Remote management of vital signs information data by Yun Long
AI to prevent ventilator dyssynchrony by Elias Baedorf Kassis
AI to guide weaning and predict the extubation success in acute respiratory failure by Samir Jaber
Good afternoon everybody I’m um Alami from um France from the teaching Hospital of R I’m a full professor in critical care and it’s my pleasure today to uh share such a prestigious panel of speakers who will be talking about a current and a future challenge for all the intens
Definitely the issue of how to use data and artificial intelligence in the ICU and we are of course all aware about the considerable amount of data collected in each day in in our intensive care units that definitely recalls the need of the uh to use Big Data uh analysis in order
To improve uh our patient outcomes in this webinar which is uh supported by MRE and we are thankful to MRE for this support who will particularly focus on the remote uh management of Vital Signs information data uh how artif artificial intelligence can prevent ventilator uh desynchrony and guide weaning and
Predict extubation success in acute uh respiratory failure before beginning just a few words to tell you that there is a chat and I encourage you to use it and there will be a dedicated time to ask all your question at the end of the webinar so please don’t miss this uh
Opportunity and it’s my pleasure to introduce the first speaker um Dr J long who is the director of the Department of Critical Care Medicine uh of Beijing Union Medical College Hospital he is also a member of um standing committee of Critical Care Medicine of Chinese Association of pathophysiology and also the vice
Chairman member of Critical Care Medicine of beijin Medical Association and and member also of the National Committee of Critical Care Medicine of Chinese Medical Association he will be speaking about um to Remote Management of vital science information data and so uh I welcome of course Dr Young and uh the floor is yours
Please thank you thank you and everybody you now I’m you in Beijing okay now the Beijing is in the evening and good afternoon everybody and I’m share my slides about the Remote Management of vital sence of information data uh as you know U mechanical ventilation is very important for critical yield
Patients you can see the numbers and about 40% of patients as you lead the mechanical ventilations treatment especially for acute respirat fers um around uh 16 7% and mortality is around the 40% during mechanical ventilation the phenomenon is human and machines simply is very common uh in some angles the IU
Uh source is a very scarce resource you can see and we have the numbers compared with United States you can see in the numbers of ASU bets in the uh China as is uh uh we have much more I BS I have much more doctors but the difference the
Ventilators is not isn’t significant in contrast equipment to the B rual and respiratory and therapist bed rals in China is are actually much lower than those in the United States so um in any stage me ventilation will brce a serious adverse effects such as discomfort and flow hunger and increase the respiratory
Efforts and especially for the D muscle vertigal and such as a lot of sense which should to pay much attentions that means U PR need mechanical time and difficult overwheling and increase Mor moralities so that is the global problems that need solved urgently and need a huge problems in manual analysis
Of patient ventilator or syus and especially for uh a syn waveform datas that’s through that you can see uh uh in the global have say some uh researchers and issues next especially in American Journal of cre Care Medicine in 2011 and the European joural respirat medicine 2017 and the American Heart announc in
2020 they pay much attention to the uh patient as and ventilator is and you can see well sure well the uh research and there has pay much attention to the patient V s you can and especially double trigger and automatic trigger and invalid triggers like that in 39 hospitals in Middle Middle East
Around 400 and 11 um I STS including the respirator surface that is interesting sense you can sh um you can see the the this uh result um comparison the TR the non TR the the personels that has no difference no difference of they recognize or simp especially only only the difference can
Be shown after the tring after the tring the vator ining TR after that you can see the different so the training is very important and we can show mechanical vintage with for identifications have hasn’t not been solved this is global problems will be solved urgently so now a lot of companies
Involved in this problems SP F and and and g health cares and they do some works and the F from 2014 to 2018 sign improv the ASU treatment they decrease the mortality and shortens the ASU St and and the numbers and um make the clinical decision so that is very big um
Significance benefit but we have still some of very shortcoming is the data pressure is very uh difficult to transfer and this is a a close sistance limited to Philips and G sance so what we do our department have developed the remote mechanical ventilation Monitor and intelligent uh analyis Technologies we can B mechanical
Ventilation and with the cloud computation and big data and the AI algorithm this technologes connects ventilators across suu in labeling and remote monitoring event warning and data management and a dynamic analysis of the mechanical ventilation datas we convert result the this ventilated data management s world so softwares has received the naal
Medical device class two registration certification so you can see this is our data input from the Su word to the uh suu Cloud flight and they have the um data management centers we give feedback to the clinical doctors and that can be uh 24 hours data analysis
And diagnosis for the critical care uh dat Center the clinicians and send back to the best side doctors including diagnosis and treatment decision and to adjust the venator parameters that is the uh closed data medical datas so this is examples you can real turn the revolve the um best side monitoring for
Patients including the monitoring system and mechanical ventilation mechanisms this is seven multiply 24 hours real time M for patients so we need the quity control the mechanical ventilation this is p p medical uh College the center we have um L of the um registration the hospitals they send that um mechanic ventilators
