Abstract
Background
Extracorporeal membrane oxygenation (ECMO) is resource intensive with high mortality. Identifying trauma patients most likely to derive a survival benefit remains elusive despite current ECMO guidelines. Our objective was to identify unique patient risk profiles using the largest database of trauma patients available.
Methods
ECMO patients ≥16 years were identified using Trauma Quality Improvement Program data (2010-2019). Machine learning K-median clustering (ML) utilized 101 variables including injury severity, demographics, comorbidities, and hospital stay information to generate unique patient risk profiles. Mortality and patient and center characteristics were evaluated across profiles.
Results
A total of 1037 patients were included with 33% overall mortality, mean age 32 years, and median ISS = 26. The ML identified 3 unique patient risk profile groups. Although mortality rates were equivalent across the 3 groups, groups were distinguished by (Group 1) young (median 25 years), severely injured (ISS = 34) patients with thoracic and head injuries (99%) via blunt mechanism (93%), and a high prevalence of ARDS (77%); (Group 2) relatively young (median 30 years) and moderately injured (ISS = 22) patients with exposure-related injuries (11%); and (Group 3) older (median 46 years) patients with a high proportion of comorbidities (69%) and extremity injuries (100%). There were no differences based on center ECMO volume, teaching status, or ACS-Level across all 3 groups.
Conclusion
Machine learning compliments traditional analyses by identifying unique mortality risk profiles for trauma patients receiving ECMO. These details can further inform treatment guidelines, clinical decision making, and institutional criteria for ECMO usage.
• ECMO utilization in trauma patients increased over the ten-year period 2010-2019; however mortality risk remained unchanged. • Machine learning analysis identified 3 unique patient groups based on patient characteristics; however, groups did not differ on mortality risk or center characteristics. • Future research should consider these unique patient groups and incorporate granular clinical and center characteristics to inform ECMO guidelines.Key Takeaways
Background
Extracorporeal membrane oxygenation (ECMO) is a highly resource intensive treatment requiring specialized providers and is only available at select trauma centers. Despite advances in critical care, including ventilator management and the use of evidence based adjunctive therapies, mortality for trauma patients who receive ECMO remains high at 30-36%.1-3
Although still infrequent, ECMO utilization has increased in trauma patients over time and the nationwide Extracorporeal Life Support Organization (ELSO) registry was established to track patients receiving this treatment.1,4 However, many centers that provide ECMO do not participate in ELSO and retrospective studies remain limited in study duration and scope.5-7 Longitudinal data regarding further adoption of ECMO and frequency of use for injured patients is not readily available. As such, attempts at identifying patients who may or may not benefit from ECMO have remained inconclusive.7,8
Given the heterogeneity of injury patterns and characteristics among trauma patients who have received ECMO, identifying specific risk factors associated with mortality has been challenging. Machine learning affords efficient analysis of large datasets without a priori factors that would impose bias in more traditional statistical analyses. K-median clustering analysis is one such type of machine learning method. Rather than advanced selection of specific variables or factors to assess associations with outcomes of interest, K-median clustering accounts for all dataset variables and assigns each data point to a cluster represented by its nearest median. The resulting analysis identifies patterns or “clusters” of datapoints, should they exist, based on similarities and differences within the data.
The Trauma Quality Programs Participant Use File is the largest available nationwide database of trauma patients. 9 The purpose of this study was to describe patient characteristics of those receiving ECMO over a 10-year period (2010-2019). We further sought to identify specific patient groups using machine learning K-medians clustering that may be associated with positive or negative outcomes following ECMO. We hypothesized that several distinct patient risk profiles would be identified based on clinical characteristics and injury severity, location, and type.
Patients and Methods
Data Source
This study received exempt status from Scripps Health Institutional Review Board and followed the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for reporting epidemiological studies. 10 Trauma patients were identified from the 2010-2019 American College of Surgeons Trauma Quality Programs Participant Use File (TQP). 9 Data available includes demographic, injury, clinical, hospital, and discharge data as well as International Classification of Diseases, Ninth Revision-Clinical Modification (ICD-9-CM) and 10th Revisions (ICD-10) procedure codes.
Study Population
Patients aged 16 years and older and ICD-9 and ICD-10 procedure codes (39.65; 5A1522F-H; 5A15A2F-H and 5A15223) for ECMO were identified. Patients missing key demographic, injury, comorbidities, hospital complications, or hospital length of stay were excluded. Patient selection schema is detailed in Figure 1. CONSORT patient flowchart.
