TiME OUT: Time-specific Machine-learning Evaluation to Optimize Ultra-massive Transfusion

Courtney H Meyer; Jonathan Nguyen; Andrew ElHabr; Nethra Venkatayogi; Tyler Steed; Judy Gichoya; Jason D Sciarretta; James Sikora; Christopher Dente; John Lyons; Craig M Coopersmith; Crystal Nguyen; Randi N Smith

doi:10.1097/TA.0000000000004187

. Author manuscript; available in PMC: 2025 Mar 1.

Published in final edited form as: J Trauma Acute Care Surg. 2023 Nov 13;96(3):443–454. doi: 10.1097/TA.0000000000004187

TiME OUT: Time-specific Machine-learning Evaluation to Optimize Ultra-massive Transfusion

Courtney H Meyer ^1,^2,³, Jonathan Nguyen ^1,⁴, Andrew ElHabr ⁵, Nethra Venkatayogi ⁶, Tyler Steed ^1,², Judy Gichoya ^1,², Jason D Sciarretta ^1,², James Sikora ^1,², Christopher Dente ^1,², John Lyons ^2,⁷, Craig M Coopersmith ^2,⁷, Crystal Nguyen ¹, Randi N Smith ^1,^2,³

PMCID: PMC10922246 NIHMSID: NIHMS1941027 PMID: 37962139

Abstract

Background:

Ultra-massive 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 life-saving 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 utilize time-specific machine learning (ML) 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 24h of admission. Cross-sectional data was obtained from the blood bank and trauma registries and time-specific data was obtained from the electronic medical record. Time-specific decision-tree models (TS-DTM) predicating mortality were generated and evaluated using AUC.

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 to 55.5 in the survivors (p<0.001) and significantly greater rates of pRBC and FFP at each time interval. TS-DTM 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.

Conclusions:

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. ML is an effective tool to investigate this concept and further research is needed to refine and validate these TS-DTM.

Level of Evidence:

Keywords: Ultramassive transfusion, massive transfusion, resuscitation

Media Summary:

Machine learning models can successfully predict mortality in ultra-massive transfusion and identify how parameters associated with survivability change over time. Such models have the potential to serve as evidence-based tool to guide appropriate resource allocation in UMT resuscitation.

#UltramassiveTransfusion #MassiveTransfusion #SoMe4Trauma

Introduction

Hemorrhage is the leading cause of preventable death after trauma [1-3], and component-based, damage control resuscitation (DCR) has become the gold standard of care [4-10]. Empiric transfusion of blood products for trauma patients with hemorrhagic shock has been shown to improve morbidity and mortality, mitigate the incidence of traumatic coagulopathy, acidosis and hypothermia and significantly decrease rates of exsanguination within the first 24h [4-12].

With DCR as the leading resuscitation strategy for trauma patients in hemorrhagic shock has come the emergence of principles like massive transfusion (MT), massive transfusion protocols (MTP) and ultra-massive transfusion (UMT). UMT, specifically, is defined as transfusion of 20 or more units of red blood cell product within 24h of admission [13-17]. Despite the widespread use of UMT as an intervention for critical trauma patients, mortality rates remain as high as 40-80% [13-23]. Recognizing the limited availability of blood, and the concerns surrounding ongoing transfusion in futile cases, studies have sought to investigate the concept of transfusion ceilings, or cut-off point for high volume transfusion [24, 25].

In 2022, Quintana et al.found a mortality plateau after 40.5 units while Morris et al in 2020 suggested futility after 80 units [20, 22]. Conversely, a military focused study in 2023 by Gurney et al., found a 24hr mortality rate of only 20.7% for patients who received over 100 units of blood [19]. While this broad range of findings demonstrates the absence of a consensus threshold in our current literature, evidence does overarchingly support that increased blood product transfusion is not associated with increased survival, and that the odds of mortality increase with increased products transfused [14-16, 18-23].

Other studies have sought a more granular approach, and to identify specific factors associated with improved outcomes. For example, a 2021 multicenter study by Matthay et al. found that GCS > 3, age < 50, the absence of an emergency department (ED) thoracotomy, and normal platelet levels on admission were factors associated with improved survival[15]. Gallastegi et al. in 2022 highlighted that transfusion rates over time were able to better discriminate survival probability than isolated transfusion volumes [14].

While such work has provided important contributions to the field, we have yet to identify consensus guidelines to direct utilization of this resource-demanding intervention and thereby triage appropriate allocation of blood product. Furthermore, while we recognize resuscitation as an inherently dynamic process, the majority of these studies have been based on cross-sectional data and little consideration has been given to how the component of time may impact these decisions. We hypothesize that the clinical and physiologic parameters associated with survivability during trauma-related UMT do change over time and should be considered in guiding resuscitation.

Machine learning (ML) predictive modeling, a subtype of artificial intelligence (AI), is defined as the development of computer systems that are able to learn and adapt without following explicit instructions, by using algorithms and statistical models to analyze and draw inferences from patterns in data [26, 27]. ML is an emerging tool in clinical research and one that may provide a novel approach to this complex problem. Given the algorithmic nature of trauma and ML’s affinity for pattern-recognition, recent review articles have highlighted the potential benefit of using ML to predict outcomes in trauma [28, 29]. Therefore, we sought to utilize time-specific ML models to identify if parameters associated with survivability change over time and to predict mortality during the active resuscitation period in trauma patients undergoing UMT.

Methods

Study Design & Inclusion Criteria

A retrospective cohort review was conducted at a large, academic, American College of Surgeons (ACS) -verified Level I trauma center from May 2018-November 2021. The dataset was generated from the institutional blood bank registry, trauma registry and electronic medical record (EMR).

First, the blood bank registry was queried for all patients undergoing massive transfusion protocol (MTP) activations during the study period. These patients were then cross-referenced with the institutional trauma registry to determine if the etiology of hemorrhagic shock was attributable to a traumatic mechanism of injury. The EMR blood product transfusion record was then used to delineate which patients met criteria for UMT (defined as 20 or greater total combined units of packed red blood cells and whole blood) within the first 24 hours of admission.

Pregnant patients were excluded. Patients who met criteria for UMT for indications other than trauma including burns, gastrointestinal bleeds (GI bleeds), primary cardiothoracic (CT), vascular or transplant surgery, obstetrics and gynecology (OBGYN) related complications were also excluded.

For patients meeting inclusion criteria, data pertaining to demographics, clinical presentation, injury classification, hospital course and outcomes was obtained from the institutional trauma registry. For the same cohort of patients, the EMR was queried to obtain time-specific data for all recorded vital signs, laboratory values and blood product transfusions throughout their hospitalization. This data was then coalesced into the working dataset for the study.

Variables of Interest

The exposure of interest was UMT and the primary outcome of interest was in-hospital mortality at 48 hours and discharge. Secondary outcomes of interest include differences in clinical presentation, hemodynamics, lab values, total blood products transfused, transfusion ratios, rates of transfusion and operative intervention between surviving and deceased patients undergoing UMT.

The blood products analyzed in this study included whole blood (WB), packed red blood cells (pRBC), fresh frozen plasma (FFP), platelets and cryoprecipitate (cryo). The corresponding standard volumes for each unit of product at the study institution were 500ml, 300mL, 250mL, 350mL and 10mL, respectively. Transfusion ratios were analyzed as rates of pRBC to FFP.

Injuries were classified as blunt or penetrating and then the mechanisms of injury (MOI) were grouped into the following categories: gunshot wounds (GSWs), motor vehicle collisions (MVCs), motorcycle collisions (MCC), pedestrian vs auto (peds vs auto), other blunt, and other penetrating. Operative intervention was classified as those who went directly to the operating room (OR) from the ED. Details of operative procedures were determined by review of attending surgeon operative notes. A resuscitative thoracotomy was classified as one that occurred in the emergency department (ED) and these patients were excluded from the analysis of those undergoing thoracotomy in the OR.

Machine Learning Modeling

A series of time-specific decision tree models were then generated with the primary outcome of predicting mortality at 48 hours, based on data from the initial 24 hours following the start of transfusion. The secondary outcome of interest was identifying how the clinical and physiologic parameters associated with survivability changed during the given time intervals between deceased (48hr mortality) and survivors. The following variables were included in the models; patient demographics, injury mechanism and severity, vital signs, laboratory values, total blood products and total component blood products transfused, TXA administration and ED thoracotomy. All data points were medians of the variable for the given time interval except for the transfusion data, which was based on the cumulative median at the given time interval. For continuous variables, the change from baseline at time 0 to that given time interval was also included. The time intervals analyzed included 0 hours, 0-2h, 2-4h, 4-8h, 8-12h, 12-16h, 16-20h and 20-24h hours.