Message to our PL plight Center we have the uh doctors and 24 hours real time uh on to analyze this uh message and to give feedback to the bside doctors so we do the data man M data processing and database management and remote monitor the data collection and
Transmission and do the event analysis include Abal event warning and warning assessment and give the data report that is every day we give the ventilator data analysis report let so this is centrer support including standardized the data connction actions M earing Waring service and intelligent data service and
Ventilation mode automations and that is our supporting service and that is algorithm implementations so this is I think this is a hard work from the best side continuous real time collection data data and uh to sure the data security in integrity and efficiency and we do U
Standardize and few fusion and we can uh prediction and accurate identify the analysis the multi level events and massive data storage and intelligent retrial and do the most even early morning intelligence ensur the treatment comfort and infection so especially we pay much attention to the patient ventilator or Sy if first
Show the Cy cycle trigger and and the expiration and on the ventilator of including flow image air TR double trigger reverse trigger and secretions conation and condensation and NE so uh this is a close uh loop from the more uh the data and personalized me ventilation and World change and
Including non protection ventilation to minimize the patient ventilator sority so and we gave the patient gave the clinical doctors report that is uh uh standardized the uh the uh ventilation report including the everyday ventilation mode and the T of the non protection ventilation data and aen and respiratory
Mechanism so it give sound case this is um male 13 years old after the cardiac surgeries they have left num strikes the ATT want to control the spontaneous brazing because of the this the air leak around as 3 lit per minute and we do some increase sedation but but the
Software and the respiratory therapist is a remote flat for recognize it might be a auto trigger through the iser pressure form so the actually recommended adjust the sensitivity from inspir trigger from two to four minute so the spontaneous PR is PR and an other patient after isure cancer
After surgeries the patients becom the resistance around 20 six and relative now run flow 30 but the high peak pressures so after we just the depth of under trial tubes we have Chang decreased the resistance we found the some Noble is an neas we do the uh track
Is um intelligence to to open the this l so other cases is like and is a delayed cycling you can see the remote for the specifical algorthm you can see uh end of inspiratory exist The increased pressure levels so that is is decre delayed cycling the thisp time is
1.3 for so we decrease the insort time short that and we can see the figure the pressure end of spat pressure decreased the the uh the pressure increase disappear after that so this other cases is is a uh rupture ad is suffer the after surgery abdominal reasons suff ARS so can see
The title volume and the respond reach the virus if we increase pressure support as to 1 so the increase tile volue decrease the respons reach and decrease p01 when we decrease pressure support and we can decrease TI volum and increase reach and increase P one so this is a very interesting case
It can shot to our doctors and training our physicians to Lear a lot of respiratory mechanisms so that’s what I inclusion the patient vator ASN is very common and gradually impressive to everyone so we do the remote ventilator management that is involving in China of speci for scarcity resource regions and AI re
Recognization especially for the patient ventilator ASN is under research the initial report is Improvement and training and feedback are necessary for further Improvement thank you everybody now finish thank you so much uh thank you for this uh nice presentation uh Dr long and as I said also the questions will be at the
End of the webinar um and now it’s also my pleasure uh to present um Dr elas bedor CIS Cassis who is assistant professor at Havard medical school and medical and scientific director of the resp atory Care at the BET Israel Deon Medical Center in Boston maets and um Dr uh bedor
Casis he he has a an expertise and research focusing on the application of the basic uh physiology principles uh to the management and to the prevention of the acute respiratory uh syndrome distress syndrome and um he’s also um he has also some uh expertise on machine learning approach to the ICU
Physiological signals and um this is uh also uh the the the subject of his uh uh topic for the uh the just coming 15 minutes uh Dr bed Cassis will talk about uh uh how artificial intelligence will prevent ventil dis synchrony so please Dr bed Cassis the floor is yours thank
You thank you so much for having me today um it’s a pleasure to to chat about this this uh subject with you all today and as I’m sure you will guess we are just at the cusp of what we are going to be using for this so some of it
Is talking about what we’re doing now but also a look to the Future as well uh here are my disclosures um and so let’s just start with a quick case so this was a 48-year-old patient with covid ards uh they were very deeply sedated with a Ras
Of negative five and they were very desynchronous despite this this high level of sedation um and I think this is a common scenario where we all find ourselves as as critical care clinicians what do we do in this scenario do what do we do in the ventilator um what is
This how do we intervene on it um and what kind of support can we um have from a A system that helps us out with this um I think you know we have a very a good understanding now that that desynchrony overall may have a a pretty good um plausible mechanism for causing
Lung injury um there’s certainly concern for overdistension injury with large swings of pressure across the lungs the transpl pressure um large title volumes it maybe uh cause of hypoxemia increased muscle uh work and discomfort for patients and cardiovascular compromise and unfortunately one of the major ways that we treat this um potentially um