Variables and Outcomes
TQP variables used include: age, injury severity score (ISS), TQP-reported complication of ARDS, and TQP-reported comorbidities as well as Abbreviated Injury Scale (AIS) “predot” and severity codes. 11 AIS 2005 codes were cross walked to 1998 codes to allow for comparability. Any AIS 2005 codes that did not have a 1998 equivalency were manually categorized by region and severity. AIS codes were then evaluated and categorized into sets according to body region and based on the most severe injury. Abbreviated Injury Scale codes within body regions were subclassified according to injury severity as “low” (AIS 1 to 2), “medium” (AIS 3 to 4), and “high” (AIS 5 to 6) based on research team consensus (see Supplemental Table 1). A body mass index (BMI) of ≥30 was used to identify the presence of obesity.
The primary outcome was to identify distinct patient profiles of those who received ECMO following a traumatic injury. Secondary outcomes were to compare mortality, hospital length of stay, ICU and ventilator days, and discharge disposition across patient profile groups. Additionally, admission year, time to ECMO initiation, total number of ECMO cases for trauma patients during the study interval, and admitting trauma center level were analyzed. The highest trauma center level classification between American College of Surgeons (ACS) verified centers and State designated centers was retained for analysis.
Statistical Analysis
Descriptive statistics were displayed as means with standard deviations, medians with interquartile ranges, or frequencies with column proportions, as appropriate. A two-stage K-medians cluster analysis was performed to create a patient grouping schema based on age, ISS, and presence or absence of comorbidities, body region injury and type, and ARDS status. The first stage included all possible TQP variables. The second stage isolated the most prevalent variables and with the highest utility from a clinical perspective. To assess the mathematical similarity between values of each variable in a multidimensional space, the Gower distance measure was employed. 12 In short, the Gower distance is a measure of similarity between the values a variable can take on that is weighted by the number of all variables used in the procedure. The Gower distance was used because it is appropriate for data containing a mixture of categorical and continuous factors. Because we had no a priori hypothesis on the number of clusters and to generate an unbiased data-driven set of groups, we employed a random starting option to assign a random observation as an initial group center in the multidimensional space. Observations were then iteratively added to the space based on their distance from the initial centers. Over the course of the clustering procedure, the median centers may move and then ultimately “settle” allowing for a number of distinct clusters to be isolated. The Duda-Hart Je(2)/Je(1) index was used to identify the number of clusters that resulted in the most distinct groupings. 13 Colloquial descriptions were subsequently applied to each group based on the quality and prevalence of attributable factors.
Initial analysis of patient demographics, clinical factors, and outcomes was performed using chi-square tests, rank-sum tests, or Kruskal-Wallis tests between clusters as appropriate. Cox proportional hazards modeling evaluated mortality risk by patient cluster groups after adjusting for covariates. Covariates used in model development included trauma level, number of ECMO cases for trauma done at a facility, patient year of admission, and days from admission to ECMO initiation, as many of the direct clinical factors were used in the clustering algorithm. Follow-up time for survival analyses utilized the time from ECMO initiation to death or hospital discharge; those with missing values for ECMO initiation time were excluded from this stage of the analysis. The Kaplan-Meier curve was used to evaluated time and outcome. The proportional hazards assumption was evaluated graphically using scaled Schoenfeld residuals whereby a plot displaying a non-random pattern against time was classified as being in violation of the assumption. A P-value of <.050 was used to identify statistical significance. Data were managed and analyzed using Stata MP version 17.0 StataCorp LP, College Station, TX).
Results
Center Characteristics
Extracorporeal Membrane Oxygenation Center Characteristics by Group.
IQR, interquartile range.
*bolded values <.05.
Patient Characteristics and Outcomes
Extracorporeal Membrane Oxygenation Group Demographics, Injury and Outcome Characteristics.
IQR, interquartile range; MVA, motor vehicle accident; ISS, injury severity score; STAC, Short-Term Acute Care hospital; LTAC, Long-Term Acute Care facility; SNF, Skilled Nursing Facility.
*bolded values <.05.
Heatmap of significant variables with percentage from cluster analysis (ps<.05).
