Science fiction or clinical reality: a review of the applications of artificial intelligence along the continuum of trauma care

Abstract

Artificial intelligence (AI) and machine learning describe a broad range of algorithm types that can be trained based on datasets to make predictions. The increasing sophistication of AI has created new opportunities to apply these algorithms within within trauma care. Our paper overviews the current uses of AI along the continuum of trauma care, including injury prediction, triage, emergency department volume, assessment, and outcomes. Starting at the point of injury, algorithms are being used to predict severity of motor vehicle crashes, which can help inform emergency responses. Once on the scene, AI can be used to help emergency services triage patients remotely in order to inform transfer location and urgency. For the receiving hospital, these tools can be used to predict trauma volumes in the emergency department to help allocate appropriate staffing. After patient arrival to hospital, these algorithms not only can help to predict injury severity, which can inform decision-making, but also predict patient outcomes to help trauma teams anticipate patient trajectory. Overall, these tools have the capability to transform trauma care. AI is still nascent within the trauma surgery sphere, but this body of the literature shows that this technology has vast potential. AI-based predictive tools in trauma need to be explored further through prospective trials and clinical validation of algorithms.

Background

The term artificial intelligence (AI) was first conceived in 1955 by John McCarthy as “the science and engineering of making intelligent machines” [1]. More colloquially, AI can be thought of as a broad term describing an algorithm that performs tasks that would normally require human intervention. Machine learning (ML) is a subtype of AI whereby these algorithms can improve their performance over time with additional experience [2]. There are many ways to classify ML algorithms, but one of the most popular ways is to group them into three main categories: supervised, unsupervised, and reinforced learning (Fig. 1). Supervised learning uses labeled inputs to produce a defined set of outputs of discrete values [2]. Examples of supervised learning includes decision trees, support vector machines (SVMs), regressions, and artificial neural networks (ANNs). Unsupervised learning creates groups from data whereby elements within each group are like each other but dissimilar to other groups; popular unsupervised algorithms include k-means clustering, singular value decomposition, and DBSCAN [2]. Finally, reinforcement learning is a technique that uses interactions with its environment to learn how to behave through trial and error; these include k-armed bandit, Markov decision processes, and SARSA [2].

Fig. 1

Overview of major types of machine learning. Overview of different types of machine learning (ML): ML is shown as a subset of artificial intelligence (AI). Within ML, there are three subtypes: supervised learning, unsupervised learning, and reinforced learning. Supervised learning is task-driven and uses labeled data to predict a predefined outcome. Unsupervised learning is data-driven and is used to find trends/outputs in unlabeled data. Reinforced learning is environment-driven and uses interaction with the environment to learn

For the purposes of trauma medicine, supervised learning algorithms have been studied the most, and therefore are the focus of this review. Supervised learning algorithms vary significantly in complexity (Fig. 2). More basic supervised ML algorithms are logistic regressions and decision trees [3]. These algorithms are interpretable (and, for decision trees, familiar clinically) but are lower in accuracy due to their relatively low flexibility. Algorithms capable of producing a range of forms include random forests and ANNs. These systems are complex and, while highly accurate, are less transparent to the user. Random forests are effectively an average of many permutations of decision trees made from a data set. ANNs emulate the connections within a brain, with inputs “synapsing” with multiple hidden layers (“interneurons”) via complex equations to deliver predictions [4]. Using AI in trauma requires balancing a model’s sophistication and complexity with transparency and usability.

Fig. 2

Overview of major types of supervised learning. Continuum of complexity of supervised learning algorithms: While not all types of supervised learning algorithms are shown here, four major illustrative examples—logistic regression, decision tree, random forest, and artificial neural network—are shown along a qualitative continuum from least to most complex. The diagrams are meant to provide a visualization of the algorithm processes whereby the blue circles and the orange squares represent different outcomes

The purpose of this review is to critically appraise and highlight the different applications of AI/ML that have been studied in trauma care in order to  provide clinicians, hospital administrators, and other non-computer scientist/non-technical audiences a basic understanding of AI/ML, the capabilities of these algorithms, and the potential ways that these may transform trauma care in the future. As injuries cause the greatest reduction in active life years globally and are the leading cause of death in people under 40 [5], ML has the potential to impact global public health through the optimization of processes and improvement of outcomes. With large-scale electronic health record implementation, an unprecedented volume of trauma data are available to train and validate new ML systems [6]. As such, trauma care is primed for AI-based transformation. This article follows the trauma patient’s journey—starting from the point of injury, through triage and arrival at the emergency department, to treatment and outcome prediction—to outline the utility of AI along the entire continuum of trauma care (Table 1) [7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53].