Incorrectly so is with deep sedation and deeper sedation we are increasingly um aware of as as critical care doctors may be a an important cause of harm in patients with deeper sedation causing worse mortality and this has been recapitulated over and over again in multiple different studies this is just
Our own work from our lab um and in terms of um the the potential implications of D synchrony um it has been associated with increased length of stay um uh increased mortality and here a study by some of our Spanish colleagues looking at the asynchron index which is a combination of combined
Ineffective efforts double triggering and other forms of D synchrony that they combined into one single IND index um where they um showed that higher levels of um of asynchrony in this case greater than 10% um index was associated with worse than outcomes in patients um and
Um in addition uh there’s been um met analyses which have shown uh the similar findings as well and this just looked at uh both asynchron IND indices as well as ineffective trigger indices um looked at that and found that greater than 10% of both um was associated with worsen
Outcomes wors mortality wors than um ventilator um uh times on the ventilator as well but I think the important thing to think about though is that um are all forms of desynchrony the same in terms of harm and how much more can we do in terms of better understanding and
Differentiating the different forms of D synchrony the as synchrony index combined five different forms of D synchrony and they all really may have a different potential for for harm and maybe some of the ones that are hiding in here are actually maybe the Kaiser so of of the group the reverse triggering
Of the group is one that maybe we should be most worried about but it’s the hardest to actually discover and find now I won’t go through all the different forms of D synchrony in the interest of time today um but the reason why I bring this up is that each of these D
Synchronies it may be a high drive D synchrony or a low Drive DN it may cause injury or maybe a nuisance it may respond to sedation or get worse with sedation um and so um lumping all these forms of D synchrony together in terms of a single way of monitoring
Patients is probably doing our patients a disservice as well as really not allowing us in our research to uncover the differences between these different forms of D synchrony and the implications for clinical care as well as implications for treatment as well and so really we have a some
Important challenges when it comes to D synchrony one of the most important challenges which I think our um our prior speaker spoke to very nicely is this concern for detection and due to Patient uh sorry due to clinician time as well as clinician training many clinicians probably are not comfortable
In terms of detecting um uh different forms of D synchrony there’s very few systems that are available to assist in this and the remote system that was spoke speaking about by our prior uh uh speaker was really very nicely um displayed and uh discussed some of the
The future um uh ways that that we can use technology to help us out additionally I think this is one of the most important um applications or or concerns that we currently have is the correct identification of D synchrony different forms and phenotypes are all very different from each other and and
Correspondingly understanding the implications from these different forms and phenotypes is is very important in terms of the clinical impact and the and the potential for harm and then lastly obviously how do we treat these and these are the current um challenges I think that really we need more help um
In terms of addressing them and the current systems that we have in place are really not adequate um most current systems um not all but most current systems use uh quote unquote rules based programs where uh different um tracings the volume the flow the pressure tracings are analyzed and they looked at
In terms of what deviates from the expected and systems like intellisync or better care um utilize often times utilize similar systems that look at the these deviations and so really you can imagine though that these this is a great system potentially used augmented or artificial intelligence um to help us
Out with it and so this was the way that we um uh took our approach in our own lab with this and what we wanted to do was really kind of focus on two forms of dis synchrony that are very obvious and also more a cult and uh a cult being
Very difficult to find or potentially missed by uh clinicians to give ourselves a little bit more of a challenge um and so when we’re looking at here you can see is breath stacking to synchrony we have a diaphragm contraction you have this very large title volume you can see in the volume
Tracing this is integrated flow so you can see these large volumes with large trans pum pressure swings um this is a breast stacking event that we’re seeing two of them in in the screen and breast stacking as you may recall comes in two flavors we can see here that on the left
We have reverse triggering secondary um to entrainment causing breath stacking and on the right we have double triggering um which is secondary to High Drive State these are very different from each other but also can appear similar if you’re looking at the ventilator tracings and so having ways
Of differentiating these two can be very useful and so when I say reverse triggering May some of you may be familiar with this but just a briefly review this is a neuromechanical coupling between the ventilator and the patient where the patient has a reverse trigger meaning their diaphragm contracts but after the ventilator
Breath is delivered um and you can see here that this is a tightly coupled event where the VOR is delivering a breath um and the patient has a corresponding uh diaphragm contraction if the ventilator is uh the as paus in terms of the delivery of the breath during breath holds then the
Contractions as you can see in the top tracing the soft Gill pressure tracing the PES tracing those actually stop this is this is reverse triggering but this can be very difficult to determine at the bedside for clinical care for patients um so what we have actually