Statistical Analysis

Data was analyzed using R version 4.1.1 (R Foundation for Statistical Computing). Significance was set at α=0.05. Comparisons were made between those who survived their UMT to discharge and those who did not. Independent t test was used for normally distributed data, Pearson’s Chi-squared test for categorical data, and Mann-Whitney U and Fisher’s exact test for non-parametric data.

An hour-by-hour time series analysis was conducted to compare vital signs, laboratory values and transfusion rates between overall survivors and deceased during the first 48h after the initiation of transfusion. The median value for the following variables was calculated at each time interval; heart rate, systolic blood pressure, diastolic blood pressure, hemoglobin, platelet count, lactate level, pH, base deficit and transfusion rates of total product, pRBC, FFP, platelets, cryoprecipitate. The time intervals analyzed included 0 hours, 0-2h, 2-4h, 4-8h, 8-12h, 12-16h, 16-20h and 20-24h.

A subsequent analysis was generated comparing those that survived the first 48h versus those that do not. The models were created based on these cohorts, and generated in R version 4.1.1 (R Foundation for Statistical Computing). Nested cross validation was used to validate the models and AUCs were generated based on the average models performance. IRB approval was obtained for this study (IRB 00115688) and the STROBE checklist for observational studies was utilized.

Results

Demographics

During the study period, there were 20,508 trauma activations. Of which, 1,164 had MTP activations from the blood bank and 196 (16.8%) of which, were trauma patients meeting criteria for UMT. 16 of these patients were excluded due to insufficient time-specific data points, yielding a total of 180 patients included in the study. The 24-hour mortality rate was 35.6% (n= 64), 48-hour mortality rate was 40.5% (n = 73) and the overall mortality rate was 52.2% (n=94). The patients were predominantly black (81.7%, n = 147) males (77.8%, n = 140) with a median age of 28.5 [24-40.5] years. There were no significant demographic differences between the deceased and survivors [Table 1].

Table 1:

Demographics, injury characteristics, clinical presentation, operative intervention and transfusion data for trauma patients undergoing UMT stratified by mortality

Modeling by survival duration	Overall	Overall Survival			48-hour Survival
Group	n = 180	Deceased n = 94	Survived n = 86	¹p-value	Deceased n = 73	Survived n = 107	¹p-value
Age	28.5 [24, 40.5]	28 [24, 41]	29 [23, 40]	0.683	29 [25, 41]	28 [24, 40]	0.678
Percent male	140 (77.8%)	72 (76.6%)	68 (79.1%)	0.826	58 (79.5%)	82 (76.6%)	0.792
Race				0.858			0.780
American Indian	1 (0.6%)	1 (1.1%)	0 (0.0%)		1 (1.4%)	0
Asian	3 (1.7%)	1 (1.1%)	2 (2.3%)		1 (1.4%)	2 (1.9%)
Black	147 (81.7%)	75 (79.8%)	72 (83.7%)		57 (78.1%)	90 (84.1%)
Native Hawaiian or PI	1 (0.6%)	1 (1.1%)	0 (0.0%)		1 (1.4%)	0
Other	2 (1.1%)	1 (1.1%)	1 (1.2%)		1 (1.4%)	1 (0.9%)
Unknown	5 (2.8%)	2 (2.1%)	3 (3.5%)		2 (2.7%)	3 (2.8%)
White	21 (11.7%)	13 (13.8%)	8 (9.3%)		10 (13.7%)	11 (10.3%)
Ethnicity				0.042^*			0.283
Hispanic/Latino	2 (1.1%)	0 (0.0%)	2 (2.3%)		0	2 (1.9%)
Non- Hispanic/Latino	174 (96.7%)	90 (95.7%)	84 (97.7%)		70 (95.9%)	104 (97.2%)
Other	4 (2.2%)	4 (4.3%)	0 (0.0%)		3 (4.1%)	1 (0.9%)
Percent penetrating	99 (55%)	46 (48.9%)	53 (61.6%)	0.119	39 (53.4%)	60 (56.1%)	0.843
Mechanism of Injury				0.374			0.878
Fall	4 (2.2%)	1 (1.1%)	3 (3.5%)		1 (1.4%)	3 (2.8%)
GSW	97 (53.9%)	45 (47.9%)	52 (60.5%)		38 (52.1%)	59 (55.1%)
MCC	15 (8.3%)	9 (9.6%)	6 (7%)		7 (9.6%)	8 (7.5%)
MVC	41 (22.8%)	26 (27.7%)	15 (17.4%)		19 (26%)	22 (20.6%)
Peds vs auto	21 (11.7%)	12 (12.8%)	9 (10.5%)		7 (9.6%)	14 (13.1%)
Stabbing	2 (1.1%)	1 (1.1%)	1 (1.2%)		1 (1.4%)	1 (0.9%)
Injury Severity Score	36 [26, 50]	40 [29, 50]	34 [25, 50]	0.166	36 [26, 50]	38 [25, 50]	0.953
Initial Systolic Blood Pressure	90 [68, 120]	80 [58, 116]	90 [78, 121]	0.001^*	80 [58, 114]	90 [76, 121]	0.004^*
Initial Diastolic Blood Pressure	57 [0, 83]	42 [0, 83]	60 [40, 82]	0.049^*	40 [0, 77]	60 [32, 85]	0.047^*
Initial Heart Rate	115 [ 82, 137]	105 [53, 133]	125 [106, 144]	< 0.001^*	89 [40, 125]	125 [106, 146]	< 0.001^*
Initial Respiratory Rate	22 [16, 30]	20 [0, 32]	24 [19, 30]	0.001^*	20 [0, 29]	24 [18, 30]	0.003^*
Initial Glasgow Coma Score	7.0 [3, 14]	3 [3, 10]	14 [6, 15]	< 0.001^*	3 [3, 7]	13 [4, 15]	< 0.001^*
Mode of Arrival				0.743			0.438
Ground Ambulance	158 (87.8%)	81 (86.2%)	77 (89.5%)		63 (86.3%)	95 (88.8%)
Helicopter	20 (11.1%)	12 (12.8%)	8 (9.3%)		10 (13.7%)	10 (9.3%)
Private vehicle	2 (1.1%)	1 (1.1%)	1 (1.2%)		0	2 (1.9%)
Time in ED (Minutes)	25.5 [15, 45]	23 [16, 40]	31 [14, 60]	0.088	21 [15, 34]	32 [15.5, 61]	0.014^*
ED CPR	17 (9.4%)	15 (16%)	2 (2.3%)	0.002^*	13 (17.8%)	4 (3.7%)	0.003^*
ED Thoracotomy	40 (22.2%)	36 (38.3%)	4 (4.7%)	< 0.001^*	31 (42.5%)	9 (8.4%)	< 0.001^*
Operative intervention?
Yes	177 (98.3%)	92 (97.8%)	85 (98.8%)	0.889	72 (985%)	105 (98.1%)	0.987
Thoracotomy**	49 (27.2%)	29 (30.9%)	20 (23.3%)	0.329	26 (35.6%)	23 (21.5%)	0.055
Clamshell	24 (13.3%)	13 (13.8%)	11 (12.8%)	0.911	10 (13.7%)	14 (13.1%)	1.000
Ex-lap	157 (87.2%)	82 (87.2%)	75 (87.2%)	1.000	63 (86.3%)	94 (87.9%)	0.938
Solid Organ Injury	105 (58.3%)	53 (56.4%)	52 (60.5%)	0.687	42 (57.5%)	63 (58.9%)	0.980
Vascular Injury	144 (80%)	77 (81.9%)	67 (77.9%)	0.628	59 (80.8%)	85 (79.4%)	0.970
Return to OR w/in 48hrs	85 (47.2%)	46 (48.9%)	39 (45.3%)	0.740	34 (46.6%)	51 (47.7%)	1.000
Whole Blood (Total)	0 [0, 2.5]	0 [0, 10]	0 [0, 10]	0.110	0 [0, 0]	0 [0, 4]	0.185
pRBC (Total)	31 [22.5, 46]	36.5 [26, 55]	26.5 [18, 36]	0.001^*	33 [26, 48]	23 [16, 32]	< 0.001^*
FFP (Total)	23 [16, 38]	33 [18, 47]	19 [15, 27]	< 0.001^*	29 [18, 42]	18 [12, 28]	< 0.001^*
Platelets (Total)	2.5 [2, 4]	2 [1, 5]	3 [2, 4]	0.155	2 [1, 3]	2 [2, 3]	< 0.757^*
Cryoprecipitate (Total)	0.5 [0, 2]	0 [0, 2]	1 [0, 2]	0.368	0 [0, 2]	1 [0, 2]	0.808
Total Blood Products	59 [41.5, 92]	71.5 [47, 111]	55.5 [39, 66]	< 0.001^*	66 [47, 98]	44 [34, 65]	< 0.001^*
TXA	157 (87.2%)	83 (88.3%)	74 (86%)	0.819	64 (87.7%)	93 (86.9%)	1.000
pRBC/FFP Ratio (Total)	1.3 [1.1, 1.6]	1.2 [1.0, 1.5]	1.4 [1.2, 1.6]	0.842	1.2 [1.0, 1.5]	1.2 [1.1, 1.4]	0.468
pRBC/FFP Ratio ≥1.5:1 (Total)	31.1%	29.8%	33.7%	0.574	31.8%	14.9%	0.013^*