Table 1 Studies utilizing artificial intelligence in trauma surgery

Injury prediction

AI applications in trauma begin before injury. While trauma and emergency physicians use heuristics and patterns to predict when injuries are most likely to happen (e.g., date, time of day, weather [54]), these approaches lack sensitivity and adaptability. AI can help refine injury prediction.

Within injury prediction, motor vehicle crashes (MVCs) are the most studied. MVC studies can be further subdivided into crash occurrence and crash severity prediction studies. Crash occurrence models are complex and appear to be earlier in development. MVC occurrence prediction has been modeled using simulators that predict crash occurrence or non-occurrence based on car movement (e.g., gravity, drift angle) and environment (e.g., weather) [14]. Elamrani Abou Assad et al. were able to achieve prediction accuracies of 92.00% for an SVM and 93.34% for a multilayer perceptron (MLP), which is a type of ANN [14]. While these accuracy levels are very high, simulator studies have limited utility in healthcare as every road, car, and driver cannot be tracked for up-to-date predictions, and limited research has been done to apply this principle to populations due to the spatial and temporal complexities of this process. Bao et al. [12] is one of the few who has approached this task by using a spatiotemporal convolutional long short-term memory network to predict short-term crash risk at a weekly, daily, and hourly level within Manhattan using historical crashes, taxi GPS, road networks, land use, weather, and population. This model achieved a 71.02–99.21% prediction accuracy based on the spatial and temporal resolution used (with increasing accuracy at lower resolutions); however, tools like this may be impractical for prospective use due to the processing power and volume of data required [12].

Crash severity studies are more established and predict injury acuity from easily identifiable crash scene characteristics [7,8,9,10,11, 13, 15,16,17]. Inputs often include road (e.g., speed limit, surface type), vehicle/driver (e.g., driver/vehicle age, vehicle type), and environmental characteristics (e.g., weather, time) [7,8,9,10,11, 13, 15,16,17]. These algorithms are trained to predict injury level based on predefined categories; these are usually either basic, such as severe/not severe, or more complex, such as no injury/possible injury/non-incapacitating injury/incapacitating injury/fatality [7,8,9,10,11, 13, 15,16,17]. The vast majority of studies have focused on different types of ANNs, such as MLPs or deep neural networks due to their capability to handle highly complex data inputs, and have found AI can predict severity with an accuracy of 0–96% depending on the study [7,8,9,10,11, 13, 15,16,17]. This enormous variability is due to several factors, including differences in data input (such as geography and quantity/quality of crash data) but also the output type. Of the MVC crash severity studies examined, those that tried to predict a greater number of output categories tended to decrease their accuracy [7,8,9,10,11, 13, 15,16,17], as there was fewer representative data per category to train the algorithm. These studies suggest that MVC severity prediction could help prepare first responders and hospitals on likelihood of injury severity if the appropriate input, output, and algorithm type is selected. Other mechanisms of injury, such as interpersonal violence and self-harm are less developed in current injury prediction research.

Building and validating these predictive algorithms has clinical and public health applications. Clinically, an AI-based application could be used by 911 dispatchers based on caller information to predict injury severity and more accurately inform EMS prioritization and response. Systemically, understanding the variables associated with injury severity can support harm-reduction and injury prevention public health strategies [12, 14, 55].

Pre-hospital triage

Once injury occurs, AI can help triage patients before hospital arrival. Currently, remote triage takes time and relies on (1) EMS to contact hospitals when high-acuity patients are en route and (2) effective communication between the EMS team and the receiving physician. AI has been shown to predict the need for critical care/life-saving interventions to help stratify incoming trauma patients pre-hospital both generally [17, 19,20,21] and in specific trauma subtypes, such as gunshot wounds and after resuscitation [22, 33]. The ability to predict the need for life-saving interventions can help inform hospital selection, allowing EMS to route to hospitals with the capacity to handle the necessary care for their patient. This could be especially useful in rural and remote settings where decisions must be made about air evacuation. Further, more detailed information and predictions about patients en route to hospital could help receiving centers prepare for the upcoming trauma activation, such as through allocating appropriate resource/operating rooms or ensuring available staff.

Algorithm inputs range in complexity from 6–8 inputs mostly comprising vitals, such as in the case of Liu et al. and Kim et al. [20, 21], to more complex analyses that consider time to dispatch, basic laboratories, and injury characteristics [18, 19, 22]. Almost all the studies that were examined for this paper employed types of ANNs to elucidate this relationship, and studies that used a greater number of variables often (but not always) had greater accuracy (AUC 0.82–0.912) as compared to those with fewer input variables (AUC 0.71–0.88) [18,19,20,21,22].