Been developing is a a tool to improve detection at the bedside and so many many of you may be familiar with using a soft gel manometry we have taken this a little bit step further and using the camel diagram which was a a basic physiological principle where we trace
The volume the plural pressure change over time and you can look at the relaxed chest wall the elastic recoil pressure of the lung and then all spontaneous forms of breathing as well um and we started looking at this when looking at reverse triggering which was significant enough to C to cause breast
Stacking events um and in this we were able to find that this was very common it LO caused large title volumes um and large pressure swings across the lung but what we’re really interested in doing this is using it to develop a more systematic way of approaching all forms
Of breath of um uh of desynchrony and developing this quote unquote fingerprint tool for breathing patterns and so we apply this to reverse triggering to different forms of reverse triggering different phenotypes looking at reverse triggering during inspiration which has the potential for causing overd distension injury uh and
Self-inflicted lung injury risk as well as reverse triggering during the exhalation phase with both of which look very different in terms of the phenotypes on Campbell diagram but can be difficult to differentiate um using standard time tracings and as you may recall the reverse during the EXP phase can cause Ecentric diaphragm
Contractions which may result in diaphragm injury and um and we know that diaphragm injury has been associated with wors mortality in several studies as well and so this led us to this desire to develop automated software for looking at and detecting this difficult form of D synchrony um known as reverse
Triggering and so we um started by developing first a um a rules-based program and then secondarily um wanted to develop off of that a uh a deep neural network and artificial intelligence-based program to detect these um these forms of dis synchrony so we trained it initially on a group of
Patients that were enriched for known RTS um uh looking at over 10,000 individual um events used manual identification as the gold standard to help train the progr and then um initially developed um systems uh rules based systems with soft manometry here you can see we took that more complicated camel diagram and
Turned it into a pus inory muscle pressure tracing um which is kind of the equivalent of it during a Time tracing to develop a um rules based software but then decided that we would also utilize this to um to develop an AI based neural deep neural network pattern recognition
And some of you may be familiar with this but neural networks’s obviously the older versions you had an input layer and output layer with this sort of um uh one simple layer of hidden layer in between but more deep learning um neural networks have at least three hidden uh
Layers and you can see here this is particularly useful for um looking at um at Imaging data you can see the input here is this picture of George Washington it looks at the um the sort of different pixel values identifying um edges combination of edges sort of further um developing these combinations
Of edges into being specific features the eye the nose again you can get into further detail how to the the shape of the nose the shape of relative to the eyes Etc and each of these represents a different layer to give you eventually the output which tells you that this is
Uh George Washington all right we can use the same approach for um uh for Imaging data and waveform data which I think represents an very exciting future use so how do we actually do this we take the full data data set we divide into a training set and a testing set
Then we use a validation set which is a subset of the training set which is not fit into the network during the parameter setting part of the training but then used after every iteration of the training while the net is frozen to then check on the performance and you
Can use this actually to look um to see if the training error is decreasing but uh for example if the training error is decreasing but the error in the validation set is increasing this can be sign up overfitting and then you have to adjust accordingly and then after you’ve
Optimized um with the training set then you then test it um in an independent set of data that has been untouched to see if the performance actually uh continues to um um to be as good as you would hope it to be um and so what we
Did is we um input initially this the sort of typical tracings you might see on a ventilator um looking at flow volume and pressure tracings reverse triggering versus non-reverse triggering on the right um and the left and the right here um and unfortunately we saw that the performance actually was really
Not as good as we would hope it be and part of that that was due to maybe the just the numbers of of events again we’re looking at um many thousands of events but maybe this simply wasn’t enough um but then we were like well
Maybe we are we need to use the system similar to the to the way that we our own brains operate so let’s input into it the same um uh fingerprints as we talked about earlier into the system that we use for our own manual identification and what was very
Exciting is when we input this in a pictorial form the performance of the system dramatically improved and you can see here these are just the different um uh um uh machine learning approaches or the neural networks that we use the res net um and
The and the net minced um at the top um input these two-dimensional um Campell diagram based um images uh versus the um three at the bottom looked at the one the standard onedimensional and the performance dramatically improved with improv Precision recall um AC um accuracy as well as the F1 score which
Is kind of our um a combination of of these with really achieving our goal F1 score over 80 which um which which we were hoping to achieve and the other exciting part about this is when you look at um these uh in um uh um in what’s called principal component
Analysis which is looking at these tensors and a tensor is a multi-dimensional vector which allows for um understanding um uh all the different components that lead to the unique nature of an individual waveform is that as we increase the number of waveforms you can start to see this
