Audio Coursesby JellypodLesson 02 of 15
Overview
Explore how artificial intelligence is reshaping trauma resuscitation, from haemorrhage prediction to workflow optimisation. Jeremy and Hamish break down recent evidence, highlight real-world clinical challenges, and discuss the future of equitable, context-aware AI systems in critical care.
Jeremy: Welcome back to the TIME Podcast. I’m Jeremy...
Hamish: …and I’m Hamish. Today, we're going to take a deep dive into the supplementary manuscript for our TIME session on AI in trauma resuscitation. We’ll explore the eight key studies that make up the current evidence base for AI in this space and break down what they actually mean for you, the clinicians, working in high-pressure, time-sensitive environments. Before we dive in, a big thanks to Clintix Pro, who’s making this conversation possible. Clintix Pro is an AI study companion designed to support doctors as they prepare for exams, and we’re grateful for their help.
Jeremy: Let's jump right in
Hamish: Trauma resuscitation is one of the most demanding clinical environments you can find yourself in. When you’re in the trauma bay, decisions need to be made quickly, often with limited information. Patients are unstable, and their condition can change in minutes, even seconds.
Jeremy: Exactly, and while trauma teams are incredible at recognizing patterns, reading the room, and making quick decisions, there's only so much that the human brain can do when you're juggling multiple data streams and trying to track subtle changes in a patient's condition. That's where AI comes in—it could help reduce some of the cognitive load on clinicians.
Hamish: The review we're looking at today focuses on three main areas: haemorrhage prediction, procedural guidance, and workflow optimisation. These are the key places where AI could potentially make a real difference.
Jeremy: But here's the crucial question we need to ask—does it work in real-world trauma settings? Not just in theory or in controlled studies, but actually in the moment, when timing and accuracy are everything.
Jeremy: Now, let’s talk about how this review was conducted. It used the Joanna Briggs Institute (JBI) scoping review methodology, which is really important. One of the main issues with AI research in trauma is that much of it focuses on how well the models work in a lab setting or under controlled conditions, but the real question is: are they clinically valid?
Hamish: Right. This review didn’t just focus on the theoretical accuracy of the models. It critically appraised the studies using tools that focus on practical, real-world factors—things that clinicians care about. For example, were the reference standards appropriate? Did the model use data that was available when clinicians actually needed it? Were confounders adequately addressed?
Jeremy: And most importantly, did the outcomes matter to clinicians? This is something that often gets lost in the rush to talk about accuracy, but it’s crucial. It’s not enough to show that the model is accurate on paper; we need to know if it actually helps clinicians make better decisions, avoid errors, or improve patient outcomes in real-time.
Hamish: Let’s start with haemorrhage prediction, which is where much of the excitement is focused. One of the strongest studies in this review is Gauss et al. (2025). It’s a prospective, multicentre study, which already gives it solid credibility. They looked at early clinical data—things like vitals, labs, and the trauma mechanism.
Jeremy: And something that's particularly impressive about this study is that the model was calibrated. If the AI model tells you there’s a 40% chance that a patient will need haemorrhage control, that 40% needs to hold true across a wide range of patients, not just in one small group.
Hamish: Yes, calibration is key here. Clinicians know that they need trustworthy risk estimates, not just predictions that look good in isolation. However, even in this strong study, the outcome it’s predicting is a process measure—activation of haemorrhage control interventions. It’s not predicting the actual underlying physiological need, which is the crux of the issue.
Jeremy: So while this study is methodologically solid, it’s still more about predicting what clinicians decide to do rather than what’s actually going on inside the patient’s body. That’s something to keep in mind when looking at the results.
Hamish: Now, moving on to Zhang et al. (2024). This study shows a lot of promise on paper—good AUROC and a broad dataset. But as we dig into the details, there’s a significant problem with temporal leakage.
Jeremy: That’s right. Temporal leakage means the model is using data points that were only available after decisions were made—like lactate levels that were drawn after a massive transfusion protocol was activated. If the data used by the model wasn’t available at the decision-making moment, the prediction is more about describing what happened rather than predicting what will happen.
Hamish: Exactly. Trauma registries record clinician behaviour, not the actual underlying physiology, which means that AI models trained on this data may learn to reflect the patterns of clinician actions rather than the reality of the patient’s condition. So, while the model performs well in retrospective analysis, it doesn’t necessarily hold up in a real clinical scenario.
Jeremy: Wangoo et al. (2025) provides a useful reality check. They tested both traditional and AI-enhanced massive transfusion prediction tools and found that calibration was poor across both. This is exactly what we expect given the dynamic nature of haemorrhage.