Open in a new tab

Exploratory Laparotomy (Ex-Lap), Gunshot wound (GSW), Motorcycle collision (MCC), Motorcycle collision (MVC), Pedestrian (peds), Pacific Islander (PI) Emergency department (ED), Cardiopulmonary resuscitation (CPR), Operating room (OR), packed red blood cells (pRBC), fresh frozen plasma (FFP), tranexamic acid (TXA). *Excludes those who had an ED thoracotomy. Continuous variables reported as median with [Q1, Q3

T-test for continuous variables, and chi-squared for categorical variables unless the data were too sparse, in which case Fisher's exact test

Injury Patterns and Clinical Presentation

The survivors presented with a greater proportion of penetrating injuries (61.6%) compared to 48.9% in the deceased. The primary mechanism of injury for the cohort overall was GSWs (53.9%) followed by MVCs (22.8%) and peds vs. auto (11.7%) with no significant differences between the survivors and deceased. Both groups were severely injured with a median ISS of 36 [26-50] [Table 1].

The deceased group presented with a significantly lower systolic (80 vs 90, p = 0.001) and diastolic blood pressure (42 vs 60, p = 0.049), lower heart rate (105 vs 125, p < 0.001) and lower GCS (3 vs 14, p < 0.001). They were also more likely to undergo ED CPR (16.0% vs 2.3%, p = 0.002) or an ED thoracotomy (38.3% vs 4.7%, p < 0.001) [Table 1].

Blood Product Transfusion

In the 24h following the initiation of transfusion, the deceased cohort received significantly more total blood products with a median of 71.5 units compared to 55.5 total units in the survivors (p < 0.001). Overall, the deceased received significantly greater units of pRBC (36.5 vs 26.5, p = 0.001) and FFP (33 vs 19, p < 0.001) and similar rates of whole blood (0 vs 0, p = 0.110), platelets (2 vs 3, p = 0.757) and cryoprecipitate (0 vs 1, p = 0.368). The median pRBC/FFP ratio for the initial 24h following the initiation of transfusion was 1.3 with no significant differences between the survivors and deceased overall. However, over 31.8% of those died in the first 48 hours had a pRBC:FFP ratio greater than 1.5:1 compared to 14.9% in the survivors (p = 0.013). Just under 90% of both cohorts received TXA [Table 1].

Operative Intervention

Nearly the entire cohort (98.3%) went emergently to the operating room from the ED. There were not significant differences between the survivors and deceased in terms of the operative interventions they underwent or intra-operative findings. Over three quarters of the cohort overall underwent an exploratory laparotomy while just over one quarter underwent a thoracotomy. Over 50% had solid organ injuries and over 80% had a vascular injury. Nearly half of both cohorts returned to the operating room within 48 hours of their sentinel procedure [Table 1].

Vital Signs Time-Series Analysis

During the initial 0-2hh and 2-4h time intervals following the initiation of UMT, the median heart rate and systolic blood between the two groups was not significantly different. However, from 4 hours and beyond, the deceased cohort has a significantly higher rate and significantly lower systolic blood pressure than the survivors [Table 2].

Table 2:

Time-specific vital signs and laboratory values during first 24 hours following initiation of transfusion for trauma patients undergoing UMT stratified by mortality