Cluster Group Analysis
The Duda-Hart index determined 3 cluster groups as optimal (J = .9998 & T = .19). These groups were distinct in age, injury severity/type, and comorbidities. Cluster Group 1 (n = 204) consisted predominately of highly injured young patients (median ISS 34 and median age 25 years). The majority of patients in cluster Group 1 had blunt injury (90%) and ARDS (77%). The predominate injury types in Group 1 were head injuries (eg, soft tissue injuries, fractures, or intercranial hemorrhage), thorax injuries (74%), lung injuries (39%), and medium level extremity orthopedic injuries (79%) (see heatmap Figure 2). Group 1 also had the longest hospital length of stay (HLOS), ventilator days, and ICU days (median 31, 21, and 25 days, respectively, ps <.001). Group 2 was the largest (n = 463) and consisted of relatively young (median 30 years) and less injured patients (ISS = 22) experiencing blunt mechanism (55%). Group 2 had the lowest portion of patients with ARDS (34%) but the highest portion of exposure-related injuries (11%). Group 3 (n = 334) was the oldest [median 46 years] and greatest prevalence of comorbidities (18% with diabetes, 26% with hypertension, and 46% were obese) and ARDS (36%). Most patients had blunt mechanism (85%), thorax injuries (71%), and extremity orthopedic fractures (84%) and extremity soft tissue injures (76%). There was no significant difference across the 3 groups for mortality or hospital characteristics of trauma level, bed size, teaching status, or ECMO patient volume (P = .579) (Table 1). However, there were significant differences between the groups in HLOS, ventilator, and ICU days.
Cox Model and Mortality
Cox Proportional Hazards Modeling Mortality.
*bolded values <.05.
Discussion
ECMO requires significant resources, highly trained personnel, and is associated with significant mortality. Previous attempts to identify patients that may fare better with ECMO use have yielded ambiguous results. This study used the ACS-TQP database to identify patients who received ECMO following traumatic injury over a 10-year period (2010-2019). We hypothesized applying ML would further characterize unique patient subgroups and correlate mortality trends. Results demonstrate that 3 unique groups of trauma patients emerge with similar mortality risk between patient groups.
Previous studies have attempted to isolate patient characteristics that are associated with either an increased need for or mortality benefit from ECMO use both in patients with and without traumatic injury.6,7,14-16 Results of these studies, while often not definitive and contradictory have informed current practice guidelines, which reliably focus on ARDS and TBI but are vague for other characteristics and further clarification is needed.16,17 Our findings show that providers are utilizing ECMO for a variety of injuries and patient characteristics. Prior publications and general ECMO practice guidelines focus on a narrow group of patients consisting of those with TBI or ARDS. Less than half of the patients in TQP placed on ECMO had ARDS documented. In trauma patients specifically, no diagnosis other than irreversible injury is an absolute contraindication of ECMO implementation. 17 Focusing on a single complication or single injury type misses the vast majority of patients receiving this procedure nationally.
Here, we not only describe patients with traumatic injury receiving ECMO and characteristics of the centers providing ECMO over a decade but also incorporate advanced machine learning analysis to further identify patient subgroups and their unique mortality risk profiles. Due to increased ECMO use over the COVID-19 pandemic, years prior to 2020 only were included. Our study identified an increasing trend in ECMO use, particularly at Level I trauma centers. Trauma type and injury mechanism reflect national trauma center trends, and the median ISS for patients requiring ECMO is consistently high. 4 ECMO case volume was associated with improved patient survival (HR = .986), suggesting institutional experience and resource availability may improve outcomes.
Regarding cluster group analysis, 3 unique patient mortality risk profiles were identified. Despite the statistical differences among some characteristics of the risk profiles, survival did not statistically differ between groups. Understandably, ARDS was present in all 3 cluster groups; however, it was statistically most prevalent in patients clustered to Group 1. Interestingly, of the 34 injury subtypes that statistically differed among the 3 groups, 2/3 (21) were most associated with Group 1. We did not find ARDS to be significantly associated with any particular injury or combination of injury patterns. Group 1 did have longer ventilator days, ICU days, and hospital length of stay as well as higher ISS than Groups 2 and 3. Given similar mortality rates across all cluster groups, we theorize that ISS is probably the driving factor behind this disparity in hospital outcomes. Taken together, the panoply of varied patient injury types clustered in Group 1 likely reflects the true, multi-injured and complex trauma patient and illustrates the vast combination of injuries and clinical amalgamation as indicators for ECMO utilization in trauma patients.