Hamish: Exactly. Haemorrhage isn’t static. Trauma patients have different compensatory mechanisms based on their age, injury type, and overall physiology. Predicting transfusion needs with static thresholds has always been a challenge, even for human clinicians.
Jeremy: So, the takeaway here is clear: haemorrhage prediction is still a work in progress. AI hasn’t solved this problem yet, and honestly, neither have we as clinicians.
Hamish: Now let’s shift focus to procedural AI. Cheng et al. (2021) looked at AI-assisted FAST interpretation, particularly for detecting free fluid in the abdomen. They found that the AI significantly improved sensitivity, especially for novice operators.
Jeremy: FAST exams are incredibly operator-dependent, especially in unstable patients. If AI can help standardise those difficult views, that could make a huge difference, not just in the speed of decision-making but in the quality of the images we get.
Hamish: The real question is: Does it translate into faster or better clinical decisions? Getting patients to CT or theatre more quickly could be a game-changer, but we don’t have that data just yet. It’s something we’ll need to see in further studies.
Jeremy: Hartline et al. (2025) takes a slightly different approach. Instead of helping interpret the images, the AI is guiding the operator to acquire the correct view in the first place.
Hamish: And this is really interesting because the study showed clear improvements in image quality and the accuracy of landmarks, which could be a huge help in smaller centres or during off-hours when senior staff aren’t available to guide novice operators.
Jeremy: But again, we’re talking about simulations here. Trauma patients don’t lie still, and their physiology is unpredictable. Real-world validation is needed before we can draw firm conclusions.
Hamish: One of the most relevant studies here is Fitzgerald et al. (2025), where AI was deployed directly inside a trauma bay to provide real-time prompts to clinicians. This is one of the few studies that really looks at how AI can work in a live trauma environment.
Jeremy: And the results were really impressive. The AI helped reduce resuscitation errors, particularly omissions, which are common when trauma teams are overwhelmed and under pressure.
Hamish: This is where AI has real potential—by stabilising team performance and reducing cognitive overload, it helps clinicians stay focused on the most important tasks.
Jeremy: Li et al. (2025) takes a unique approach by looking at team dynamics. Instead of focusing on the patient, the AI here was designed to recognise what phase of resuscitation the team was in, based on their actions.
Hamish: What Li et al. did was to use AI to model and classify what phase the trauma team was in during resuscitation. They used deep learning to analyse video footage of trauma resuscitations and looked at team behaviours—how the team interacted, what actions they were taking, and what was happening at that particular point in the resuscitation process.
Jeremy: This is fascinating because it’s a completely new way of looking at AI in trauma care. Rather than just predicting patient outcomes, this model focuses on recognising what the team is doing. It aims to understand the context of the situation—what’s happening in the room, what decisions the team has already made, and what needs to happen next.
Hamish: The idea is that AI could provide real-time contextual support, recognising what phase of resuscitation the team is in and offering relevant prompts or suggestions. It’s not about making clinical decisions, but about helping the team stay aligned with the patient’s evolving needs.
Jeremy: That’s the key. It’s about enhancing situational awareness for the team. AI could essentially be the extra set of eyes, helping to ensure that no step is missed and that the team is functioning as efficiently as possible.
Hamish: However, the study was retrospective, and that’s an important distinction. The model was trained on past trauma cases, so it’s not yet proven in real-time clinical scenarios. There’s no evidence yet to say how this would work when the pressure is really on, with a patient actively coding or crashing in front of the team.
Jeremy: Exactly. The potential is there, but like most studies we’ve seen, real-world validation is the next step. Could this type of contextual AI really improve team dynamics in the trauma bay? That’s the big question.
Hamish: Moving on to Zia et al. (2022), this study looked at the role of AI in triaging CT scans, specifically for suspected intracranial haemorrhage. Now, trauma bays see a lot of patients who need head CTs, particularly those with head injuries where intracranial haemorrhage is a concern. Zia et al. focused on how AI can speed up the process and improve diagnostic accuracy.
Jeremy: The AI model in this study was designed to flag suspected intracranial haemorrhage on CT scans, helping radiologists prioritise which scans need urgent review. The results were promising—the model demonstrated high diagnostic accuracy and helped reduce reporting times.
Hamish: And that’s where the real value lies—particularly in a busy emergency department or trauma centre. Faster reporting means quicker decisions, and quicker decisions can make a huge difference in patient outcomes, especially when you’re dealing with brain injuries where time really is of the essence.
Jeremy: Absolutely. Zia’s study didn’t just focus on accuracy; it also looked at how the AI could optimise the workflow. In a high-volume ED, when you have a steady stream of patients with CT scans piling up, every second counts. The AI doesn’t interfere with clinical reasoning; instead, it helps improve efficiency by highlighting cases that require more urgent attention.