Remote triage systems may be efficient if they require minimal data input by the EMS team and can be used to ensure appropriate resources available for patients on arrival to the hospital. However, remote triage applicability may be limited as there is a trade-off between increased accuracy and necessary data volume. There may be a threshold where the inconvenience to EMS of managing high data volume surpasses the relative accuracy increase in triage. Kim et al. [20] used data that could be collected on wearable devices—including systolic blood pressure, heart rate, respiratory rate, and a modified consciousness score—in addition to patient age to predict patient likelihood of death. Using an ANN, they were able to achieve an AUC of 0.89 with this method, showing highly accurate prediction with minimal human intervention [20]. The clinical meaning of a difference between an AUC of 0.89 based on this minimal-input wearable and 0.912 based on an intensive high-input algorithm would need to be elucidated in future study; however, the use of wearable sensors appears promising and may allow for dynamic prediction.

Emergency department volumes

AI has been shown to predict trauma volumes within the emergency department (ED) [23,24,25,26]. Inputs reflecting human activity and environmental conditions such as date, traffic, special events, precipitation, temperature, and air quality, are the basis of these algorithms have been used in previous studies [23,24,25,26]. Like algorithms predicting crash occurrence in MVCs, predicting ED volumes relies on ANNs to capture the complexity of the relationship between large-scale patterns of human behavior and individual center-level outputs. Further, unlike in previously discussed algorithms in this paper that use AI to predict categorical variables, these algorithms are tasked with continuous outputs.

Unfortunately, differences in data reporting/statistical analysis in the current literature makes cross-comparison between studies difficult. Stonko et al. [26] and Dennis et al. [23] used correlation coefficients to show that ANNs could be used to predict mean ISS, total number of traumas, and number of penetrating traumas in a given day with a correlation of 0.87–0.89. Menke et al. [24] and Rauch et al. [25] used deviations from the true value/average error to show efficacy, showing the predicted ED volume falling within 20 visits of the true volume 95% of the time and a mean average error of 2.32–3.25 patients, respectively. Overall, these systems show lower accuracy at extreme ends of the spectrum (very low- and high-volume days). Each study requires more robust statistical analysis to show accuracy, and thus, the ability to draw conclusions about the future applications in trauma is limited; however, based on information available, they appear to be able to predict volume and acuity on average, which has important implications for trauma care optimization.

Better prediction of trauma volumes can not only improve resourcing for cost savings to the healthcare system but can also lead to better patient outcomes when there is appropriate capacity available to treat each case. However, more investigation will be needed about the adaptability of these algorithms to shifting patterns in human behavior in the wake of the COVID-19 pandemic.

Initial assessment

Once a patient arrives at the hospital, AI can support initial diagnostic and therapeutic decision-making through patient severity assessments. Predicting patient severity at presentation is a broad category of algorithms that includes prediction of prognosis and decision support.

Predictive analytics

Prediction of prognosis can help differentiate patients who are critically ill versus those who are not and identify those who will need life-saving interventions. In practice, an AI-based tool in this setting would use easily assessed variables (e.g., vitals, GCS) inputted by the receiving trauma team to initially determine patient prognosis and need for interventions. Studies have taken dramatically different approaches to addressing this task. Liu et al. [32] developed an MLP that used vitals, demographics, and GCS to determine the need for life-saving interventions at ED presentation with an achieved AUC of 0.99. Although limited by small sample size, the algorithm’s performance is promising for further development. Batchinksy et al. [27] used ECG data alone to determine the need for life-saving intervention with an AUC of 0.86. Importantly, in both cases, these variables are readily available in the ED, making these algorithms amenable to use in high-stakes presentations.

Diagnostics and clinical decision support

Concomitant with determining need for life-saving intervention, clinicians in the ED often need to investigate and rule out injuries on CT, such as cervical spine injury (CSI). Extensive work has gone into creating guidelines to aid physicians in determining when imaging is needed as part of CSI workup, [56]. Despite evidence-based guidelines, imaging is often inappropriately used, with clinically relevant injury found in just 2% of imaged patients [56]. Bektas et al. [28] compared a logistic regression with an ANN to supplement CT in detecting CSI. The ANN had a significantly better negative predictive value than the logistic model at 97.3% versus 87.9%, respectively [28]. The ANN also had a positive predictive value of 100% and detected 2 CSIs that were missed on CT alone [28]. Furthermore, AI can support decision-making in pediatrics where imaging over-use is of greater concern due to carcinogenic irradiation [29]. Using GCS, age, gender, and injury mechanics, an optimal classification tree algorithm predicted CSI in patients < 3 years old with a 93.38% sensitivity and 82.34% specificity [29]. Other studies have demonstrated that AI can assess pelvic hematoma on CT imaging [31]. Volume of pelvic free fluid is used clinically to predict the need for transfusion and angioembolization yet is challenging and time-consuming to quantify on CT [31]. Dreizin et al. [31] developed AI capable of segmenting these CTs to produce reliable volume measurements; this algorithm had results on par to physician judgment with much less time and effort investment, with AUC of 0.81 as compared to an AUC of 0.80 when manually done by radiologists.

