Literature Review: May 2026

NTSP Literature Review Podcast: May 2026 (S7, Ep5)

This months papers cover:

• Assessing swallowing with electronic stethoscopes
• Can Pro-BNP influence timing of tracheostomy?
• Weaning programs
• LLMs for paediatric tracheostomy
• Training and testing LLMs for tracheostomy care

Link to supporting PDF: https://tracheostomy.org.uk/Podcast-Resources

The UK National Tracheostomy Safety Project (NTSP) is committed to providing education, information and resources to improve patient safety and the patient experience for those with tracheostomies and laryngectomies. All of our resources are housed on our website www.tracheostomy.org.uk, accessed by over 30,000 visitors each month from around the world.

This is the only podcast to bring you literature reviews, hot topic discussions and interviews with healthcare staff, patients and families.

Our goal is to improve the safety and quality of care for patients with tracheostomies and laryngectomies through education. We work closely with patients, families and healthcare professionals to develop new resources to improve care. We’ve collaborated with key stakeholders in tracheostomy care since 2009, and developed freely accessible resources, supported by online learning developed with the UK Department of Health. We’ve worked with the Global Tracheostomy Collaborative since 2012 to improve care for patients and their families everywhere.

We are funded by grants, donations and in partnership with medical device companies through unrestricted awards. This podcast series is supported by unrestricted education funding from the Atos Learning Institute. The funding supports the professional production of the podcasts and videos, and the medical device companies that support us do not have any creative influence over the content that we record. All of our work is undertaken by volunteer healthcare staff, patients and their families. 

Most of our content is supported by videos. You can access our training videos and resources for Basic Care, Emergency Care and Vocalisation & Swallowing. Download and print bedhead signs and emergency algorithms from our resources. 

You can support our work by watching or clicking any of the advertising links that appear via the NTSP YouTube Channel. You can also donate directly to the NTSP through the NTSP website, or by clicking the Buzzsprout podcast hosting "support" links. You can support our work by watching or clicking any of the advertising links that appear via the NTSP YouTube Channel.