Model		Overall	Overall Survival			48-hour Survival
Time Interval	Variable		Deceased	Survivors	p-value	Deceased	Survivor	p-value
0-2 hours		n = 180	n = 94	n = 86		n = 72	n = 107
	Heart Rate	120 [98, 137]	119 [89, 137]	123 [104, 136]	0.177	114 [86, 134]	125 [106, 138]	0.018*
	SBP	94 [84, 109]	96 [82, 108]	93 [84, 110]	0.964	96 [85, 112]	93 [84, 107]	0.455
	Lactate	10.8 [7.5, 14.5]	13.7 [8.2, 15.7]	10.2 [5.2, 12.4]	0.126	14 [13.3, 15.7]	10 [6.8, 12.7]	0.461
	Base Deficit	−16.8 [−21.5, −11.4]	−20 [−25.0, −13.0]	−16 [−19.0, −10.8]	0.215	−20 [−25, −13]	−16 [−19, −10.8]	0.215
	pH	7.0 [6.9, 7.2]	6.9 [6.8, 7.1]	7.1 [7.0, 7.2]	0.077	6.9 [6.8, 7.1]	7.1 [7, 7.2]	0.077
	Hemoglobin	12 [10.7, 13.2]	11.7 [10.4, 12.6]	12.1 [11.1, 13.5]	0.208	11.6, [10.1, 12.4]	12.2 [11.1, 13.3]	0.027*
	Platelets	206 [151, 239]	187 [134, 245]	206 [173, 227]	0.810	151 [120, 209]	217 [175, 247]	0.001*
2-4 hours		n = 180	n = 92	n = 86		n = 71	n = 106
	Heart Rate	105 [96, 127]	113 [103, 128]	104 [83,127]	0.143	113 [104, 128]	103 [84, 127]	0.284
	SBP	111 [104, 121]	108 [98, 116]	115 [110, 121]	0.243	113 [104, 120]	111 [105, 121]	0.677
	Lactate	10 [7.2, 16.2]	12.6 [9.5, 25.0]	7.5 [5.3, 10.0]	0.063	18.8 [12.5, 25]	9.8 [6.5, 14.4]	0.415
	Base Deficit	−11.8 [−17.0, −5.0]	−14.3 [−17.0, −8.4]	−7.0 [−14.0, −3.0]	0.033*	−15.5 [−18.2, −10.2]	−8 [−13.5, −4.0]	0.006*
	pH	7.2 [7.0, 7.3]	7.1 [7.0, 7.2]	7.2 [7.2, 7.3]	0.012*	7.1 [7.0, 7.1]	7.2 [7.2, 7.3]	0.004*
	Hemoglobin	10.7 [9.7, 12.6]	10.5 [9.6, 12.2]	11.1 [10.2, 13.1]	0.941	10.4 [8.7, 12.2]	10.9 [10.3, 12.6	0.848
	Platelets	149 [82, 251]	86 [59, 148]	199 [141, 312]	0.003*	78 [59, 117]	179 [137, 280]	0.013*
4-8 hours		n = 180	n = 83	n = 86		n = 62	n = 107
	Heart Rate	113 [91, 129]	117 [96, 130]	110 [85, 127]	0.038*	117 [101, 130]	111 [87, 128]	0.032*
	SBP	117 [102, 136]	107 [93, 127]	126 [110, 141]	< 0.001*	102 [91, 117]	127 [109, 140]	< 0.001*
	Lactate	8.0 [5.0, 15.1]	14.6 [8.5, 16.5]	5.2 [3.8, 7.7]	< 0.001*	15.2 [9.4, 17.1]	5.4 [4, 10]	< 0.001*
	Base Deficit	−7.2 [−11.4, −1.8]	−9.8 [−15, −5.8]	−3 [−8.4, 0.6]	< 0.001*	−13 [--16.8, −8.3]	−3.5 [−8.5, −0.6]	< 0.001*
	pH	7.2 [7.1, 7.4]	7.2 [7.0, 7.2]	7.2 [7.1, 7.4]	< 0.001*	7.1 [7.0, 7.2]	7.3 [7.2, 7.4]	< 0.001*
	Hemoglobin	10.6 [9.5, 12.2]	9.7 [8.6, 10.9]	11.7 [10.2, 13.1]	0.011*	9.5 [8.2, 10.5]	11.6 [10.2, 12.9]	0.032*
	Platelets	65 [42, 86]	59 [36, 80]	73 [50, 88]	0.073	62 [34, 77]	67 [48, 88]	0.076*
8-12 hours		n = 180	n = 56	n = 86		n = 35	n = 107
	Heart Rate	111[99, 126]	118 [100,131]	110 [99, 121]	0.099	123 [112, 132]	109 [97, 122]	0.080*
	SBP	117 [102, 131]	106 [98, 126]	121 [107, 132]	0.009	102 [85, 117]	121 [105, 133]	0.001*
	Lactate	7.2 [4.0, 12.1]	12.2 [9.1, 14.7]	5.1 [3.2, 8.1]	< 0.001*	13.2 [9.8, 16.9]	5.6 [3.4, 9.4]	< 0.001*
	Base Deficit	−3.8 [−8.9, −0.6]	−9.4 [−14, −3.7]	−1.9 [−4.9, −1.2]	< 0.001*	−11.2 [−15, −6]	−2.2 [−4.9, −0.1]	< 0.001*
	pH	7.3 [7.2, 7.4]	7.1 [7.1, 7.3]	7.3 [7.3, 7.4]	< 0.001*	7.1 [7.1, 7.3]	7.3 [7.3, 7.4]	< 0.001*
	Hemoglobin	10.9 [8.9, 12.4]	10.1 [8.2, 11.5]	11.1 [10, 12.8]	0.013*	9.3 [8.2, 11.2]	11.1 [9.7, 12.8]	0.007*
	Platelets	61 [42, 79]	53 [30, 71]	66 [51, 81]	0.027*	42 [27, 59]	67 [51, 83]	< 0.001*
12-16 hours		n = 180	n = 41	n = 86		n = 20	n = 107
	Heart Rate	112 [98, 125]	118 [108, 127]	109 [95, 122]	0.016*	119 [106, 127]	111 [97, 124]	0.406
	SBP	117 [102, 130]	108 [99, 122]	119 [106, 130]	0.047*	105 [99, 119]	117 [104, 130]	0.182
	Lactate	4.9 [3.2, 8.2]	9.9 [7.6, 13.4]	3.6 [2.9, 5.8]	< 0.001*	13.3 [10.6, 18.2]	3.8 [3.0, 6.9]	< 0.001*
	Base Deficit	−1.4 [−5.4, 1.6]	−5.3 [−8.2, −1.6]	0 [−3.0, 2.0]	0.139	−7.2 [−13.0, −5.0]	−0.8 [−3.0, 1.7]	0.351
	pH	7.4 [7.3, 7.4]	7.3 [7.1, 7.4]	7.4 [7.3, 7.4]	0.019*	7.2 [7.1, 7.4]	7.4 [7.3, 7.4]	0.086
	Hemoglobin	11.3 [9.9, 12.9]	10.3 [9, 11.8]	11.5 [10.3, 13.2]	0.007*	9.1 [7.9, 11.6]	11.4 [10.1, 12.9]	0.002*
	Platelets	66 [50, 86]	51 [35, 66]	71 [56, 96]	0.009*	51 [39, 55]	68 [55, 93]	0.001*
16-20 hours		n = 180	n = 34	n = 86		n = 13	n = 106
	Heart Rate	112 [97, 123]	114 [104, 132]	109 [91, 121]	0.023*	116 [107, 130]	111 [95, 122]	0.167
	SBP	113 [105, 126]	116 [106, 130]	113 [105, 125]	0.718	115 [99, 120]	113 [106, 127]	0.351
	Lactate	5.8 [3, 8.5]	8.3 [7.1, 12.8]	3.5 [2.6, 6]	< 0.001*	13.1 [8.1, 15.9]	5.3 [2.8, 7.7]	0.003*
	Base Deficit	−0.4 [−5.5, 1.4]	−5.1 [−8, −1.5]	0.5 [−3.2, 2.2]	< 0.001*	−6.4 [−11.4, −0.3]	−0.1 [−4.2, 1.6]	0.018*
	pH	7.4 [7.3, 7.4]	7.3 [7.2, 7.4]	7.4 [7.3, 7.4]	< 0.001*	7.3 [7.1, 7.4]	7.4 [7.3, 7.4]	0.064
	Hemoglobin	10.9 [9.9, 12.6]	10.2 [8.7, 11.8]	11.1 [10.2, 12.7]	0.011*	9.9 [9.2, 11.5]	11.0 [10.1, 12.6]	0.078
	Platelets	69 [45, 92]	44 [37, 83]	76 [59, 96]	0.047*	43 [40, 45]	72 [57, 98]	0.001*
20-24 hours		n = 180	n = 31	n = 86		n = 10	n = 107
	Heart Rate	110 [94, 120]	108 [99, 129]	113 [91, 119]	0.008*	117 [108, 135]	109 [92, 120]	0.046*
	SBP	112 [105, 127]	108 [99, 129]	113 [107, 127]	0.630	103 [99, 129]	113 [106, 127]	0.244
	Lactate	4.4 [3.3, 7.2]	7.6 [5.4, 10.1]	4 [3, 4.8]	0.001*	12.6 [8.2, 17.2]	4.3 [3.2, 6.5]	0.067
	Base Deficit	−0.7 [−5, 1.8]	−5 [−8.8, −0.3]	0.0 [−2.9, 2.4]	0.001*	−8.8 [−11.0, −6.1]	−0.6 [−2.9, 2.0]	0.001*
	pH	7.4 [7.3, 7.4]	7.3 [7.2, 7.4]	7.4 [7.4, 7.4]	0.019*	7.2 [7.2, 7.3]	7.4 [7.3, 7.4]	0.014*
	Hemoglobin	11.1 [9.6, 12.8]	10.4 [9.5, 12.1]	11.6 [10.1, 12.8]	0.443	10.0 [7.5, 10.8]	11.6 [9.8, 12.8]	0.037*
	Platelets	65 [52, 90]	46 [37, 67]	68 [58, 94]	< 0.001*	32 [22, 45]	67 [56, 91]	< 0.001*

Open in a new tab

Systolic blood pressure (SBP), not applicable (n/a). Continuous variables reported as median with [Q1, Q3]

Laboratory Values Time-Series Analysis

At each time interval following the initiation of UMT, the deceased group have a higher lactate level, more acidic pH and a lower base deficit than the survivors. The difference in pH and base deficit becomes significant at the 2-4h time interval and persists through the duration of transfusion while the difference in lactate becomes significant at the 4-8h interval and persists through the duration of transfusion [Table 2, Figure 1B].

Figure 1: — Trends of clinical parameters (**1A:** Vital Signs, **1B:** Lab Values, **1C:** Blood Products Transfused) during first 48hrs following initiation of transfusion for trauma patients undergoing UMT, stratified by mortality

Cryoprecipitate (cryo), Fresh Frozen Plasma (FFP), Heart Rate (HR), Interquartile Range (IQR), Packed Red Blood Cell (pRBC), Systolic Blood Pressure (SBP), Median trends of clinical parameters (1A: Vital Signs, 1B: Lab Values, 1C: Blood Products Transfused), IQRs represented by shaded ribbons

The deceased and survivors had similar starting hemoglobin (111.7 vs 12.1, p = 0.208) and platelet (187 vs 206, p - 0.810). The deceased then have a significantly lower hemoglobin than the survivors in all time intervals between 4-24h, with both cohorts stabilizing to a similar median beyond the 24h time window (9.3 vs 9.4, p = 0.657). Both the deceased and survivors experience a significant decrease in platelet counts during the initial 24h, but is significantly lower in the deceased from hour 8 and onward. This difference becomes most pronounced in the 20-24h time interval (55 versus 327, p < 0.001) [Table 2, Figure 1B].

Blood Product Transfusion Time-Series Analysis

In the first 0-2h, the deceased cohort received 27 total units of blood product compared to 19 total units in the survivors (p < 0.001), 16 vs 10 units of pRBC (p < 0.001), 10 vs 6 units of FFP (p = 0.002) and 1 vs 0 units of platelets (p < 0.001). In the 2-4h time interval, the deceased received a median of 25 total units compared to 15 in the survivors (p < 0.001), with significantly greater amounts of pRBC (12 vs 8, p < 0.001) and FFP (10 vs 7, p < 0.001) transfusion. [Table 3, Figure 1C].