AI can supplement other imaging in trauma workup [30, 31, 52]. Ultrasound, while indispensable in trauma evaluation, is limited by its wide sensitivity range (28–100%) [30]. Cheng et al. [30] designed a model that interprets free fluid in Morison’s pouch during FAST exams after torso trauma. The model was trained to not only to detect free fluid, which it did with 96.7% accuracy, but to also determine if the image captured on ultrasound was qualified to make such predictions, which it could determine with 94.1% accuracy. Further studies have used ML to accelerate workup, improve diagnostic accuracy [31], and reduce unnecessary imaging [28, 29].

AI has a potential role in imaging workup in situations where clinicians must rapidly interpret imaging to inform patient management. In high-acuity settings where time is of the essence, these algorithms could evaluate images faster than and with equal or superior accuracy to human review, allowing for the identification of pathology more rapidly and precisely.

Outcomes

Trauma patients are a heterogeneous group at high risk of complications, including but not limited to organ failure, cardiac arrest, infection, respiratory distress, shock, stroke, and death [40, 45]. Owing to their heterogenicity and rapidly changing status, it remains challenging for physicians to predict a clinical course for these patients in hospital. Numerous non-AI-based risk prediction tools exist for these complications and outcomes, but these tools lack the ability to intelligently adjust the weight of input variables and instead are linear and additive [40]. As such, it is unsurprising that much of the body of the literature around AI in trauma is centered on intervention and outcome prediction. Within intervention and outcome prediction, there are three main types of algorithms: complication prediction, survival prediction, and discharge prediction.

Complication prediction

ML has been studied to assess its ability to perform risk prediction and accurate prognostication of clinical outcomes in trauma patients [37, 40, 45, 48]. An ideal AI-based tool for complication prediction would either (1) use variables that are readily available to the trauma team after a brief workup (i.e., vitals, comorbidities, injury factors/TRISS scores ± laboratory results) to help identify which complications a patient is most at risk of and which interventions have the greatest possibility of mitigating these complications or (2) use variables that are collected post-intervention (i.e., using all the same variables above but with invention-related inputs) to identify likely downstream postsurgical complications. While complication prediction would have high clinical utility, this type of prediction is technically challenging due to the many potential complications these patients can encounter. To predict an output, there needs to be sufficient examples of that output in the training dataset to determine the relationship between the inputs and the output. With a perfect dataset and infinite computing power, all outputs could be predicted with equal accuracy; however, in practice, data limitations can result in wildly variable capabilities to predict individual outputs within a single algorithm. Christie et al. [37], for example, looked at 7 complications and was able to predict their occurrence with an AUC 0.45–0.74. Maurer et al. [45] looked at 11 complications and achieved c-statistics of 0.689–0.835. As such, individual algorithms that have been developed may be able to accurately predict certain complications but may not be reliable at predicting the full suite of complications that may befall a patient.

Further, with the ever-changing condition of trauma patients, the ideal risk calculator could dynamically alter predictions in real time and identify modifiable factors to change outcomes [37]. Christie et al. [37] designed the “SuperLearner,” an algorithm that incorporates data across time and re-evaluates mortality and complication risk. Although, SuperLearner’s variability in prediction by complication means it may not yet be ready for clinical application, tools that can dynamically adjust predictions over time would have exciting applications in trauma care [37]. In the future, machine learning merged with causal inference methods may be able to predict which treatment would provide the best outcome and could be the basis of precision medicine in trauma.

Survival prediction

Within outcome prediction, survival prediction is by far the most studied. As such, there is diversity in algorithm choice, input variables, and prediction accuracies within this space. The most basic of these algorithms use inputs such as comorbidities, demographics, GCS, vitals, and injury data [36, 41, 50, 53]. While studies use several different algorithms, they are able to consistently achieve accuracy levels of > 89% with some as high as 97% [36, 41, 50, 53]. As these algorithms become more complicated, they also incorporate laboratory results, imaging findings, currently available scoring systems (e.g., TRISS), and interventions [35, 38, 39, 44, 49, 51]. However, increasing the input complexity does not always increase accuracy. These models are consistently able to predict at accuracies of > 82% and as high as 98% [35, 38, 39, 44, 49, 51]. The small gains in accuracy (and in some studies, a drop in accuracy) relative to simpler models may be due to overfitting, whereby increasing the number and specificity of input variables creates an algorithm that is perfectly trained to predict based on the training dataset but is unable to generalize to new datasets.

