Add eSign Hospital Discharge with airSlate SignNow

Add esign Hospital Discharge

we're going to start the afternoon session with a presentation from narendra ahuja from the university of illinois at urbana-champaign and he'll be speaking about his project uh titled adding audio visual cues to signs and symptoms for triggering for triage suspected or diagnosed covet 19 patients we are working on triaging of patients for efficient use of the hospital facilities and our work is based mainly on using audio visual cues to help with the triaging process so the team our team is consists of three universities um i'm the pi and professor david weisser from university of chicago uh professor david chester and jady krishnan from university of illinois in chicago and professor mark has johnson my colleague at uiuc and i see them in the audience so if there are any questions we can uh we can go to them uh first of all this doesn't need to be restated the situation is is bad and the hospitals have been having difficulty keeping up with the patient load and i've been thinking about how to handle the the load with the limited resources with the emergency departments the early one of the early things that was discovered in this process or thinking process of how to handle this one thing that became clear was that a lot of the patients were apparently okay but there was still a chance of them relapsing and so it was not clear whether to discharge them or not to discharge them if you discharge them it's good because they will recover better at whom but then what happens if they decompensate so the problems were solved by letting them overstay at the hospitals which means more beds which means fewer beds for the new patients so in the context of this the attention to telemedicine became more prevalent people were beginning to think more and more about how to use how to bring telemedicine into the fore it's not been as as useful it had not been as useful as one would have liked because of the need being not as severe but now with the smartphone being a you know great enabler without the wi-fi without the computer necessary so the telemedicine became a favorite solution and because it was uh um not going to require the the a lot of equipment if at all we could do that the uh need because of the infancy of the of these telemedicine solutions that were available the uh there was a need to bring in ai basically to tell whether the at home when a patient is at home whether the patient is sick or not sick and if there is a reason to believe that things may get worse than to escalate the situation you know calling 9-1-1 or or sending them to er without doing that but that capability has to be there in order for telemedicine to be workable so that is the context in which our project was conceived our goal is to have ai algorithms that can estimate health parameters just like a staff in hospital would do to estimate the various vitals for example uh diagnose and as i just said because uh because prediction is critical if the person is at home then prediction capability is very important and of course these are all the tasks that uh are the objectives of ai if the person is to successfully stay at home and still not suffer the consequences of being not at the hospital now once you do that you don't have the doctor next to you you don't have all the other experts next to you so you want the system the algorithm system to be accurate to be consistent to be non-intrusive as nigel's talk uh emphasized this morning so it should be fast should be scalable and most of all easy to use by ordinary people and of course inexpensive by the same token which means it should be implementable on commodity hardware to minimize the cost and finally remote contact with the system should be easy it should be the whole thing should be communication friendly now initially it's okay if it only assists physicians and the mid-level providers and later on we can once trust is built once the system has been known to perform well has been seen to perform well then trust will come automatically and hopefully autonomy will follow so this is the context and our approach to doing this is to use a whole bunch of different sensors uh each being capable of doing a good job in on different aspect of the patient's health uh all those things that we listed earlier we would want uh this multi-sensor system to bring the strengths of different sensors to overcome their weaknesses so that the overall system does as desired so a targeted system is one that would monitor predict communicate all of this using commodity tools to to address those various aspects of cost and usability etc and what we have done in our last several months of work uh probably around starting around october um what we have done is divide it up divided the our project into two parts the first part is only vitals estimation meeting those requirements that i just listed and second is on diagnosis and prediction with that stage we are expecting to follow uh for now and our focus is on on vital's estimation and stage two will follow after that depend once we are done with stage one whenever that is hopefully by summer for validation you know the algorithms that we develop for validating them we will have vital signed monitors that are in emergency departments currently being used as the ground truth as doctors the way to find out and second there will be physicians to review it also of course let's see again yeah and then the deployment for testing purposes would be at the same places where the data are being collected namely university of illinois at chicago hospital and university of um illinois at uh university of chicago hospital those are the two places where the data are coming from as well as the validation and deployment will happen the specific signs that we are aiming for are heart rate respiration rate uh pulse oximeter temperature hydration level and hemoglobin level and also we are including in our preview the physical and neuro exam parts of physical and neural exam that is typically given to a patient so this is the system we are targeting our goal is to be able to do these things in a uh in a in a way that we i we just we just listed namely the system should have all those characteristics that will make it useful and practical so let me just quickly uh give you uh in a couple of minutes where things stand today not in our work but overall in the field uh and and i mean audio visual estimation because that's our uh our focus and really that is also has been the focus of almost all work so far a lot of work has been done on visual method based methods some on audio based methods uh and so here is just a quick review of where things stand so the four areas that we have reviewed before we start making our own contributions are heart rate estimation respiratory respiration rate estimation blood pressure and blood oxygen the methods that have been around that are around the literature is not very large maybe dozens of papers on a spectral base the color based analysis of the patient skin for example movement how things move whether it's because of blood or because of um respiration whatever the cause but motion is a cue for heart rate stability for respiration rate chest motion muscle motion color changes temperature changes for blood pressure color changes in skin when the blood flows wherever the blood pulse or burst is that's where the skin changes color uh how much oxygen