Transcript

SPEAKER_00 0:07

Hello and welcome to Trachee Talk from the National Tracheostomy Safety Project. My name is Brenda McGrath and it's May 2026. Thanks for joining us for another monthly look at some of the recent literature around tracheostomy care, airway management, swallowing, rehabilitation and critical care recovery. There's a strong theme this month running through several of the papers, the role of technology and artificial intelligence in tracheostomy care. And as ever, we should approach this with a degree of caution, but that's why we're here to critically appraise the literature that we've got. But definitely some thought-provoking ideas around how technology might support assessment, education and decision making in the future. We've also got a few papers looking at prolonged ventilatory weaning and predictors of weaning failure, and those are topics that remain very relevant for the multidisciplinary teams who we have looking after our complex patients. So today specifically we've got a pilot study looking at neck-worn electronic stethoscopes to monitor spontaneous swallowing frequency. We've got a large retrospective study looking at something called ProBMP as a biomarker for ventilatory wheeling failure, a randomised controlled trial comparing two prolonged wheeling strategies, and then two papers exploring artificial intelligence in tracheostomy education and caregiver support. So, let's get started. Our first paper is from Matsumoto and colleagues, published in the Journal of Clinical Medicine. This is the one looking at monitoring of spontaneous swallow after tracheostomy using a neck-worn electronic stethoscope. The group looked at 18 tracheostomised patients undergoing dysphagia assessments, and they use this neck-worn stethoscope combined with artificially intelligent assisted analysis to measure spontaneous swallowing frequency over a 10-minute period. Now, spontaneous swallow frequency is something that's increasingly recognised as potentially important, and reduced spontaneous swallowing might reflect impaired airway protection, it might reflect bad secretion management or may be influenced by the severity of any dysphagia. So the interesting thing here is that the technology attempts to provide a more objective and then potentially scalable and reproducible monitoring of swallowing activity. The team found several clinically intuitive associations. So a spontaneous frequency of swallowing increased when cuffs on tracheosomissues were deflated compared to inflated. And that ties in with a lot of the work we've described and discussed on this podcast. We know that restoring upper airway flow restores normal physiological function to the upper airway and to the swallowing reflexes. And so this is something we see all the time in our practice when you take the cuffs down, things start to work again. And so, you know, that this wasn't a surprise, but it's nice to see it reinforced in the literature. Spontaneous swallowing frequency was also higher in patients with better oral intake scores, and they use something called the functional oral intake scale. But most interestingly, there was a moderate to strong negative correlation between spontaneous swallowing frequency and pharyngeal saliva retention when they looked on fees. So, in simple terms, that means that patients who were swallowing more frequently they tended to have less pooled secretion at the back of their throat. So that's something that you would expect to find on fees, and that was a strong sort of finding suggesting that the swallowing that was going on was actually impactful and doing some job. So none of that will be a surprise to experienced swallow clinicians, but what might matter here is this attempt to create this reproducible monitoring tool. And one of the challenges in tracheosometrophage and management is that so much of our assessment is intermittent, it requires a person or persons who may not be the same people each time making that same assessment. So whilst FEES gives us very rich detailed information, it's only a snapshot and a moment in time. And what this paper hints at is the possibility of longer duration physiological monitoring, which is operator independent, which means it should be consistent. So the key questions this throws up is: could future systems help identify deterioration earlier? Could they help monitor the response to cuff deflation? And could they support more ward-based surveillance outside of a highly monitored environment like an intensive care unit? Well, I think the answer to all of those questions is possibly. We've got to be a little bit cautious because this was a very small single-center observational pilot study, only had 18 patients. Measurements were also taken without speaking valves in place, which does limit the generalizability to quite a lot of rehab pathways out there. But the technology doesn't tell us whether swallowing is definitely safe or definitely effective, only that swallowing activity is occurred. But the correlation with the fees clearing the secretions sort of suggests that there's some positive effects going on. So we're nowhere near replacing instrumental assessment or that careful bedside assessment. But as a proof of concept paper, I think it's really interesting. And it fits with that broader trend that we're seeing and that we discussed on this podcast that moving away from purely episodic assessment towards this continuous physiological monitoring and airway swallowing and care is the direction of travel. Our second paper comes from Lehman Colleagues, which was published in a journal called Biomedicines. This looked at something called NT ProBMP, which is a prognostic biomarker for heart failure, but they used it in this context for ventilator weaning failure, and they also discussed implications for tracheostomy timing. Now, NT ProBMP stands for N-terminal, Pro B type, natureetic peptide. This is a blood test that some clinicians, if you're listening, will be familiar with. And if you've got somebody with acute heart failure