The trauma outcome predictor (TOP) is one such algorithm that has been validated to predict mortality, as well as 9 other complications [45, 57]. It uses data such as demographics, vital signs, mental status, comorbidities, and injury characteristics to feed an optimal classification tree algorithm, which can predict mortality and morbidity with c-statistics up to 0.941 [45]. TOP is an excellent example of what a survival prediction tool would look like in clinical practice, adjusting the necessary input questions based on previous answers in order to predict mortality and morbidity (Fig. 3) [57]. Tools such as TOP could be used by clinicians to assess survival risk in order to plan the intervention and management of patients, as well as inform palliation and end of life discussions.

Fig. 3

Trauma outcome predictor (TOP) example screen shot. This screenshot of the TOP interface shows how clinicians can input variables based on clinical assessment to predict mortality after blunt injury. The differences between the left and right panels are due to the algorithm’s ability to adjust the questions asked based on answers to previous questions; in this case, the differences in GCS answers prompt the algorithm to diverge in its input variable requirements. Reprinted from Surgery, Vol 171/6, El Hechi M, Gebran A, Bouardi HT, Maurer LR, El Moheb M, Zhou D et al. Validation of the artificial intelligence-based trauma outcomes predictor (TOP) in patients 65 years and older, Page 1689., Copyright (2022) with permission from Elsevier and the original authors

Discharge prediction

Understanding discharge disposition and length of stay can help health systems prioritize bed allocation, begin discharge planning in advance, and set realistic expectations for patients and families post-injury. This is especially important post-traumatic brain injury (TBI) where long-term effects can be unpredictable. Pang et al. [47] and Nourelahi et al. [46] compared the efficacy of several algorithm types to predict Glasgow Outcome Scoring for patients’ post-TBI. Both used demographics, GCS, and pupillary responses (with some additional unique inputs per study) and were able to achieve accuracy of 63–78% [46, 47]. Compared to other clinical applications of AI in trauma discussed in this paper, TBI outcome appears to be of lower accuracy, likely due to the high variability in patient recovery post-injury. Length of stay is also a burgeoning area of discharge prediction studies. Staziaki et al. [52] and Ji et al. [43] used several algorithms, including SVM and ANN, to predict duration of hospitalization. Both papers tested a variety of combinations of input variables, including but not limited to demographics, GCS, vitals, and injury scoring, to achieve accuracy levels of 58–79% for SVM and 65–79% for ANN [43, 52].

Speech interpretation: supporting the trauma care continuum

Narrative clinical documentation in trauma is often difficult to analyze in real time as it is not entered as discrete and time stamped data elements, which is critical for clinically relevant algorithms. Natural language processing (NLP) and automatic speech recognition (ASR) are two fields of AI that can relieve the burden and time of converting speech notes to text and provide higher-quality data input. Research by Blackley et al. [58] found that speech recognition saves time, increases efficiency, and allows for quicker and more relevant documentation. AI related to speech/audio can also help diagnose pneumonia, asthma, and other infections. For example, investigators using cough data [59] achieved 100% asymptomatic COVID-19 detection rate and 88% accuracy on all subjects. Converting narrative data to structured data using NLP/ASR would be potentially transformative in fast-paced and data-rich trauma resuscitation environments, where critical decisions are often made without integration of all available information.

Conclusions

AI in trauma surgery has numerous applications and proven efficacy. As more studies validate new or existing algorithms, trauma analytics are likely to shift away from rudimentary scoring methods toward more dynamic and accurate AI decision support tools. These tools are applicable from the point of injury through to surgical follow-up.

In order to begin to fulfill the potential of AI, trauma systems must adapt  compatible electronic health records and reporting systems to support real-time data collection and integration. Existing AI systems must be evaluated prospectively to demonstrate replicability as compared to algorithms trained on retrospective data. More ML systems must be able to dynamically adjust their predictions as patient status changes. Algorithms need to be paired with interpretable graphical user interfaces so that they can be used by clinicians and not just computer scientists.

AI has a promising role within trauma surgery practice and is worth the time and investment needed to prove and establish its specific uses. Given the technical expertise required to design, evaluate, and validate these algorithms, this endeavor will require interdisciplinary collaboration between physicians, computer scientists, statisticians, and administrators. These tools have the promise of changing clinical practice and improving patient outcomes and population health.