Table 3:

Time-specific blood products transfused during first 24 hours following initiation of transfusion for trauma patients undergoing UMT stratified by mortality

Model type		Overall	Overall survivor			48-hour survivor
Time Interval	Variable		Deceased	Survivors	p-value	Deceased	Survivors	p-value
0-2 hours		n = 180	n = 94	n = 86		n = 72	n = 107
	Total Product	26 [11, 36]	32 [19, 47]	22 [6, 31]	< 0.001*	34 [21, 49]	22 [7, 31]	< 0.001*
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 2]	0.174	0 [0, 0]	0 [0, 2]	0.138
	pRBC	12 [5, 18]	16 [8, 20]	10 [3, 16]	< 0.001*	17 [10, 22]	10 [3, 16]	< 0.001*
	FFP	9 [2, 13]	10 [4, 15]	6 [1, 11]	0.002*	11 [6, 16]	6 [0, 11]	< 0.001*
	Platelets	6 [0, 6]	1 [0, 6]	0 [0, 1]	< 0.001*	1 [0, 6]	0 [0, 6]	0.002*
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.592	0 [0, 0]	0 [0, 0]	0.203
	pRBC:FFP	1.3 [1.0, 1.8]	1.3 [1.0, 1.8]	1.3 [1.0, 1.8]	0.768	1.4 [1.0, 1.8]	1.3 [1.0, 1.8]	0.922
2-4 hours		n = 180	n = 92	n = 86		n = 71	n = 106
	Total Product	24 [14, 39]	30 [16, 45]	21 [8, 33]	< 0.001*	29 [16, 47]	23 [10, 36]	0.006*
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.517	0 [0, 0]	0 [0, 0]	0.898
	pRBC	9 [4, 17]	12 [6, 19]	8 [3, 15]	< 0.001*	12 [5, 20]	9 [3, 15]	0.005*
	FFP	9 [4, 15]	10 [5, 18]	7 [3, 12]	< 0.001*	10 [5, 19]	7 [3, 14]	0.002*
	Platelets	6 [0, 6]	6 [0, 12]	6 [0, 6]	0.079	6 [0, 12]	6 [0, 6]	0.527
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.511	0 [0, 0]	0 [0, 0]	0.821
	pRBC:FFP	1.0 [0.8, 1.3]	1.0 [0.8, 1.3]	1.0 [0.8, 1.4]	0.605	1.0 [0.8, 1.3]	1.1 [0.9, 1.3]	0.521
4-8 hours		n = 180	n = 83	n = 86		n = 62	n = 107
	Total Product	6 [0, 17]	8 [0, 27]	3.5 [0, 10]	< 0.001*	9 [0, 30]	5 [0, 11]	0.005*
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.749	0 [0, 0]	0 [0, 0]	0.980
	pRBC	1 [0, 7]	5 [0, 11]	0 [0, 3]	< 0.001*	5 [0, 12]	1 [0, 5]	0.002*
	FFP	1 [0, 6]	2 [0, 11]	0 [0, 3]	< 0.001*	3 [0, 12]	1 [0, 4]	0.004*
	Platelets	0 [0, 6]	0 [0, 6]	0 [0, 6]	0.139	0 [0, 1]	0 [0, 1]	0.147
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.182	0 [0, 0]	0 [0, 0]	0.249
	pRBC:FFP	1.1 [0.9, 1.5]	1.1 [1.0, 1.5]	1.0 [0.7, 1.8]	0.670	1.2 [1.0, 1.5]	1.0 [0.8, 1.5]	0.873
8-12 hours		n = 180	n = 56	n = 86		n = 35	n = 107
	Total Product	0 [0, 6]	0 [0, 11]	0 [0, 6]	0.105	0 [0, 12]	0 [0, 6]	0.237
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.985	0 [0, 0]	0 [0, 0]	0.633
	pRBC	0 [0, 1]	0 [0, 3]	0 [0, 0]	0.027*	0 [0, 3]	0 [0, 1]	0.174
	FFP	0 [0, 2]	0 [0, 3.5]	0 [0, 1]	0.030*	0 [0, 3]	0 [0, 1]	0.125
	Platelets	0 [0, 0]	0 [0, 0]	0 [0, 6]	0.479	0 [0, 0]	0 [0, 0]	0.917
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.754	0 [0, 0]	0 [0, 0]	0.494
	pRBC:FFP	1.0 [0.0, 1.3]	1.1 [0.9, 1.3]	0.7 [0.0, 1.2]	0.819	1.0 [0.9, 1.1]	0.9 [0.0, 1.4]	0.911
12-16 hours		n = 180	n = 41	n = 86		n = 20	n = 107
	Total Product	0 [0, 5]	1 [0, 8]	0 [0, 3]	0.011*	0 [0, 6.5]	0 [0, 4]	0.546
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	n/a	0 [0, 0]	0 [0, 0]	n/a
	pRBC	0 [0, 1]	0 [0, 2]	0 [0, 1]	0.021*	0 [0, 1]	0 [0, 1]	0.685
	FFP	0 [0, 1]	0 [0, 2]	0 [0, 0]	0.008*	0 [0, 1]	0 [0, 1]	0.593
	Platelets	0 [0, 0]	0 [0, 6]	0 [0, 0]	0.066*	0 [0, 0.5]	0 [0, 0]	0.583
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.437	0 [0, 0]	0 [0, 0]	0.509
	pRBC:FFP	1.0 [0.4, 1.0]	1.0 [0.4, 1.0]	0.8 [0.2, 1.5]	0.859	1.0 [0.4, 1.0]	1.0 [0.5, 1.0]	0.816
16-20 hours		n = 180	n = 34	n = 86		n = 13	n = 106
	Total Product	0 [0, 6]	0 [0, 6]	0 [0, 2]	0.293	0 [0, 6]	0 [0, 6]	0.711
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.325	0 [0, 0]	0 [0, 0]	0.320
	pRBC	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.446	0 [0, 0]	0 [0, 0]	0.754
	FFP	0 [0, 0]	0 [0, 1]	0 [0, 0]	0.311	0 [0, 2]	0 [0, 0]	0.487
	Platelets	0 [0, 0]	0 [0, 6]	0 [0, 0]	0.036*	0 [0, 0]	0 [0, 0]	0.961
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.393	0 [0, 0]	0 [0, 0]	0.490
	pRBC:FFP	0.8 [0.0, 1.0]	0.8 [0.0, 1.1]	0.9 [0.0, 1.0]	0.782	0.8 [0.4, 0.9]	0.8 [0.0, 1.1]	0.376
20-24 hours		n = 180	n = 31	n = 86		n = 10	n = 107
	Total Product	0 [0, 1]	0 [0, 1]	0 [0, 0]	0.838	0 [0, 0]	0 [0, 1]	0.760
	Whole Blood	0 [0, 0]	0 [0, 0]	0 [0, 0]	n/a	0 [0, 0]	0 [0, 0]	n/a
	pRBC	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.157	0 [0, 0]	0 [0, 0]	0.406
	FFP	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.410	0 [0, 0]	0 [0, 0]	0.387
	Platelets	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.012*	0 [0, 0]	0 [0, 0]	0.012*
	Cryo	0 [0, 0]	0 [0, 0]	0 [0, 0]	0.555	0 [0, 0]	0 [0, 0]	0.025*
	pRBC:FFP	0.0 [0.0, 1.2]	0.0 [0.0, 1.5]	0.0 [0.0, 1.2]	0.871	1.2 [1.0, 1.5]	0.0 [0.0, 1.2]	0.152

Open in a new tab

Packed red blood cells (pRBC), fresh frozen plasma (FFP), cryoprecipitate (Cryo). Continuous variables reported as median with [Q1, Q3]

This trend persisted in the 4-8h, 8-12h and 12-16h time intervals, with the deceased receiving more total product and significantly higher rates of pRBC and FFP. In the 12-16h, 16-20h and 20-2h4 hour time intervals, the deceased received not only higher rates of pRBC and FFP, but also of platelets. There were no significant differences in transfusion ratios between the deceased and survivors during any time intervals. [Table 3, Figure 1C].

Decision Tree Models

Overall, the decision models were able to predict mortality with 59.12-81.5% accuracy. The baseline model had an AUC of 62.7 [SD 5.3] for any given time point after the initiation of transfusion. The highest performing models were the 4-8h (AUC 81.5 [SD 4.2]) and 8-12h (AUC 78.0 [SD 6.2]) time intervals. The lowest performing model was 16-20h (AUC 59.1, [SD 8.9]). In the 0-2h and 2-4h time intervals, presence of an ED thoracotomy was selected as the primary decision tree node while in hours 4-8h, 8-12h and 12-16h, this shifted to arterial pH value. At 16-20h and 20-2h4, lactate was chosen as the top node. [Table 4, Figure 2]

Table 4:

Time specific- decision tree models predictive accuracy

Time Interval	N at Risk	Time Interval-Specific Mortality Rate	Cumulative Mortality Rate	AUC, [SD]	Decision Tree Primary Node
0 hours	180	0%	0%	62.7 [5.3]	Pulse
0-2 hours	179	0.6%	0.5%	60.5 [5.8]	ED Thoracotomy
2-4 hours	177	1.1%	1.7%	69.3 [4.2]	ED Thoracotomy
4-8 hours	169	4.5%	6.1%	81.5 [4.6]	Arterial pH
8-12 hours	142	10.1%	21.1%	78.0 [6.2]	Arterial pH
12-16 hours	127	10.6%	29.4%	63.7 [5.8]	Arterial pH
16-20 hours	120	5.5%	33.3%	59. 1 [8.9]	Lactate
20-24 hours	116	3.3%	35.6%	75.9 [9.2]	Lactate

Open in a new tab

Model-specific summary characteristics predicting patient surviving past 48 hours.. Number at risk decreases over time as patients expire in earlier time periods. Time interval-Specific Mortality Rate represents the mortality rate of the remaining population during that time period. Cumulative Mortality Rate represents the overall mortality rate of the entire population as time progresses. Area under the curve (AUC) [SD] as estimated via 10-fold cross-validation is reported as measure of model performance. AUC values close to 1 are ideal and represent perfect classification; values close to 0.5 signify a model that is "guessing" mortality or survival. A decision tree's primary node is its first choice in variable selection which leads to the best outcome in model performance within the training data.