Hamish: Right. And this is a perfect example of AI enhancing the clinical process without trying to replace clinician judgment. The AI system isn’t making decisions for the radiologists or clinicians, it’s just helping them prioritise and flag the critical cases faster.
Jeremy: That said, while the study shows great promise in terms of improving workflow, there are still questions around how this would play out in a real-world clinical setting. The study was observational, and it wasn’t clear whether the AI intervention actually led to improved patient outcomes—whether it helped save lives or improve functional outcomes in the long run.
Hamish: Exactly. The diagnostic accuracy is great, but the real question is whether this translates into better clinical decisions and outcomes for the patients. It’s one thing to speed up the process, but we need to see the impact on patient recovery or survival before we can fully embrace it.
Jeremy: So while this is a step in the right direction, much like the other studies, the key next step will be validating these models in real-world settings and measuring their true clinical impact.
Hamish: When you step back and look at these studies together, there are a few important lessons we can take away. The first is about reference standards. Many of these studies relied on practice-derived standards—things like massive transfusion activation or expert consensus—which are useful but don’t always capture the true physiological processes at play.
Jeremy: Exactly. If the reference standard is based on what clinicians are already doing, then the AI is just learning those behaviours. It’s not necessarily learning to predict the physiological events that the clinicians are trying to respond to. So, while these tools look good in terms of performance, they might not be addressing the core issue in trauma care.
Hamish: The second lesson is temporal alignment. Trauma resuscitations are fast-paced, and every second counts. If the model uses data that’s not available when a decision needs to be made, then it’s not truly reflective of what happens in real time.
Jeremy: Right. If a model is trained on data points that only come through after a decision has already been made, then it’s not testing the real-world application of those decisions. It's just saying, "This is what happened after the fact." And that’s not helpful in a clinical scenario.
Hamish: The third takeaway is that most of these studies focus on process outcomes, like diagnostic accuracy or error reduction. But what’s missing is a true measure of clinical impact—things like mortality or long-term patient outcomes.
Jeremy: That’s the big gap. We need to know if these tools actually improve survival rates, functional outcomes, or even shorten ICU stays. Until we have that data, we can’t say for sure that AI is making a meaningful difference in patient outcomes.
Jeremy: Now, let’s address a critical issue that hasn’t been discussed enough in the current AI trauma literature: equity and real-world translation. All of the studies we reviewed come from well-resourced, urban trauma centres. The AI tools being tested were implemented in settings with access to advanced resources—high-end imaging, dedicated trauma teams, and established clinical pathways.
Hamish: That’s right, and while these urban centres offer some of the best clinical environments for AI testing, it doesn’t mean that these tools will automatically work everywhere. Trauma care in rural or remote areas, or in low-resource settings, operates differently. They often lack the same technological infrastructure, and healthcare teams may face challenges that aren't present in larger, more resource-rich hospitals.
Jeremy: Exactly. One of the biggest questions for the future of trauma AI is: will these systems perform as well in diverse, real-world settings? If AI tools are only validated in a handful of urban, well-resourced trauma centres, we can’t assume they’ll be equally effective in hospitals with fewer resources. What works in a state-of-the-art trauma bay might not work in a rural or underserved hospital.
Hamish: That’s an excellent point. We need AI tools that are tested and adapted for the full spectrum of healthcare environments. In rural or resource-limited settings, there could be issues with connectivity, access to the latest imaging technology, or even having enough staff on hand to properly use and monitor the AI tools.
Jeremy: Right, and this becomes even more complicated when you consider diverse populations. Many of the studies we reviewed did not assess AI performance across different cultural or socioeconomic groups. Without this data, we can’t be certain that the AI is not introducing biases that could affect underrepresented or marginalized groups differently.
Hamish: Yes, bias in AI is a huge issue. If these models are trained mostly on data from certain demographic groups, say, predominantly white, urban populations, they might not work as effectively—or as safely—on patients from different ethnicities, age groups, or socio-economic backgrounds. That’s something that hasn’t been addressed in most of the trauma AI studies we’re seeing right now.
Jeremy: And let’s not forget algorithmic transparency. Many AI systems are designed as "black boxes," meaning we don’t always know how they’re making their decisions. In critical care, where human lives are on the line, it’s important that these systems are transparent, especially when they’re being implemented in diverse healthcare environments. Clinicians need to understand the logic behind AI recommendations, not just trust them blindly.
Hamish: Exactly. And without the proper transparency and understanding, clinicians may hesitate to use AI in their decision-making or, worse, use it inappropriately, potentially causing harm. So, AI in trauma care not only needs to be validated in a variety of settings but also needs to be culturally competent and transparent to be truly effective across all populations.