or signs of heart failure, you can get exactly the same signs. And so what the group were trying to understand was effectively how much of sort of a failure to wean effectively is due to the heart and how much of it is due to the lungs. I mean the answer I can tell you now is it's usually a bit of both, but let's go through the data. This was a pretty big retrospective cohort study. It involved over 700 patients with tracheostomies in a South Korean medical ICU. And the team say in their opening remarks that you know weaning failure is simply not a respiratory problem. And we know there's lots of literature there, and our experience tells us that when you transition from positive pressure ventilation, where the ventilator does all the work to hybrid modes where the patient does a bit of work and the ventilator does some work, or whether the patient is breathing totally spontaneously, that does place cardiovascular stress on patients. And a lot of patients fail to wean because of this cardiac dysfunction rather than primary respiratory mechanical dysfunction. And so the authors examined whether the admission levels of this pro-BMP could predict ventilator weaning outcomes. And they found that higher levels of ProBMP on admission were strongly associated with weaning failure. So the number that they came up with was if you had a Pro-BMP above 3,270 picograms per milliliter, patients had nearly three times the odds of weaning failure, even when they just did for other factors in their analysis. Higher biomarker levels were also associated with longer ICU stays, which goes hand in hand with a longer weaning duration, and increased 90-day mortality. So this is identifying groups of patients who've got bigger problems with their weaning. Equally interesting though was their finding regarding tracheostomy timing. So patients who received tracheostomy within 10 days of starting to be on a ventilator had much lower odds of weaning failure. Now, this is where things get a bit more nuanced because we've been talking for years around the benefits of early versus late tracheostomies, and many trials haven't shown any particular benefits. But this paper does raise some interesting possibilities. That timing probably matters more importantly when you've got specific physiological phenotypes or different groups of patients. And patients with these sort of high cardiovascular vulnerabilities represented a group who behaved differently during prolonged weaning. And that was a speculative sort of hypothesis, but I think potentially quite important. This was retrospective, observational work from a single center, so the usual sort of limitations apply. And pro-BMP is not just influenced by your heart function, it's influenced by your kidney's ability to clear the peptide, sepsis, multiple non-cardiac factors that are extremely common in the IC population. So it's not just like a magic number that tells you there's a problem with the heart. And of course, when you find association in a retrospective data set, that doesn't necessarily mean that one has caused the other. So association does not mean causation. But I think the key message is a bit more broader. When weaning repeatedly fails, we need to think beyond the lungs. And cardiac function, fluid balance, cardiovascular reserve are often very central to the problem. And certainly in our practice, it's very unusual for us to get to a stage where we're trying to wean a patient who isn't on a load of treatment for heart failure, like beta blockers, ACE inhibitors, angiotensin 2 antagonists, and a patient who isn't aggressively diures. And that is our experience that we we tend to assume everyone's got a bit of heart failure, and then we we treat that aggressively. We try and link that with dynamic studies like echocardiography, and if a patient is really struggling, then we'll we'll get straight on and think about can we try and fix that with um you know uh angiograms and coronary interventions because you know that the heart does have a really big impact on respiratory weaning. And for any tracheostomy teams out there, I think this just reinforces that multidisciplinary perspective uh is absolutely essential when you're thinking about the care for our patients. It's not just what we're gonna set the ventilator on today, it's it's what else can we do for this patient to try and get them better? And often the answer is have a look at the heart as well as loads of other things, all the sort of the psychology and the analgesia management, getting people moving, eating, drinking, all the sort of things that we've been uh banging on about for many years. Okay, our third paper is a randomised clinical trial from uh Dolalay and colleagues. This compares two prolonged ventilator weaning strategies in patients with tracheostomies. They looked at patients requiring prolonged mechanical ventilation, and lots of these patients have already failed multiple spontaneous breathing trials. They looked at two fairly common approaches, so something called pressure support ventilation-based weaning, and then a therapist-led patient-specific program that they call TIPS. Now, there's a bit of subtlety in the differences here, but you know that there wasn't a huge amount of difference that I could detect between the two programs. If you want the details, the links to the paper are in the epitaphs to this podcast. But what the authors do say is that you know our patients are often highly deconditioned, they've got neuromuscular weakness, the critical illness, uh, frailty, multiple comorbidities, the heart failure that we've just discussed, and despite even specialized weaning units, outcomes do remain quite poor. And what's striking here, the authors didn't find any major differences between the two weaning pathways. So liberation from the ventilator's success of 30 days was roughly similar between the two groups, as were mortality and discharge outcomes. So at first glance that might seem disappointing, so you know they haven't found that one weaning pathway is better than the other, but