Figure 2: — Representative time-specific decision tree models for trauma patients undergoing UMT, **2A:** 0-2 hours, **2B:** 4-8 hours, **2C:** 12-16 hours, **2D:** 16-20 hours

Glasgow Coma Scale (GCS), heart rate (HR), systolic blood pressure (SBP), Injury Severity Score (ISS), packed red blood cells (pRBC), fresh frozen plasma (FFP), emergency department (ED), length of stay (LOS), diastolic blood pressure (DBP), “changc in” (▲), whole blood (WB), hemoglobin (Hgb), platelets transfused (Pit), base deficit (BD)

*Each terminal tree leaf represents probability of mortality

In the 0-2h model, additional factors most predictive of mortality included a shorter time in the ED, a lower heart rate, greater ISS and more units of platelets transfused. In hours 2-4h, receiving more pRBC or FFP, a higher pRBC:FFP ratio, a lower heart and lower GCS were all predictive of mortality. In 4-16h, a lower blood pressure, lower heart rate, lower hemoglobin and lower platelet value became the predictors selected for by the model. In 16-24h, a higher TEG R time, older age and greater FFP units transfused became the more prominent variables for predicting mortality [Table 4, Figure 2].

Discussion

Ultra-massive transfusion (UMT) is a potentially lifesaving yet incredibly resource demanding intervention for trauma patients in severe hemorrhagic shock. Identifying the threshold beyond which UMT transitions from a necessary therapy to a costly expenditure of a limited resource has been challenging, and there are currently no consensus guidelines in the literature directing these decisions. This study is one of the first to employ machine learning predictive modeling as an innovative strategy to address this question and to also incorporate how these factors change over time during an active resuscitation. The results not only support that ML can successfully integrate complex data to predict mortality, but also highlight a few key principles of UMT resuscitation that have the potential to improve the way we both care for this patient population and research this management strategy.

First, our findings support much of the current literature in suggesting that increased blood product transfusion is not associated with increased survival [7,9,11]. At each time interval during the initial 24 hours of resuscitation, the deceased cohort received significantly more total blood, total pRBCs, and total FFP than survivors. We recognize this likely serves as a proxy for time to hemorrhage control. Additionally, in keeping with published data, patients who did not survive tended to present with greater hemodynamic instability and in a more advanced state of shock [7-10]. However, while we observed this trend in patients’ single presenting set of vitals, the median vitals among survivors and deceased at 0-2h and 2-4h did not differ significantly. We interpret this as support for the efficacy of early and aggressive resuscitation in temporizing the hemodynamics of critically ill trauma patients. While not observed in the overall survival analysis, survivors past 48-hours more often had transfusion ratios less than 1.5:1 pRBC:FFP, reinforcing the importance of a balanced resuscitation.

Additionally, our time-specific decision tree models support the notion that resuscitation is a dynamic process and that the clinical and physiologic parameters associated with survival do change over time. Current literature has focused on identifying a transfusion ceiling, or cutoff point for patients undergoing UMT but perhaps this decision is more nuanced than a single numeric value. Our models demonstrate that in the early hours of a resuscitation, factors such as lower heart rate, greater ISS, more total pRBC, more total FFP, a lower GCS and the presence of an ED thoracotomy are associated with a higher likelihood of mortality. However, past 4 hours, a lower pH, lower hemoglobin, greater lactic acidosis and more severe base deficit became more predictive of mortality. This variability highlights a critical concept - that the answer of “when to stop” is not a single numeric value. Instead, it is likely the interaction of all of these foundational concepts of trauma resuscitation; 1:1:1 transfusion ratios, administration of TXA, early correction of acidosis and traumatic coagulopathy, etc.. considered in a patient specific and time-sensitive manner.

Machine learning predictive modeling is an innovative strategy that enables this type of complex analysis. Our models were able to predict mortality with 59-82% percent accuracy for any given patient at any given time point after the initiation of transfusion. This range reflects a moderate to high performing modeling in the context of ML. Though these models are not in a stage of development that is ready for clinical use, this robust performance supports them as a promising tool for the future. The envisioned utility of these models is to add a layer of objectivity to the complex and ethical challenges of a UMT resuscitation. They are not intended to dictate clinical care, but rather to serve as evidence-based tools to help guide providers and better inform practice. They can also be used to guide appropriate resource allocation, triage in busy trauma centers or mass causality situations and the blood bank optimize their daily supply.

Furthermore, machine learning is powered by data, and prospective multicenter studies are required to expand upon the sample size of this study and thereby the robustness and validity of these models. A larger dataset would aid in correcting many of the limitations discussed below, allowing for external validation and potential prospective use.

Limitations

There are several limitations to this study the authors would like to acknowledge. First, it was retrospective in nature and therefore limited in the ability to control for confounding variables such as duration of pre-hospital time, pre-hospital blood product and medication administration, cause of death, presence of a traumatic brain injury, indication for return to the operating room and markers of adequate resuscitation. Additionally, our institution implemented an empiric emergent whole blood transfusion program in 2019. Therefore, the whole blood product transfusion rates depicted in this study are limited and may not accurately reflect current practice.

Furthermore, while the time-specific dataset was comprehensive, there was significant variability in the time intervals and data points captured for each patient. This required workarounds in the coding to normalize the data in a way that could standardize its input format for the models. Our sample size was also limited, which does impact the robustness of the time-specific analysis. There are inherently fewer patients and thereby fewer data points at each subsequent time point, potentially allowing outliers to skew the results more substantially. A larger dataset through a multi-institutional study is required to validate the models.

Additionally, our models were run based on survival through 48 hours of admission. This was considered the time period for active resuscitation, and attempts to exclude those that expire due to in-house complications such as sepsis, cardiac arrest, pulmonary embolisms, etc. However, it is possible some patients expired shortly after the 48 hour period from reasons associated with their UMT that would not be reflected in the model. Lastly, the authors recognize the role of ML in this setting is limited to predictive and descriptive and requires further refinement and multi-center validation prior to being applied in any clinical setting.

Conclusions

Our results suggest that instead of seeking a single, numeric value as a transfusion ceiling, that the answer to “when to stop” is likely based on the complex integration of patient and time-specific factors during a UMT resuscitation. Furthermore, this is a study of feasibility, supporting machine learning modeling as an innovative and pragmatic tool to address this difficult question. With multicenter studies, the authors envision external validation and improved robustness of these models, allowing them to add a layer of objectivity and data to help guide the clinical and ethical challenges of a UMT resuscitation. If providers had evidence-based survival probabilities tailored specifically to their patient during each time interval following the initiation of transfusion, we believe the answer of “when to stop” may become more clear and resources could be more appropriately allocated.

Supplementary Material

Supplemental Data File (.doc, .tif, pdf, etc.)_1

Strobe Checklist: STROBE checklist for observational studies with page number and relevant text for each item

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___1.docx^{(39.1KB, docx)}

Supplemental Data File (.doc, .tif, pdf, etc.)_2

Supplemental Figure 1: Flow-diagram of cohort generation

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___2.jpg^{(116.8KB, jpg)}

Supplemental Data File (.doc, .tif, pdf, etc.)_3

Supplemental Figure 2: Survival curves for trauma patients undergoing UMT stratified by (2A) gender, (2B) age, (2C) mechanism of injury and (2D) injury severity score

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___3.jpg^{(183.1KB, jpg)}

Acknowledgements:

Thank you to the Grady Trauma Registry team for their assistance in data acquisition

Funding:

C Meyer supported by NIH T32 Training Grant, NIGMS (5T32GM095442-11)

Footnotes

Conflicts of Interest: J Nguyen receives honoraria from Prytime Medical, Zimmer Biomet, and Teleflex for educational events.

The authors have no additional conflicts of interest to report.