Jeremy: We need studies that specifically test AI in diverse environments—rural, urban, under-resourced, and across a variety of patient demographics. And those studies should include a robust evaluation of algorithmic fairness, so we can be sure that AI isn’t inadvertently causing harm by perpetuating existing biases or disparities in care.
Hamish: This is where the future of trauma AI lies—creating systems that are adaptable, transparent, and equitable. If AI is going to have the kind of impact we all want it to, it needs to work for everyone, not just a select group of patients or hospitals. It needs to be tested, refined, and validated across all types of healthcare settings, with a strong emphasis on patient safety, equity, and fairness.
Jeremy: Looking to the future, the real promise of AI in trauma care lies in its ability to integrate various data points in real-time—what we call context-aware AI. Imagine a system that doesn't just react to one stream of data, like vital signs or imaging, but integrates all available information to understand the full context of a resuscitation. It could look at a patient's vital signs, their imaging results, and even the team’s behaviour to make more accurate predictions about what should happen next.
Hamish: That’s right. AI systems that understand the complete context of a trauma resuscitation could really make a difference. They wouldn’t just look at a snapshot of the patient’s condition—they’d look at the patient’s evolving status, and how the team is interacting with them. It could alert clinicians when they're missing a critical piece of information, or suggest next steps based on the patient's trajectory.
Jeremy: We’re also likely to see more anticipatory AI, systems that can predict the next step in the resuscitation process before it's needed. If the AI can understand what’s likely to happen next, it can prompt the team to take action before things become critical. This could be especially useful in high-stakes scenarios like trauma where the pace is rapid and the margin for error is slim.
Hamish: But, as you mentioned earlier, for this kind of AI to be successful, we need multicentre trials. We need to see real-world validation, not just in a single institution or population but across different settings. This is critical. AI needs to be able to work in the dynamic, unpredictable environments of trauma care.
Jeremy: And let’s not forget governance and regulatory frameworks. For AI to be truly integrated into trauma care, it needs strong oversight. We need to establish guidelines for when and how AI tools should be used, as well as how they should be updated. Without these frameworks, AI could easily add complexity and confusion rather than improving care.
Hamish: Exactly. For AI to be a tool that genuinely supports clinicians, it has to be seamlessly integrated into the clinical workflow. We need systems that don’t overwhelm the clinician with information, but instead provide the right information at the right time. The AI has to fit naturally into existing workflows, and not add another layer of cognitive load.
Jeremy: So, what should clinicians take away from all of this? First and foremost, AI is a tool that can help with discrete tasks in trauma care. It’s useful for things like improving image interpretation, optimising workflows, and reducing errors, especially under pressure. But it's important to remember that AI is not a substitute for clinical expertise. It’s not replacing clinicians—it’s supporting them.
Hamish: Exactly. While AI can handle certain repetitive or data-heavy tasks, it can’t replace the expert judgement and critical thinking that clinicians bring to the table. Trauma resuscitation is complex, and it requires the intuition and experience of clinicians to make the best decisions. AI can provide insights, but it can’t—and shouldn’t—be relied upon to make the final call.
Jeremy: AI is also a stabiliser in chaotic, high-pressure environments. It can help clinicians stay on track, reduce errors, and ensure that nothing important gets missed. It’s there to augment human performance, not replace it.
Hamish: Clinicians should also be aware that AI is still in its early stages in trauma care. While the potential is there, we don’t yet have enough evidence to say that AI tools can consistently improve outcomes in real-world trauma settings. The technology is promising, but it’s still evolving.
Jeremy: And as we’ve seen, AI tools need to be validated in diverse settings, tested across different populations, and evaluated for equity and fairness. Clinicians should stay informed about how these tools evolve and be ready to critically assess their usefulness as new evidence emerges.
Hamish: In the end, the most important thing is to remember that AI is a tool, not a magic solution. It’s there to help clinicians provide the best care possible in the most efficient way possible, but it’s not a replacement for expertise, judgment, and human interaction
Jeremy: Exactly. AI is about enhancing the clinical decision-making process, not taking it over. Clinicians need to continue to lead in the decision-making process, using AI as an extra set of eyes and a source of insight when needed.
Jeremy: That’s about us for today. We’ve covered a lot, from shiny algorithms to messy realities. If you liked what you heard, subscribe and stay tuned—there’s plenty more coming your way, including episodes on AI in sepsis and digital triage tools. Hamish, it’s always a pleasure talking this stuff through with you.
Hamish: Cheers Jeremy, you too. And thanks to everyone for listening—keep sending in your questions and experiences with tech on the frontlines, and we’ll see you soon on the next TIME podcast.
Jeremy: Take care, folks.