Clustering that occurs by phenotype and what we’re very excited about is using these clusters to help better train systems and then identify at the bedside um for clinical care as well as for research purposes and so ultimately I think where is this taking us with our
Care um I like to think of artificial intelligence really in terms of our clinical care is more augmented intelligence um how do we use these systems to support our iCal care um are we letting the systems take over for us or are we instead using it to help
Optimize and improve care and I think we are going to be doing this in a number of different ways we’re going to be using automated detection identification of D synchrony using artificial intelligence or augmented intelligence we’re going to be using closed loop systems to optimize the event in
Response um and then also using U artificial andart augmented intelligence to better understand the different forms of dysynchrony that are harmful versus benign that can be ignored and to really use it as a helper to provide decision support and reminders to clinicians about Best Care um as you can see from
Our prior speaker we’re maybe closer to this than we um would uh would anticipate but the the key is now expanding it and normalizing this in in uh clinical care across all centers to really improve um the par the care for our patients using using AI uh so thank
You very much uh for your time and um uh I have you to take questions at the end thank you thank you so much uh thank you for for this uh nice presentation and also to give us some U uh some of the vision about the future and how can
You use this actually at the bedside in our um units so um now it is my pleasure uh also to to present um uh a professor but only a big friend uh the professor uh Samir jabber who is uh Samir is the head of the critical care and anesthesia Department of
S in France um he’s uh of course uh a very brilliant researcher in the field of the peroperative medicine from the operating room to the Intensive Care Medicine so obesity Airway uh invasive and noninvasive ventilation and irds and Samir will be telling us in the 15 coming minutes about um how the
Artificial intelligence will guide the winning and predict the exibition success in acute respiratory failure so Samir the floor is yours thank you very much my dear ala it’s a great pleasure thank you for the invitation from my colleagu of isman thank you for the mry support before I
Start I just want to to express my to all my Moroccan brother and especially to the victims of the fqf my solidarity during this difficult period we think all of our Moroccan people so I will start uh uh this is my disclosure I propose you in the next 15 minutes to speak
About why I in medical ventilation could help the physician what outcome might we expected using EI in and how work in patient with acute respiratory failure I will give you some the more clinical evidences and why clinician unfortunately do not use uh EI and the limits of this uh technology and some
Take home messages first of all this is a remember how works the ventilation control system in a ventilator in a machine there are three levels of control first of all the classical ventilator mode that you use pressure support control vum mode Etc it works with one parameter so it’s called the
Theory of control and this the closed loop choose one Cycle One respiratory Circle this is the unit of millisecond the second could be at least two independent parameters this is very close to a physiological approach fixed and this the Clos Loop works and cycle or during one cycle after another one so
The machine calculate what happened the cycle before and adapt the next cycle or during the St that’s work like the proportional acid ventilation or Nava Etc and finally the third one which is the main topic of today is how the global view could be done using the artificial intelligence then called the
Therapeutic strategy using the medical reasoning models that means a machine could or will replace the physician to decide and make the proposition and this needs need few minutes or few hours and then uh different modality now are available I will present you then what might uh what outcome
Might we expect using the automated mode with artificial intelligen I can try to resume to summarize it in 10 Point first of all we hope that it could reduce the ventilatory and L injury using a protective mechanical ventilation with the AI to use an automated protective ventilation reduce
The diaphrag anduse injury the viid use less sedation needed and then from the beginning to the end of invasive ventilation during throughout the ventilation period then we hope that uh it could would reduce the winning time for duration of wind mechanical ventilation decrease the workload of the
Nurse team the physici team it could be safe and cost Effectiveness and at the end we hope that could improve the outcome and especially the mortality of the ICU patient so cold an automated mode with I provide a total protective ventilation from the intubation to extubation period i i this slide to
Summarize how could Works an a machine using an artificial intelligence First Step first of all in the acute respiratory filler we try to set to have an Target of tidal volume classically is around 6 m per kilogram between 5 to seven even uh controlled or spontaneous breathing and use a positive and
Expiratory pressure from 5 to 15 cm for the majority of the and then the second step is to use some therapeutic adjustment using some recuitment maneuver in some selected patient who need recuitment maneuver and then this try to optimize the L volume with to avoid the volum on barot troma
To have the minimum level of high respiratory rate or low respiratory rate than the optimal respiratory rate to have the optimal minut ventilation that could the target of tco2 on the the work of breathing of the patients and then to have the good oxygenation with a target
Of fi do not too high not too low and have to to minimize the um baru and Vol trauma by maintaining a plateau pressure less than 30 a driving pressure than 15 and now more recently we showed that mechanical Power should be or could be less than 27 Jew per minute and then