negative trials can still be quite informative, I think. And what this study highlights is just how difficult prolonged weaning actually is. These patients had very high illness burdens, long durations of ventilation, and substantial ongoing morbidity. And so the biggest message here, I think, is that no single protocol is likely to solve prolonged weaning in isolation. And success probably depends on the integrated and multidisciplinary team. So physiotherapy, speech and language therapy, secretion management, nutrition, psychological support, optimizing sleep, mobilising the patients, being careful about the respiratory loader that you place on, but consistency in the approach. And one of the interesting observations here was that continuing to wean beyond those 30 days did improve that liberation success. So sticking at it and being consistent and continuing with that sort of weaning pathway eventually leads to success in a lot of our patients. So prolonged weaning trajectories are often measured in uh weeks to months in the real world rather than the days that you often see in some of the studies that are limited to say 30 days. Now, this is a small study, just over 50 patients, it wasn't blinded, but I think it's still a valuable prospective randomized uh data in a population where evidence remains relatively thin. And I think it reflects again what we do in our practice that we recognize that not everyone can follow a protocol. So we tend not to write down our weaning sort of policy, if you like. We don't actually have one. Everyone buys into a set of principles, and then we've got the team, we've got the expertise, and we've got the equipment to try and deliver that uh weaning pathway consistently. And we're open to review, we're open to change, we're open to obviously working with the patients. And if you know, sometimes we give patients complete rest for a week, take a week off and focus on you know your strength and conditioning, but we need to get back to sort of working, and you know, we also recognise that not everybody can be weaned from a ventilator, and I think you know that the more you do it, that the the better you get at it. I thought this was a really interesting paper that should prompt anyone involved in this sort of area to think about your own weaning practice. Now the final two papers this month both focus on artificial intelligence in tracheostomy care, and this will probably generate a bit of discussion. The first is from Silicon colleagues, and they compared chat GPT and Claude responses to pediatric tracheostomy caregiver questions. So these are two large language models with public interfaces that I suspect many of you will have come across. But what the authors did here was they submitted 17 caregiver-focused questions to both of these large language models, and then they asked pediatric otolaryngologists or ENT surgeons to rate the responses for accuracy, completeness, clarity, and relevance. If we zoom straight the results, ChatGPT scored slightly higher for accuracy and readability, but Claude scored slightly higher for completeness. And overall, both systems perform reasonably well. But before anyone panics, this isn't meaning we're going to replace specialist nurses, medical staff, or tracheostomy educators with tracky chatbots anytime soon, but the paper does reflect a reality that we probably all recognise that patients and caregivers are increasingly using AI tools right now. So the question is not whether they'll use them, the question is is the information they receive safe, accurate, and appropriately contextualized. And that's where things get a bit more challenging because AI systems can generate very convincing answers even when they're not right. And I've certainly come across situations where I've asked uh Chat GPT or another large language model to help me out with uh with a reference, and when you look at the reference, it's made it up, and so clearly you need a human in the loop to make sure that these things are providing correct information. The systems can lack nuance, they might not recognise an emergency if someone's asking for for help, and they don't understand the clinical context of the individual, in this case as a child. They deliberately didn't assess urgent or emergency scenarios in this particular study, and that matters importantly. But I think this paper raises important questions about you know how we're going to develop our supplementary education tools in the future, particularly for sort of routine reinforcement of learning or this sort of low-risk information support, and that's where I think AI has a real role. That leads us nicely on to the final paper of the month, which will probably feel um very familiar to regular listeners of the podcast. This is from uh Ben Gorse, the lead author and colleagues, which include uh several familiar NTSP uh friends of ours. We evaluated uh an expert-trained generative AI system. Now, I resisted this initially, but it's called AI Brendan. So an AI version of me that you may have seen on the social media uh or on the website, and the premise behind that was we had an AI bot that could generate uh AI content that we had trained to be appropriate, trained to be accurate, and and we'll go on about the testing about this in a minute. But the the notion that we just described above that you want your AI to be trusted and to be familiar, um that's why we ended up making an AI sort of avatar clone of me, uh, because the colleagues uh in this team felt that having someone that was a recognizable face in uh tracheostomy care that was uh hopefully trusted uh would benefit people when they were sort of hearing these responses, and obviously that comes a bit of responsibility, and so what we set about to do was to make sure that the responses that AI Brandon was giving were accurate and appropriate. So, what the AI Brandon system is, it's ChatGPT-4 is its base, uh which gets updated as this OpenAI releases sort of newer versions of their large language model. But we added on top a curated dataset of