Author COI Forms: Conflict of interest forms completed independently by each author

Meeting Affiliation: This research has been accepted for quickshot oral presentation at AAST 2023

References

1.Park Y, Lee GJ, Lee MA, Choi KK, Gwak J, Hyun SY, et al. Major causes of preventable death in trauma patients. Journal of Trauma and Injury. 2021; 34(4):225–32. [Google Scholar]
2.Kalkwarf KJ, Drake SA, Yang Y, Thetford C, Myers L, Brock M, et al. Bleeding to death in a big city: An analysis of all trauma deaths from hemorrhage in a metropolitan area during 1 Year. Journal of Trauma and Acute Care Surgery. 2020; 89(4):716–22. [DOI] [PubMed] [Google Scholar]
3.Carroll SL, Dye DW, Smedley WA, Stephens SW, Reiff DA, Kerby JD, et al. Early and prehospital trauma deaths: Who might benefit from advanced resuscitative care? Journal of Trauma and Acute Care Surgery. 2020; 88(6):776–82. [DOI] [PubMed] [Google Scholar]
4.Chapman MP, Moore EE, Chin TL, Ghasabyan A, Chandler J, Stringham J, et al. Combat: Initial experience with a randomized clinical trial of plasma-based resuscitation in the field for traumatic hemorrhagic shock. Shock. 2015; 44(0 1):63–70. [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Lammers Danny T. MD; Holcomb John B. MD. Damage control resuscitation in adult trauma patients: What you need to know. Journal of Trauma and Acute Care Surgery. 2023; 95(4):p 464–471. [DOI] [PubMed] [Google Scholar]
6.Black JA, Strickland VP, Juneja K, Holcomb JB. Complications of hemorrhagic shock and massive transfusion – a comparison before and after the damage control resuscitation era. Shock. 2021; 56(1). [DOI] [PubMed] [Google Scholar]
7.Cannon JW, Khan MA, Raja AS, Cohen MJ, Como JJ, Cotton BA, et al. Damage control resuscitation in patients with severe traumatic hemorrhage. Journal of Trauma and Acute Care Surgery. 2017; 82(3):605–17. [DOI] [PubMed] [Google Scholar]
8.Cole E, Weaver A, Gall L, West A, Nevin D, Tallach R, et al. A decade of damage control resuscitation. Annals of Surgery. 2021; 273(6):1215–20. [DOI] [PubMed] [Google Scholar]
9.Holcomb JB, Fox EE, Wade CE. The prospective observational multicenter major trauma transfusion (PROMMTT) study. Journal of Trauma and Acute Care Surgery. 2013; 148(2):127–36. [DOI] [PubMed] [Google Scholar]
10.Spahn DR, Bouillon B, Cerny V, Duranteau J, Filipescu D, Hunt BJ, et al. The European guideline on management of major bleeding and coagulopathy following trauma: Fifth Edition. Critical Care. 2019; 23(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Thompson P, Strandenes G. The history of fluid resuscitation for bleeding. Damage Control Resuscitation. 2019; 3–29. [Google Scholar]
12.Holcomb JB, Tilley BC, Baraniuk S, et al. Transfusion of Plasma, Platelets, and Red Blood Cells in a 1:1:1 vs a 1:1:2 Ratio and Mortality in Patients With Severe Trauma: The PROPPR Randomized Clinical Trial. JAMA. 2015;313(5):471–482 [DOI] [PMC free article] [PubMed] [Google Scholar]
13.McQuilten ZK, Flint AW, Green L, Sanderson B, Winearls J, Wood EM. Epidemiology of massive transfusion – a common intervention in need of a definition. Transfusion Medicine Reviews. 2021; 35(4):73–9. [DOI] [PubMed] [Google Scholar]
14.Dorken Gallastegi A, Secor JD, Maurer LR, Dzik WS, Saillant NN, Hwabejire JO, et al. Role of transfusion volume and transfusion rate as markers of futility during ultramassive blood transfusion in trauma. Journal of the American College of Surgeons. 2022; 235(3):468–80. [DOI] [PubMed] [Google Scholar]
15.Matthay ZA, Hellmann ZJ, Callcut RA, Matthay EC, Nunez-Garcia B, Duong W, et al. Outcomes after ultramassive transfusion in the modern era: An Eastern Association for the surgery of trauma multicenter study. Journal of Trauma and Acute Care Surgery. 2021; 91(1):24–33. [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Dzik WS, Ziman A, Cohen C, Pai M, Lozano M, Kaufman RM, et al. Survival after ultramassive transfusion: A review of 1360 cases. Transfusion. 2015; 56(3):558–63. [DOI] [PubMed] [Google Scholar]
17.Meyer CH, Bailey NM, Leslie SL, Thrasher K, Grady ZJ, Smith RN, et al. Defining Ultra-Massive Transfusion through a Systematic Review. The American Journal of Surgery. 2023. Epub ahead of print. [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Yu AJ, Inaba K, Biswas S, De Leon LA, Wong M, Benjamin E, et al. Supermassive transfusion: A 15-year single center experience and outcomes. The American Surgeon. 2018; 84(10):1617–21. [PubMed] [Google Scholar]
19.Gurney JM, Staudt AM, Holcomb J, Martin M, Spinella P, et al. Finding the Bleeding Edge: 24-hour Mortality by Unit of Blood Product Transfused in Combat Casualties from 2002-2020. J Trauma Acute Care Surg. 2023. Epub ahead of print. [DOI] [PubMed] [Google Scholar]
20.Quintana MT, Zebley JA, Vincent A, Chang P, Estroff J, Sarani B, et al. Cresting mortality: Defining a plateau in ongoing massive transfusion. Journal of Trauma and Acute Care Surgery. 2022; 93(1):43–51. [DOI] [PubMed] [Google Scholar]
21.Velmahos GC. Is there a limit to massive blood transfusion after severe trauma? Archives of Surgery. 1998;133(9):947–52. [DOI] [PubMed] [Google Scholar]
22.Morris MC, Niziolek GM, Baker JE, Huebner BR, Hanseman D, Makley AT, et al. Death by decade: Establishing a transfusion ceiling for futility in massive transfusion. Journal of Surgical Research. 2020; 252:139–46. [DOI] [PubMed] [Google Scholar]
23.Liu S, Fujii Q, Serio F, McCague A. Massive blood transfusions and outcomes in trauma patients; an intention to treat analysis. Bulletin of Emergency and Trauma. 2018; 6(3):217–20. [DOI] [PMC free article] [PubMed] [Google Scholar]
24.McGann PT, Weyand AC. Lessons learned from the COVID-19 pandemic blood supply crisis. J Hosp Med. 2022;17(7):574–576. Epub 2022 Jun 21. [DOI] [PMC free article] [PubMed] [Google Scholar]
25.Ang D, Fakhry SM, Watts DD, Liu H, Morse JL, Armstrong J, et al. Data-driven blood transfusion thresholds for severely injured patients during blood shortages. Journal of Surgical Research. 2023; 291:17–24. [DOI] [PMC free article] [PubMed] [Google Scholar]
26.Beam AL, Manrai AK, Ghassemi M. Challenges to the Reproducibility of Machine Learning Models in Health Care. JAMA. 2020;323(4):305–306. [DOI] [PMC free article] [PubMed] [Google Scholar]
27.Jakhar D, Kaur I, Artificial intelligence, machine learning and deep learning: definitions and differences, Clinical and Experimental Dermatology. 2020; 45(1): 131–132 [DOI] [PubMed] [Google Scholar]
28.Peng HT, Siddiqui MM, Rhind SG, Zhang J, da Luz LT, Beckett A. Artificial Intelligence and machine learning for hemorrhagic trauma care. Military Medical Research. 2023;10(1). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.Liu NT, Salinas J. Machine learning for predicting outcomes in trauma. Shock. 2017; 48(5):504–10. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplemental Data File (.doc, .tif, pdf, etc.)_1

Strobe Checklist: STROBE checklist for observational studies with page number and relevant text for each item

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___1.docx^{(39.1KB, docx)}

Supplemental Data File (.doc, .tif, pdf, etc.)_2

Supplemental Figure 1: Flow-diagram of cohort generation

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___2.jpg^{(116.8KB, jpg)}

Supplemental Data File (.doc, .tif, pdf, etc.)_3

Supplemental Figure 2: Survival curves for trauma patients undergoing UMT stratified by (2A) gender, (2B) age, (2C) mechanism of injury and (2D) injury severity score

NIHMS1941027-supplement-Supplemental_Data_File___doc___tif__pdf__etc___3.jpg^{(183.1KB, jpg)}

[R1] 1.Park Y, Lee GJ, Lee MA, Choi KK, Gwak J, Hyun SY, et al. Major causes of preventable death in trauma patients. Journal of Trauma and Injury. 2021; 34(4):225–32. [Google Scholar]

[R2] 2.Kalkwarf KJ, Drake SA, Yang Y, Thetford C, Myers L, Brock M, et al. Bleeding to death in a big city: An analysis of all trauma deaths from hemorrhage in a metropolitan area during 1 Year. Journal of Trauma and Acute Care Surgery. 2020; 89(4):716–22. [DOI] [PubMed] [Google Scholar]