is in the blood that also can is revealed in spectral characteristics and for example movement of eyes through the level presence of oxygen until you know it affects the movement of sclera so uh these are the the kinds of things that have been done in the visual cues area uh in there and of course we want to do it we also want to uh include audio features we have done a survey on audio features they have been used also in many of these things so so the the goal that we have set for ourselves is that we want to use all of these things combined in a manner that we take the strength of all and weaknesses to overcome the weaknesses of all so uh and this is in contrast with what has happened so far which is uh people have done x from y you know one particular vital or sometimes two vitals from a single modality and so we hope that by doing all of this we'll come up with a system that is that is a practical because it has all those properties that i mentioned now let me just give you some recent results we have not obtained uh publishable results yet or published publication level results so far from the work here and i'll tell you the details in a minute but what we have done is we have looked at what of our past work which aspects which projects in our past work can give us a head start namely which results that we have can be immediately used to to at least partly enable the objectives that we have set for ourselves so i'll go over the image and video and audio and machine learning and audio visual integration all are examples of these to give you a sense of uh what the problems involved are and what uh we have done that could be uh can can be used as a starting point so here is a uh an example of given an image not a video just an image and also video so this one is an image example where you can get hierarchical uh syntactic structure in it so that then you can use it for for example learning whether the image is uh you know complicated or complicated in structured and complicated in illumination for example the shadow here um estimation of 3d post sequence that is very important for exams so here for example we have a doctor showing a doctor a showing the uh gate that happens with parkinson's characteristic of parkinson's um it is you're given video frames you're given which are obviously 2d and you want to generate a 3d human post sequence because then you can reason about whether things are going okay or not so you can generate all these joint positions etc and from input you can get an output 3d pose shown here the input is this video and the output is shown through visualization uh in this right cube and hopefully since we know these parameters here on the right you can then relate it to the clinician's knowledge and see whether or model it or predict what is happening or classify whether this is serious or not here is an example of voice conversion where you can take a voice input voice and you can break it up into pieces and control them the pieces being timber pitch rhythm etc so you can get access to them individually which is to say that we can then in a similar way analyze sounds for for health analysis here we are talking about screen detection uh given an audio signal can you detect screens so again those things could be translated to cough and other respiration aspects of health here is an example of audio visual synthesis where we have input frames coming in input audio coming in and sync and suddenly the video stops but the audio continues can you use that audio to create the video frames knowing how they are related from before so in other words can you produce these following three frames that are shown below without seeing them because they were occluded or something so again the the the relationship to what our project wants to do is obvious finally on the machine learning side our focus of the work that we are doing uh and a lot of this should be available to report uh by summer uh we are working on an architecture that is uh interpretable this is learning architecture which is tractable which means analytic uh compact which means it can be embedded uh in edge hardware for example and all of this is done in a tractable optimal manner so that you can change it uh extend it analyze it as needed coming to patient data collection uh we have a total of 250 patients targeted at the two hospitals usc and uc hospitals emergency departments uh and they are mostly covered but we are not ruling other diseases out pulmonary sorry that should be pul um uh we have data collection script that we have debugged collectively we have come up with a method of exactly what we'll do it's about 30 35 minute long data collection script and the irb approvals are expected anytime now uh hopefully within the next 10 days and so that we can then actually start the data collection in the meantime we are doing things informally with uh you know of collecting our own data and seeing it for ourselves so there are no irb problems and finally what we expect in the next six months is a system that would have as much of what i mentioned earlier in our plans we will have that system we would be testing them on the patient data that we are collecting that we are going to be collecting soon by the end of the month hopefully and then we would have validated it using the ground troop that is available at the time of data collection and finally in the light of our experience here we hope that we'll have a nice framework for taking this forward to stage two which is of course diagnosis and and prediction so that brings me to the uh to the end of the overview great thank you very much arinda um we maybe have a time for one question if there's something i don't see any in the chat i have a quick question um sure so uh you talked about vital signs monitoring i think these days apple watch has a lot of these features i was just curious the audio visual uh measurement would be more accurate or could complement something like that or i guess it has more capabilities then yes yeah more capabilities there are because there is much more information in in in vision and sound and and these have been proven individually uh for example we developed an app that is now being used in number of hospitals all it does is uses cough and it's so effective that it has gone through political trials and it's now widely being used in hospitals because but it's only for cough to detect better dry cough now that's an important problem but that's one problem the question is can we use the visual and audio features which is of course what doctors do uh which is not to say that doctors do only those you know they use knowledge so can you synthesize the the vast knowledge that the doctors have and the video and audio cues hopefully they will be more than what your watch can hear when when it is listening to your wrist okay thank you one other comment related to that this is uh dave chester one of the co-eyes at uic i'm an emergency medicine doctor one of the things that we're finding with covet as well is that respiratory rate is a big predictor and that's something that you can't get from a wrist watch both at rest and also like walking so if we just have a patient walk in place for a minute which is actually part of our um our protocol here and then if the respiratory rate shoots up they're very deconditioned that's a huge predictor of hospital admissions so that's one more thing that a wristwatch wouldn't be able to get thank you