Using all of this technique could improve the patient and at the end the first step is try to have the early winning phase as soon as possible start winning as soon you could extubate the patient then you could the quick fast winning process finally we can summarize my slides by from Total controlled
Ventilation at the start of the acute phase of the pathology at the end during the spontaneous ventilation from start of winning at the end of extubation then now the physician could do this but the question is can we translate what physician do with his or her brain to the machine with
EI why it’s necessary why automated mode are required for the future because as you see on this Dynamic Slide the population over the world increases and more ICU patient will be in our unit than more elderly patient with more comorbidities then more and more now especially in France and other country
We have unfortunately less nurses less physician less expert in mechanical V that we need and we the experience of the co learn us what we should do then and Care try to be more expensive with human resource today so we in we can speculate that the automated mode with
AI could improve all of this and this in this 20 years old dere conus publish this paper using how the demand and Supply there is no the same uh kinetic and you see the number of clinician will unfortunately not increase and this is absolutely the case in France and then
However the patient on mechanical ventilation will increase significantly So Co an automa mod with I provide the optimal on the adaptative level of ventilatory support from intubation to extubation because the challenge is to not give to less or too much ventilatory support and this can summarize on this
Slide as you see you have the optimal mechanical ventilation if you put in the yapis the risk for example what happened when you have the under assistance you have several things as we just previously presented by alas that the asynchrony the discomfort Etc the P Sil Etc and if you have over assistance
Again you could have discomfort with asynchrony and Vol Bara and at the end you can have an outcome very worse so the machine could should adapt the adaptative support for each breath to the patient because the ventilatory is not the same during the day and the main limitation of the usual ventilatory mode
That we use to day volume control mode or pressure support mode or others that all of them give a fixed delivery vatory support when you set 500 mL you have 500 mL even what happened by the inspiratory effort of the patient and this is the same when you set a
Pressure support mode using a one pressure this is the C why the patient ventilatory demand over times is variable is not the same when the patient sleep when the patient is awake when you have nurse Etc so we can speculate and that’s what the automated ventilatory mode with II exist today
Because using one or several close loop could deliver assistance according the patient ventilatory demand of over the time so what we have about the clinical evidences this is very important for us today to have as you see the two mod that exist today is the ant event from Hamilton and the
Smart care and both of them try to adapt the ventilatory demone the antant mode is to have a different close loop with ventilation control and oxygation controller and the smart care only a pressure support controller the smart care Works to have a patient with increase or decrease the PSV level to
Obtain a respiratory rate in the comfort zone from 12 to 28 but now in the last version you can set your optimal respiratory rate you have also you can set the your tile volume and this you can also set the ENT CO2 to have the Clos Loop Works about this this is the
Summarize what happened with the II and the first study was uh this study performed when Laur Brar Works in worked in CR with Michelle Doja as you see this is a crossover study tested in 10 patient the inadequate ventilation the percentage of total ventilation when the patient have an high respiratory rate
Low tidal volume and high anti as you see here for each patient you have above 30 or 40% time and when you switch with a pressure support automatic EI mode like a smart care all the patients you see here decrease in Comfort Zone this show that it works and than Canada why
Worse because in the real life The Physician change one or two times a days the pressure support setting this is the real life what you do in your ICU in my ICU and when you use the the automatic mode you have more than 15 uh times than change and sometimes more than 100
Because it adapts the Layel and this could be available for example for this patient in our unit you see the airway pressure change over the Time Each cycle the work is that the pressure support decrease if the patient tolerate well they continue to decrease if it’s not possible it’s increased and then when
The patient arrives at 7 cm they stop and can propose you to exate the patient as you can see now we have some clinical evidence mainly for short period in surgical patients however some study are negative for example from Australia because the nurse ratio is not the same
Like in France is around 31 in France at the period of the study and for example in Australia is 1 121 ESV anent works in the same way but from start to the end using different variables to have the optimal UC see oxygenation and tidal CO2 tidal volume respiratory pressure
Respiratory this is a green part and then if it’s not in the green part if could be an accepted level and if it’s accepted high is this toate in the gray Zone if it’s not in this case it change what it change level of assistance and moreover they have a close loop of
Oxygen because we know that hypox is not goodin is not good and then with this Clos Loop control of F2 and P it can prevent hemia and hypoxemia and then using the Clos Loop of the RDS Network as you see here we can have Improvement of vaccination we evaluate this mod few
Years ago and have exactly the same that previously reported is the Smart care using an artificial intelligence mode we have less time in nonoptimal Comfort Zone in this patient as you see here and you see here the between the PSV is the one times a dead change and here you