Group 2, while being the largest, was relatively younger than both the entire cohort median age as well as the oldest cluster (Group 3), and was the least injured. It is interesting to note that exposure and esophageal injuries were most-significantly and almost exclusively represented in this second group, along with a handful of other injury types and patient characteristics. Still, age and low injury severity may have been the most unifying features of this group, but those did not increase survival rates.
Finally, in reference to Group 3, underlying comorbidities and age, which are typically closely related at baseline, are the most notable features of this cluster. Again, the varied and unique characteristics of these groups reiterate the multifactorial components pertaining to ARDS and/or acute cardiorespiratory failure in trauma patients leading to ECMO as a treatment. Despite the emergence of the 3 distinct groups, overall mortality between the groups and compared to historic trends did not differ. Interestingly, this was also irrespective of individual trauma center characteristics.
Again, current ECMO guidelines are vague in terms of identifying candidates for the procedure, and for all patients (not just trauma) focus on ARDS and TBI as the main indicators. Less than half (44%) of trauma patients who received ECMO had an ARDS diagnosis and results from the cluster analysis show it was prevalent in only one profile group, indicating ECMO is being used for other reasons. The data also demonstrate that mortality did not statistically differ between the 3 risk profiles, suggesting factors other than ARDS and TBI are playing a role in mortality risk. Injuries varied across profile groups, however, of cluster analysis identified exposure and esophageal injuries as dominant in one of the profiles. Additionally, age and underlying comorbidities characterized the third risk profile. This analysis provides further clarification on the unique patient profile of those that have been and may be candidates for ECMO, which includes additional characteristics beyond ARDS.
ECMO patients with traumatic injury are a varied and heterogenous group in terms of their injury types, severity, and health status. Traditional statistical methods are used to determine the presence and strength of associations between a risk factor of interest and an outcome. It is increasingly challenging to assess these associations given the large number of variables available in datasets and the risk of Type I error. Conversely, selecting only some of the variables injects bias into the study design. Here, K-medians clustering allowed for an unbiased evaluation of all possible injury-specific variables within the TQP database to determine if unique patient risk profiles exist within the larger cohort. The results add to the current literature by identifying additional characteristics within specific profiles that are not considered in practice guidelines for ECMO. Evaluating mortality risk across these profile groupings indicates they are statistically equivalent; however, the cause of mortality may be different within each profile. Providers should consider these profiles when treating their patients.
The limitations of this analysis include the lack of detailed patient clinical information, inability to compare patients who did not receive ECMO as a control group, survivor bias, and the inability to track patients beyond the index admission among others. Specifically, the database does not distinguish elements of patient care prior to ECMO including the specific clinical factor(s) leading to ECMO initiation, ventilator parameters, days prior to decannulation, or the use of adjunctive treatments or salvage therapies for ARDS. Unfortunately, VA or VV ECMO type is only delineated in procedure codes after 2018; thus, we were unable to identify type of ECMO procedure over the 10-year study period. Additional center and provider characteristics, such as background and training of ECMO providers, open vs closed intensive care unit designation, and overall ECMO center volume (to include non-trauma patients) would also lend further insight to our analysis.
Conclusion
Trauma patients are increasingly receiving ECMO and attempts to differentiate patient characteristics that are associated with greater survival benefit have remained elusive. Although mortality did not differ between groups or over time, machine learning identified 3 distinct trauma patient risk profiles. Future research should consider these groupings and utilize both traditional and machine learning analysis models with additional granular clinical patient and center characteristics to delineate survival benefit and better inform ECMO practice guidelines.
Supplemental Material
Supplemental Material - Machine Learning Differentiates Extracorporeal Membrane Oxygenation Mortality Risk Profiles Among Trauma Patients
Supplemental Material for Machine Learning Differentiates Extracorporeal Membrane Oxygenation Mortality Risk Profiles Among Trauma Patients by Bryan R. Campbell, Alexandra S. Rooney, Andrea Krzyzaniak, Richard Y. Calvo, Kyle D. Checchi, Alyssa N. Carroll, Michael J. Sise, Vishal Bansal, and Michael J. Krzyzaniak in The American Surgeon™
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Supplemental material
Supplemental material for this article is available online.
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
