Okay, so I thought the way we would walk through this is is look at some questions to answer. So the first is is why efficiency. How to measure efficiency. What are the AI options to optimize performance. What is the ideal throughput. What models can optimize report generation and revenue cycling. And are there additional actions or AI solutions. So as you all know, healthcare costs continue to rise in the United States, annual expenditures account for approximately 136 billion US dollars, and we've seen endoscopic procedures increasing by volume, year over year. When it comes to quality indicators this is really gained traction over the last two decades because of alarming medical error rates and high degrees of variation, and the quality of healthcare can essentially be measured by comparing performance with benchmark structural measures process measures outcome measures. And so the question comes, why efficiency. Well efficiency is really defined as the use of resources in such a way as to maximize the production of goods and services, and the National Academy of Science and Institute on medicine defined efficiency as a key element of quality care delivery with a continuous decrease in waste, and we've seen that improvements in efficiency and essentially enhance patient experience enhance employee workspace satisfaction, allow for more patient access to necessary endoscopic procedures. And so the challenge really is is there's always an appropriate limited use of appropriate resources that we've seen, especially in academic medical centers, where delays can be up to 30 to 50% of the time. And so, how do we measure efficiency. Well this can be through patient preparation procedural time and patient recovery and delays can happen in the ways of patients being late, or patients are appropriately prepared staff room unavailable recovery unit is full, or the there's a delay in discharge by the providers. And so these metrics can really be drilled down to reception time which is really our time from reception to pre op. The time to wheels in which is oftentimes our first case on time start at the beginning of the day, our anesthesia time from anesthesia start to anesthesia ready our scope time from scope in to scope out. And then ultimately our turnaround time which is our wheels out to our next wheel in, and our discharge time. And so when we've looked at our own data what we've seen is is that there's oftentimes error rates that occur when with manual entry with up to 19% missing at least one time standpoint and up to 50% of the time that time standpoints are inconsistent they don't they don't make sense they don't follow appropriate chronologic order. And so the question then becomes is, how do we measure efficiency, excuse me, how do what are the AI options to optimize our performance. And so, several models are now being developed to essentially use unbiased workflow to measure what's happening in the OR suite. And so this is through sensors to look at patient location bed location scope utilization and different types of air particle monitoring even. And this can be integrated with endoscopy systems and integrated with the EHR. And so we utilize such a system in our in our center, and essentially you can see where these units are mounted and can appropriately measure what's happening at the time of the endoscopy procedure in the pre op space. And it's provides a more fact driven opportunity and it's built on deep learning and realistic computer reasoning for decision making. There's multiple data capturing methods and it essentially creates a digital twin of the endoscopy suite. And so with this model we can then create outputs of our occupancy turnover time and also get all our metrics in terms of optimizing the procedural efficiency. And so what we found was when we utilize such a system we had an increase in our ability to perform volume with a reduction in with a continued reduction in our procedural end time and you can see before we utilize that model what we found was is that we weren't really able to increase our capacity through the unit despite being able to continuously improve through the end of the day. And so, I think that what and just to add to that a little bit further, what we found was that we were able to save about 1.3 staff hours per day per procedure room, and this increased the capacity of usage of the room by about 9%. And so, in essence, you know, this is these are the types of systems that we're now seeing developed to help optimize efficiency. Now again data is really lacking on this in terms of comparison to improved education or what it essentially does in terms of cost savings or time savings and multi center evaluations of this but in essence, these are some of the models that we're seeing that are around efficiency. What other questions were we going to answer well what models can optimize report generation and revenue cycling. And so automating ADR is one of these key methods. There's an inverse association as you know, between ADR and interval colorectal cancer And the Stanford currently exists of carefully reviewing all medical records through manual entry, natural language processing models allow us to use computers to essentially search out key terms, match them to medical records and essentially make recommendations beyond that. And so the way we weigh this these models performances in terms of creating identifying first screen colonoscopies identifying corresponding path reports, and the data, then can be presented and provide an outcomes report. And