Have each cycle a change of the level of pressure support increase or decrease and you see a a higher significant variability which could Al also decrease the aasis by Auto recruitment we ALS Al observed in this study an improvement of oxygenation in the artificial intelligence mode finally now the last
Version also integrat the driving pressure and you see here we have less driving pressure and the last one publ two years ago you have also the mechanical Power this is very interesting and incredible to have an adaptation with the artificial intelligence to have the lower driving pressure and mechanical Power delivered
By the V to have the less baru and Trauma now we have some evidences very few randomized control with very few uh patient in more often single Center and you can see some positive and some negative study about the total duration of ventilation but a lot of study were
Were positive for post cardiac surgery patient because this mod could be for interest for very short period as shown in this different uh meta analyis so finally I will finish on why clation do not use it as we we see This Way St study managed by the AIG
We learned that the physician delay to recognize Readiness to win patients the Cal of sedation and unfortunately automated mod with I are totally not used by the patient why because the classical Obstacle of barrier to change not easy to understand for some physician not trust to replace the
Expert physician the F resing the care team and not yet convincing clinical evidence this is the case of airplane autopilot we have some limitation and this why the limits are could not solve all the ventilation problem it cannot be used in all patients and the major
Mistake this is my main message made by physician is that the current automated mode with AI could not not should not be dedicated for the difficult win patient but for the more easy patient so the physician use them for the more difficult patient and this is the case
Here for example the pilot when you are in the plane when the weather condition are good the pilot use the automated pilot with the II it is safe but in case of storm and S turbulance the pilot the pilot must return to manual piloting it’s exactly the same when you have a
Patient difficult to we so in conclusion C artificial inent guide mechanical ventilation my answer is yes probably in selected very easy and non difficult patient it could be helpful for example the very simple postoperative period uh patients or or very short intubation and mechanical ventilation duration but we need more
Friendly userness friendly userness artificial mode and need more convincing reproductive Riz control trailer thank you for your attention oh thank you thank you so much Samir for this excellent presentation and U the real cases that you showed us and um how you Illustrated the your your
Key message I like I love this because it’s really uh very illustrative uh to understand the problem of the use of the artificial intelligence nowadays so uh waiting for the uh for the questions uh from the the floor um I would address for uh all of
You this this question when I was just listening to your excellent talks I I was just wondering I know it is about uh of course ventilation and the use of artificial intelligence in in ventilation and how it works and how can we trust this for um for the patients
But I’m um just I have a naive question and um naive um observation of um uh of uh my practical uh anyway practical um experience uh when we have a problem sometimes of winning uh it can be of course because there are problems to adjust the ventilator and
The the support but it can be also for example um because the patient is anxious because he has a delir delirium um because um because also sometimes some hemodynamic issues do you think really that um just using uh algorithms or artificial intelligence based only on ventilator uh Loops or volume pressure
Loop may help for uh to to solve all these problems for winning uh in uh in patients in the ICU I I I I I need to I really want to hear for the from the three of you maybe I can start if you are agree my colleagues
I I think the question uh of ala is the main question that we have all of us every day when when we have a a difficult patient to win in fact as I tried to conclude that when you have a difficult patient to win that it’s very difficult so the artificial
Intelligencia intelligence now I think today it could not replace the physician because you need expertise we need experience it can help but never replace so if we use only artificial intelligencia with mechanical ventilation close loop we have only a view of what happened about ventilation so the patient is more
Difficult that only lung of respiratory muscles you have the problem of what you said anxiety you have the problem of the pathology what is the pathology exactly there patient have another problem another ventilatory associate pneumonia another sepsis Etc and then we need to have all of you the
Hamic what happen so at this moment it’s not possible with that we have but in the future we need to have something like a GPT machine and integrated the hamic parameter the ventilatory parameter the biological parameters the psychology parameter and then if we have all of
This then we can help us more and more but at this moment the the II automated mode available in the market some of them can combine with a Clos Loop of hemodynamic some of them try to combine with a closed loop of sedation to adap like in anesthesia
Because we have this in anesthesia using the the Beast sedation Etc but I think all of this this is my conclusion could be only available now for the more easy P Pati for the more easy patient that in aend to spend time to avoid to spend
Time and then you can win time for other things but for the more difficult patient it’s very difficult to replace the brain of physic this is my point of view I don’t know if my colleague share this View thank you Samir I think elas wants
Also to to comment yes thank you um I think the your point about weaning is exactly the same as with the synchrony um the um the data we are getting for assisted decision-making usually comes from monitors of the patient but these monitors do a pretty poor job compared