over 900 tracheostomy care questions that we'd gathered and scooped up during our training courses over the years, going back, I think about as far as 2009. So, all the questions that we've been asked, we'd wrote them down and we answered them with an expert panel, and then we fed them into the large language model. And it's not publicly out there on ChatGPT, it's it's contained within this particular model, so it's like a subset that is sort of specific to AI Brendan. And what we figured was that most of the questions that people ask on our courses represented most of the questions that people would ask of our chatbot. We then examined the responses generated from AI Brendan and we compared them with those from ChatGPT 4.0, which was uneducated, if you like, with this curated dataset. And we also asked human experts. We used a validated tool called Equip, which stands for Evaluation of Quality of Information for Patients. And this is a tool which is used as an assessment framework to systematically rate the accuracy, completeness, clarity, and usefulness of healthcare information provided to the public. The answers were anonymized and they were rated by blinded members of the public who were university students, and they had reviewed and scored the answers without knowing which source had generated them. We compared the human experts in one group with ChatGPT 4 with AI Brendan, so three groups of answers. So interestingly, AI Brendan scored higher than both the standard Jat GPT and the human experts using the modified equip tool. And you'll see in the discussion we go on to offer some reasons for that. And the the equip tool sort of favours more complete answers, and our human experts were probably a little bit pushed for time, although they received the same instructions as the AIs, but perhaps the AIs were a little bit better at following the instructions, so that the human expert responses would tend to be a little bit briefer, which I think probably led to them scoring less accurately. We also looked at the content of the responses, and the authors read each of the responses to check for inaccurate or alarming responses. Unfortunately, we didn't find any from any of the groups, including the experts. So before I get too carried away with the AI ego, there's some really important caveats here. This doesn't mean that AI is suddenly superior to clinicians. What it does suggest is that with a carefully curated, domain-specific, artificial intelligence sort of chatbot, if you like, training can be delivered consistently and educational quality can remain high in certain contexts. Okay, and I think consistency is probably one of the key concepts here because human teaching varies, experts vary, which is what we saw in this study. We've all got different phrasings. Emphasis, occasional emissions, and different opinions, and the expert trained AI system can potentially help to standardize this sort of baseline educational content. There's obviously risks with this approach, and the training data set matters enormously, but we've tried to get a genuine training data set which is collected from our real life experiences on our courses. There's biases there, so what you're seeing is our opinion if you like, but it was a multidisciplinary team opinion from a group that is well regarded in the field of tracheostomy care. Hallucinations may remain a problem. We didn't see any, but as I alluded to before, that is a problem with generative artificial intelligence. Governance matters, so we had a human in the loop, and we will continue to have a human in the loop when we're assessing and evaluating the responses here. And educational support is not the same as clinical decision making, which is really important. So I like this paper, uh, written by my colleagues, but I had a little bit of input into it, because it frames AI not as a replacement for clinicians but as a possible adjunct to improve accessibility and scalability of training. And we'd like to be in a position at the NTSP to deliver consistent training at scale to anyone who needs it. And we think this paper is the first small step towards doing that in our very niche area of tracheostomy care. And you know, anyone who works in this space or has experienced tracheostomy care will know that workforce training challenges are very real and tools like this may prove very useful, especially perhaps in geographically remote settings for out-of-hours educational reinforcement or in settings where people, staff, family members find it difficult to access high-quality, bespoke, perhaps face-to-face training. So we think this is a step forward in appropriate use of artificial intelligence, but I look forward to hearing any comments anyone's got. So we've seen all these papers tied together this month with the intersection between technology and human expertise. Whether we're talking about physiological monitoring, biomarker, guided risk stratification, ventilator weaning pathways, or AI-assisted education, none of these technologies replace multidisciplinary clinical care, but they might augment it. And perhaps the bigger lesson from all these papers is that successful tracheostomy care still depends on integration, and technology alone is not the answer. Careful, clinical reasoning, MDT collaboration, patient-centred rehab and good communication are all fundamental pillars of good tracheostomy care. And if we apply these thoughtfully, as some of these papers have demonstrated, that some of these tools can help us deliver care more consistently, more objectively, and potentially increase access, particularly where access is not equitable. As always, there's links to all these papers in the show notes, and you can find descriptions on our website, which is tracheostomy.org.uk. All the opinions you hear on our pods are our own and don't necessarily reflect the views of the organisations that pay our wages. Thanks very much for listening. That was Tracky Talk from the National Tracky Safety Project, produced by Simon Williams. Take care and see you next month.