so with this data extraction, this can, this really requires matching procedures and pathology dates identifying previous procedure reports to ensure that it's a first screening colonoscopy. But in essence, you know what we've seen in some of these models is that this one by Roger and colleagues, essentially the natural language processing model performed on par or better than manual entry in terms of identifying screening colonoscopies identify measuring adenoma detection rate and identifying SSAs as well. Additionally with natural language processing we're seeing voice annotation as an emerging area as well. We all know that post procedure documenting procedural report or the outcomes of the procedures time consuming. It's flawed by distractions and there's oftentimes incomplete or inaccurate documentation that happens with this. While real time voice annotation allows us to convert to text and provide smart responses potentially for coding and optimal documentation that can be delivered to endo writers. And this is something that multiple companies are working on right now, but in essence, the endoscopies is wearing this form of a headset to really document as they go what they're finding Paul up 10 millimeters at the hepatic flexure removed cold snare. And in essence, then that voice annotation gets converted to text and auto populates a procedure report that they can come over just review and hit return and save several minutes and potential inaccuracies that are associated with delay documentation. Auto documentation is actually another area that we're seeing we're utilizing convolutional neural networks to make documentation without the endoscopies input altogether. And here's just an example of a sequel intubation that's being been been cited. So in essence, SQL detection, as you know, reduces post colonoscopy, colorectal cancer development, the sequel intubation rates in the literature range from anywhere from 58.8 to 100% with photo documentation varying from 6% to 81%. There's few automated means to really record and confirm colonoscopy completion to date, but this can provide a mechanism for automated withdrawal time calculations. I'm going to skip to this slide, and we'll go back to the other side, my apologies. But in essence, this was a study that looked at 11,900 images performed by low end colleagues for training and validating a convolutional neural network to differentiate appendiceal orifice recognition from non appendiceal orifice images. And what they were able to demonstrate was a very high sensitivity, specificity, positive predictive value and negative predictive value with respect to appendiceal orifice documentation or identification with the AI model. They also were able to show good performance when it came to varying degrees of bowel preparation with this. Now, the model didn't perform as well when the preparations were poor, in essence, to determine non appendiceal orifice images, but did do well in demonstrating in high bowel prep images. So in essence, this model created an accuracy of 94% with an area under the receiver operating curve for this neural network of 0.98. And I think this is a good example of how we're seeing areas of the ability to use artificial intelligence to document some of these findings at colonoscopy. Another area where we're seeing auto documentation become more prevalent is bowel preparation scoring. You know, Boston Bioprep scale still remains our most validated means of determining the quality of the bowel preparation. There are limitations, however, in terms of variability exists between providers in terms of BBPS reporting. Those procedure reporting often, again, is introduced by inaccuracies and distractions and bias may exist in our reporting from how long it takes to perform the procedure itself. And so several groups have evaluated convolutional neural networks to determine adequacy of bowel prep and auto populate or auto document this finding. One group, Lee and colleagues, looked at defining inadequate bowel preparations if the images showed a Boston Bioprep scale of 0 to 1 and adequate bowel preparation scale if they were 2 to 3. And essentially, they looked at 200 colonoscopies in the training set with over 73,000 images and then looked at the validation set of 30 colonoscopies with 281 10-second video clips. We then looked at 10 in the test set of 10 withdrawal colonoscopies and 30 10-second video clips from the right colon, left colon, and transverse colon as well. And what they were able to demonstrate was the model was able to distinguish adequate from inadequate bowel preparation with an area under the curve of 0.91. And importantly, it was able to document an inadequate bowel prep 100% of the time. And so I think, again, this demonstrates more as to how we're seeing quality and auto documentation of artificial intelligence models to help us in demonstrating the quality of the colonoscopy exam. So in conclusion, efficiency is a key element of quality and provides increased access as well as patient and provider satisfaction. AI can provide an opportunity to measure current workflow to better determine deficiencies in flow. Advances in automated bowel preparation and SQL detection allow for improved standardization of endoscope completion. And natural language processing models perform on par or better than manual entry for the purpose of report generation. Thank you for your time.

efficiency in healthcare

artificial intelligence

patient location tracking

procedural reports generation

standardizing documentation