[R3] 3.Carroll SL, Dye DW, Smedley WA, Stephens SW, Reiff DA, Kerby JD, et al. Early and prehospital trauma deaths: Who might benefit from advanced resuscitative care? Journal of Trauma and Acute Care Surgery. 2020; 88(6):776–82. [DOI] [PubMed] [Google Scholar]

[R4] 4.Chapman MP, Moore EE, Chin TL, Ghasabyan A, Chandler J, Stringham J, et al. Combat: Initial experience with a randomized clinical trial of plasma-based resuscitation in the field for traumatic hemorrhagic shock. Shock. 2015; 44(0 1):63–70. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R5] 5.Lammers Danny T. MD; Holcomb John B. MD. Damage control resuscitation in adult trauma patients: What you need to know. Journal of Trauma and Acute Care Surgery. 2023; 95(4):p 464–471. [DOI] [PubMed] [Google Scholar]

[R6] 6.Black JA, Strickland VP, Juneja K, Holcomb JB. Complications of hemorrhagic shock and massive transfusion – a comparison before and after the damage control resuscitation era. Shock. 2021; 56(1). [DOI] [PubMed] [Google Scholar]

[R7] 7.Cannon JW, Khan MA, Raja AS, Cohen MJ, Como JJ, Cotton BA, et al. Damage control resuscitation in patients with severe traumatic hemorrhage. Journal of Trauma and Acute Care Surgery. 2017; 82(3):605–17. [DOI] [PubMed] [Google Scholar]

[R8] 8.Cole E, Weaver A, Gall L, West A, Nevin D, Tallach R, et al. A decade of damage control resuscitation. Annals of Surgery. 2021; 273(6):1215–20. [DOI] [PubMed] [Google Scholar]

[R9] 9.Holcomb JB, Fox EE, Wade CE. The prospective observational multicenter major trauma transfusion (PROMMTT) study. Journal of Trauma and Acute Care Surgery. 2013; 148(2):127–36. [DOI] [PubMed] [Google Scholar]

[R10] 10.Spahn DR, Bouillon B, Cerny V, Duranteau J, Filipescu D, Hunt BJ, et al. The European guideline on management of major bleeding and coagulopathy following trauma: Fifth Edition. Critical Care. 2019; 23(1). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R11] 11.Thompson P, Strandenes G. The history of fluid resuscitation for bleeding. Damage Control Resuscitation. 2019; 3–29. [Google Scholar]

[R12] 12.Holcomb JB, Tilley BC, Baraniuk S, et al. Transfusion of Plasma, Platelets, and Red Blood Cells in a 1:1:1 vs a 1:1:2 Ratio and Mortality in Patients With Severe Trauma: The PROPPR Randomized Clinical Trial. JAMA. 2015;313(5):471–482 [DOI] [PMC free article] [PubMed] [Google Scholar]

[R13] 13.McQuilten ZK, Flint AW, Green L, Sanderson B, Winearls J, Wood EM. Epidemiology of massive transfusion – a common intervention in need of a definition. Transfusion Medicine Reviews. 2021; 35(4):73–9. [DOI] [PubMed] [Google Scholar]

[R14] 14.Dorken Gallastegi A, Secor JD, Maurer LR, Dzik WS, Saillant NN, Hwabejire JO, et al. Role of transfusion volume and transfusion rate as markers of futility during ultramassive blood transfusion in trauma. Journal of the American College of Surgeons. 2022; 235(3):468–80. [DOI] [PubMed] [Google Scholar]

[R15] 15.Matthay ZA, Hellmann ZJ, Callcut RA, Matthay EC, Nunez-Garcia B, Duong W, et al. Outcomes after ultramassive transfusion in the modern era: An Eastern Association for the surgery of trauma multicenter study. Journal of Trauma and Acute Care Surgery. 2021; 91(1):24–33. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] 16.Dzik WS, Ziman A, Cohen C, Pai M, Lozano M, Kaufman RM, et al. Survival after ultramassive transfusion: A review of 1360 cases. Transfusion. 2015; 56(3):558–63. [DOI] [PubMed] [Google Scholar]

[R17] 17.Meyer CH, Bailey NM, Leslie SL, Thrasher K, Grady ZJ, Smith RN, et al. Defining Ultra-Massive Transfusion through a Systematic Review. The American Journal of Surgery. 2023. Epub ahead of print. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.Yu AJ, Inaba K, Biswas S, De Leon LA, Wong M, Benjamin E, et al. Supermassive transfusion: A 15-year single center experience and outcomes. The American Surgeon. 2018; 84(10):1617–21. [PubMed] [Google Scholar]

[R19] 19.Gurney JM, Staudt AM, Holcomb J, Martin M, Spinella P, et al. Finding the Bleeding Edge: 24-hour Mortality by Unit of Blood Product Transfused in Combat Casualties from 2002-2020. J Trauma Acute Care Surg. 2023. Epub ahead of print. [DOI] [PubMed] [Google Scholar]

[R20] 20.Quintana MT, Zebley JA, Vincent A, Chang P, Estroff J, Sarani B, et al. Cresting mortality: Defining a plateau in ongoing massive transfusion. Journal of Trauma and Acute Care Surgery. 2022; 93(1):43–51. [DOI] [PubMed] [Google Scholar]

[R21] 21.Velmahos GC. Is there a limit to massive blood transfusion after severe trauma? Archives of Surgery. 1998;133(9):947–52. [DOI] [PubMed] [Google Scholar]

[R22] 22.Morris MC, Niziolek GM, Baker JE, Huebner BR, Hanseman D, Makley AT, et al. Death by decade: Establishing a transfusion ceiling for futility in massive transfusion. Journal of Surgical Research. 2020; 252:139–46. [DOI] [PubMed] [Google Scholar]

[R23] 23.Liu S, Fujii Q, Serio F, McCague A. Massive blood transfusions and outcomes in trauma patients; an intention to treat analysis. Bulletin of Emergency and Trauma. 2018; 6(3):217–20. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] 24.McGann PT, Weyand AC. Lessons learned from the COVID-19 pandemic blood supply crisis. J Hosp Med. 2022;17(7):574–576. Epub 2022 Jun 21. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R25] 25.Ang D, Fakhry SM, Watts DD, Liu H, Morse JL, Armstrong J, et al. Data-driven blood transfusion thresholds for severely injured patients during blood shortages. Journal of Surgical Research. 2023; 291:17–24. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R26] 26.Beam AL, Manrai AK, Ghassemi M. Challenges to the Reproducibility of Machine Learning Models in Health Care. JAMA. 2020;323(4):305–306. [DOI] [PMC free article] [PubMed] [Google Scholar]

[R27] 27.Jakhar D, Kaur I, Artificial intelligence, machine learning and deep learning: definitions and differences, Clinical and Experimental Dermatology. 2020; 45(1): 131–132 [DOI] [PubMed] [Google Scholar]

[R28] 28.Peng HT, Siddiqui MM, Rhind SG, Zhang J, da Luz LT, Beckett A. Artificial Intelligence and machine learning for hemorrhagic trauma care. Military Medical Research. 2023;10(1). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.Liu NT, Salinas J. Machine learning for predicting outcomes in trauma. Shock. 2017; 48(5):504–10. [DOI] [PubMed] [Google Scholar]

PERMALINK

TiME OUT: Time-specific Machine-learning Evaluation to Optimize Ultra-massive Transfusion

Courtney H Meyer, MD, MPH

Jonathan Nguyen, DO

Andrew ElHabr, PhD

Nethra Venkatayogi

Tyler Steed, MD, PhD

Judy Gichoya, MD, MS

Jason D Sciarretta, MD

James Sikora, MD

Christopher Dente, MD

John Lyons, MD

Craig M Coopersmith, MD

Crystal Nguyen, PhD

Randi N Smith, MD, MPH

Abstract

Background:

Methods:

Results:

Conclusions:

Level of Evidence:

Media Summary:

Introduction

Methods

Study Design & Inclusion Criteria

Variables of Interest

Machine Learning Modeling

Statistical Analysis

Results

Demographics

Table 1:

Injury Patterns and Clinical Presentation

Blood Product Transfusion

Operative Intervention

Vital Signs Time-Series Analysis

Table 2:

Laboratory Values Time-Series Analysis

Figure 1:

Blood Product Transfusion Time-Series Analysis

Table 3:

Decision Tree Models

Table 4:

Figure 2:

Discussion

Limitations

Conclusions

Supplementary Material

Acknowledgements:

Funding:

Footnotes

References

Associated Data

Supplementary Materials

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases