital blood quickly with a parsimonious (small) set of data.We con rmed these ndings using Bayesian analysis to identify strong predictors of early in-hospital transfusion.Prehospital lactate emerged as a strong predictor for transfusion need from both the FFT and Bayesian approaches among patients who were not hypotensive.This is consistent with recent study by Griggs et al. who also predicted in hospital transfusion using prehospital lactate concentration. 15ministration of prehospital blood products to patients in hemorrhagic shock reduces mortality. 4A systematic review and meta-analysis by Rijnhout et al. describes the administration of prehospital blood products as feasible and safe but describes the evidence as low quality and di cult to compare because there are no standard indication for transfusion. 16While tools have been developed to identify patient at risk of Trauma Associated Severe Hemorrhage (TASH) and for massive transfusion (ABC score), they rely on data not readily available in the prehospital environment (hemoglobin and ultrasound) and neither was developed for the prehospital environment. 17,18 nd the simplest decision model to identify people who need prehospital blood transfusions, we are faced with two competing considerations: 1) correctly identifying the greatest number of people who need blood (i.e., maximizing the sensitivity of the model), and 2) conserving limited resources of blood.Using these considerations, an EMS Medical Director may conclude that the model with a lactate concentration threshold of 2.5 mmol/L (Sensitivity = 0.89, Speci city = 0.48, Table 2 row 1.6) is more appropriate for use in a rural setting with delayed access to a trauma center and subsequent damage control resuscitation, while a model with a 4 mmol/L threshold (Sensitivity = 0.82, Speci city = 0.62, Table 2 row 1.8) could be more suitable for urban settings with short prehospital times.Similar trade-offs can be made with the thresholds for SBP and shock index.
We adjusted the model thresholds to create simple rules for quick reference in the eld (Table 2).The results depicted in  14,22 This threshold is more conservative than the one found by the FFT algorithm (2.5 mmol/L).
The last cue identi ed to trigger potential transfusion is severe chest injury.In the data we used AIS > 2; however, recognizing this is not available in the eld setting as an objective number, this would rely on clinical exam evidence, much in the way the anatomic triage criteria for the national eld triage guidelines are identi ed.We suggest operationalizing this as ail chest, unstable chest fractures, or need for needle decompression (Fig. 3).Local medical directors certainly would have discretion to operationalize this cue in an alternative way given the personnel, resources, and trauma population seen by his or her EMS agency.We do show comparable accuracy if this cue is omitted, allowing further adaptation to local circumstances given it is the most subjective cue in operational form.
A key limitation of our study is that decision to transfuse blood is not always synonymous with the need to transfuse blood.Also, our analyses are retrospective and derived from a single EMS agency serving a regional trauma system.The dataset was limited to patients who had lactate sampled which imparts bias among patients with hemorrhagic shock.Selection bias may result when treatment priorities preclude sampling of lactate in the sickest patients.EMS data is rarely inputted into the record contemporaneously with care and is subject to recall and reporting bias.We import data electronically (vital signs, times and point of care labs) into the prehospital health record which mitigates these biases.
There is likely a selection and sensitivity bias as our critical care organization is called for patients with more severe injuries or those who are geographically distant from trauma care.

Conclusion
We developed a parsimonious, clinically relevant algorithm to identify patients who may require prehospital transfusion.This algorithm accounts for prehospital lactate concentration which is useful for identifying patients with occult shock not meeting the conventional threshold for hypotension.Thresholds of decision factors should be adjusted to meet the needs and resources of a given prehospital trauma system.Further work is necessary to externally validate this algorithm for prehospital blood transfusion.
We are including the Appendix describing the FFT algorithm, Blood Administration protocol, and the study checklist for adhering to STROBE guidelines as Supplemental Digital Content.SBP_min -minimal SBP (mmHg), nlacven -prehospital lactate concentration (mmol/L), si_minminimum SI (bpm/mmHg), ais_ab -AIS for abdomen (0/1, equal to 1 if the AIS is greater than 2).

Figure 1 Flow diagram illustrating cohort selection Figure 2 Pilot
Figure 1

Table 1 .
Cohort characteristics* n (%) shown for categorical variables, median (IQR) shown for continuous variables within 24 hours of hospital admission + the rest of population was transported from scene • the rest of population had blunt injuries ♦ means for categorical variables were compared using Fisher's exact test, for continuous variables - The median prehospital lactate concentration was 4.85 mmol/L for the subjects who received blood products (IQR = 2.30-5.80),and 2.48 mmol/L for the subjects who did not require hospital blood products within 4 hours of arrival (IQR = 1.30-2.98).Of the subjects who received hospital blood products, 19 (10%) died within 24 hours of admission.Only 1% of the subjects who did not require hospital blood died within 24 hours of admission (n = 15).Consistently, a greater percentage of subjects who received hospital blood products needed other hospital life-saving interventions (LSIs) (Table

Table 2
. Effect of using deduced, conventional, or mixed thresholds on FFT parameters # -indicates the FFT model number; models with number '2.1' or higher did not include lactate as a variable; Sens. -sensitivity, Spec.-speci city

Table 3
Cross-tabulation of prehospital transfusions by 24-hour hospital transfusions 4h hospital ED blood = YES 4h hospital ED blood = NO Table 2 have broad implications for prehospital clinicians, ranging from urban and rural EMS systems to austere military environments that might require prolonged eld care.Using the four variables derived from our models, prehospital system leadership can decide what thresholds are appropriate for transfusions in their respective systems, based on existing resources and trauma center access.
21,20ous studies associate prehospital lactate with mortality and morbidity in trauma patients.19,20Subsequentworkdemonstratedthe association between lactate and need for life saving interventions.21Recentwork by Fukuma et al. and Galvagno et al. established that prehospital lactate threshold of > 4 mmol/L is associated with the need life-saving interventions for hemorrhage control.