To our Visual and personal assessment of how the patient’s doing um you know we’re getting ventilator data uh we’re getting Vital sign data and these are easily input into systems that can then uh Provide desision support however it’s clearly limited to the kind of data that that these systems are able to actually
Utilize and it’s missing a vast amount of data from the uh direct patient assessment the the physical exam the clinical assessment of the patient which maybe eventually we’ll have video monitors and such that are able to collect this information and then utilize filters and other automated artificial intelligence systems to then
Utilize that data but that’s far away from where we are right now um the data that we have right now is is um already more than our systems can handle and know and we know what to do with and so that really realistically is quite quite
Far away and so I think um uh Dr jaar’s um um comment about you know that differentiating between the very complicated versus the run-of the M patient is really important because these systems will provide decision support and um and provide extra notification and um let us know which
Patients we might be the most concerned about but in the end these systems will not take away our job at the bedside for for patients and I think that’s always a big concern about AI is this going to replace our our job and in no way will
It I really believe that for some Fields it might but not for hours um uh and um and if any think it’ll just help provide better care for patients and let us be aware of when we should be more concerned but in the end it really requires the the clinical judgment and
At the bedside thank you Dr long you want to comment supplementary comment I think it’s very interesting question I think um in my opinion I always pay much attention to patients respirator status I think the if the patient situation is changed the respirator um must be changed in the first and we
Won’t pay much attention to the Mal ventil uh Mal uh uh volume malot volume ventilation volume if the ventilation volume increased we might pay L attention to the U wave four so especially for the covid-19 patients you can see if if the patients exist the inspirat effort strong inspirator effort maybe
The patients suffers hypoxia and unstable so I think hamic and respirat mechanisms is involved together so and for the SE patients for the chemotherapic unstable patients the waveform must be changed so and I like that and want to use the W to monitor the patient situations and to to um
Early to detect the patient situation that changes thank you I have some questions from um from the audience so one question to Dr bed casis um so one commment thank you for your excellent talk uh is an esophagal balloon needed for the AI Asen uh detection and is
There a way uh of um a way of artificial intelligence can work with length compliance to reduce the the chance of um volume baroma yeah that’s a great question um uh a soft manometry um ultimately I don’t think will be needed um but what
It allows us to do is get a much a better signal for these earlier iterations of detecting difficult D synchronies so we have a rules based program which is not your traditional neural network um automated program which uses um signals from the flow v um pressure and volume
Tracings that does not need a soft sh manometry that works fairly well but it’s but it’s um but the you get a lot of additional information of the soft manometry so so if you have the ability to use optional manometry can really help with clinical practice but It
Ultimately will be needed I hope not I don’t think it I don’t think so either but um but it certainly improves the care I also think that automated systems to help as I think was talked about earlier um U the concern for lung injury looking at bar trauma and markers of bar
Trauma and Vol trauma Etc um using soft trometry AO an automated format with AI assistance um what does make a lot of sense as well um and will certainly be helpful thank you and a last question for Samir from the audience um there is mment and and also a question um the
Comment is since smart care have been introduced 20 years ago uh why is still not used widely in hospital I think partly you also answered for your in your uh talk but also the question um is which would you prefer personally ASV or SM or smart care uh this is the the
Question thank you for this question I think that I try to answer why automated mod is not used as I presented in two slides as I explain different things previously and I think uh there is no answer of the question what is better than the the
Other the best is to use I I recommend you to use what you know very well what the ventilator mode you know what is the more safe for your patient and please never change your practice if you don’t have the experience of a new things today we have
No stronger no data no clinical evidence that one mod is better than the other so is up to date the best is to apply the recommendation of the societies that try to have less baroma less vola less sedation less to decrease the duration of mechanical ventilation use what what
You know the best for your patient this is my conclusion all this mod are for the not now available in the clinical critic care care but we should wait more to have a very simple uh automated word thank you so much thank you for uh all
Of you I want just to to conclude uh by the the last key messages that you all delivered is that uh uh artificial intelligence probably um something which will be for the future but needs to be experienced need to be learned uh in order to use it uh really
Uh use it with a with a benefit for uh patients and I like also the the sentence of Dr bov CIS said it doesn’t replace the clinical judgment and I will also repeat the sentence of Professor ja Sam jaab use what you know and apply what you what you know for
Your uh patience and more uh for your more importantly in the more difficult patients so thank you so much thank you for um the great talks um I’m really very very happy to to share this um very interesting webinar about artificial intelligence with you and uh thank you I
Want to thank also of course mindre for the support uh the European Society of intensive care and all our audience for the great uh for the great questions thank you so much bye bye bye thank you bye bye thank you bye