1 00:00:00,090 --> 00:00:02,055 - Hello everyone, this is Jacob Emerson 2 00:00:02,085 --> 00:00:03,255 with Becker's Healthcare. 3 00:00:03,395 --> 00:00:05,695 Thanks so much for tuning into the Becker's Healthcare 4 00:00:05,695 --> 00:00:08,695 podcast series where today I'm very pleased to be joined 5 00:00:08,715 --> 00:00:12,695 by Akiles Bpu, who is founder and CEO at Deep Scribe. 6 00:00:13,205 --> 00:00:15,695 He's here with me today to talk a little bit more about why 7 00:00:16,075 --> 00:00:18,295 AI note customization is key 8 00:00:18,595 --> 00:00:21,135 for widespread clinical adoption at large 9 00:00:21,135 --> 00:00:22,295 healthcare organizations. 10 00:00:22,515 --> 00:00:23,975 So Akiles, thank you so much 11 00:00:24,075 --> 00:00:25,615 for taking the time to be with us today. 12 00:00:26,535 --> 00:00:28,955 - Thanks, Jacob. I'm super glad to be here as well. 13 00:00:29,425 --> 00:00:30,675 - Yeah, we're glad to have you with us. 14 00:00:30,775 --> 00:00:32,235 And, and to get us started, 15 00:00:32,615 --> 00:00:34,275 I'm wondering if you could tell us a little bit more about 16 00:00:34,715 --> 00:00:36,515 yourself, your background in healthcare 17 00:00:36,855 --> 00:00:39,195 and what it is that you do today at Deep Scribe. 18 00:00:40,765 --> 00:00:43,455 - Yeah, absolutely. So prior to Deep Scribe, 19 00:00:43,935 --> 00:00:45,095 I was actually exposed to the problem 20 00:00:45,095 --> 00:00:46,935 of documentation at a pretty young age. 21 00:00:47,315 --> 00:00:48,575 You know, my dad was an oncologist 22 00:00:49,115 --> 00:00:50,695 and as a son of an oncologist, 23 00:00:51,095 --> 00:00:52,735 I actually would say I was desensitized 24 00:00:52,835 --> 00:00:53,935 to the problem of documentation. 25 00:00:54,315 --> 00:00:58,295 You know, he would spend upwards of half his day writing 26 00:00:58,825 --> 00:01:00,775 notes on everything his patient said to him, 27 00:01:01,435 --> 00:01:03,175 and a lot of that he would bring home. 28 00:01:03,675 --> 00:01:07,015 But as a kid, not much you can actually do about it. 29 00:01:07,235 --> 00:01:11,055 So I admitted defeat and a few years later, 30 00:01:11,355 --> 00:01:13,295 or roughly decade later, 31 00:01:14,055 --> 00:01:16,695 I eventually found myself at Bear, 32 00:01:16,945 --> 00:01:18,935 which is Berkeley's AI Research Lab. 33 00:01:19,435 --> 00:01:21,655 And I was working on summarization at the time. 34 00:01:22,355 --> 00:01:26,625 And you know, back in 2017, summarization 35 00:01:27,485 --> 00:01:32,025 was starting to do amazing things, including being able 36 00:01:32,025 --> 00:01:36,345 to distill one to two page CNN articles into nice, 37 00:01:36,455 --> 00:01:37,745 concise abstracts. 38 00:01:38,405 --> 00:01:42,785 It was pretty amazing. But every time I traveled back home, 39 00:01:43,365 --> 00:01:45,025 it was like I was going 10 years into the past. 40 00:01:45,405 --> 00:01:47,865 My dad was using the same tools, the same tech, 41 00:01:48,085 --> 00:01:49,785 and spending the same amount of time on documentation. 42 00:01:50,325 --> 00:01:52,265 And so I started to ask the question, you know, 43 00:01:52,925 --> 00:01:55,465 why not build a tool that could listen 44 00:01:55,465 --> 00:01:57,825 to my dad's conversations and summarize them 45 00:01:57,965 --> 00:01:59,665 and fill out his documentation for him? 46 00:02:00,205 --> 00:02:02,185 And if I were to do so, would 47 00:02:02,185 --> 00:02:03,905 that actually save him the time he 48 00:02:03,905 --> 00:02:04,945 was spending on documentation? 49 00:02:05,055 --> 00:02:06,705 Because obviously it meant a lot to me. 50 00:02:07,165 --> 00:02:09,545 So I took the same first 51 00:02:09,545 --> 00:02:11,145 principles approach we did in the lab. 52 00:02:11,985 --> 00:02:14,865 I tried to collect the data and then curate it 53 00:02:15,005 --> 00:02:16,185 and then train a model on it. 54 00:02:16,445 --> 00:02:18,705 But the data step is where I got stuck. 55 00:02:19,035 --> 00:02:22,265 There was no data set of patient physician conversations, 56 00:02:22,595 --> 00:02:23,745 there was no data set 57 00:02:23,805 --> 00:02:26,345 of the transcripts, the annotated notes. 58 00:02:26,645 --> 00:02:29,745 And so it was virtually impossible to train an AI to, 59 00:02:29,745 --> 00:02:32,145 to do this work at the bar of accuracy 60 00:02:32,145 --> 00:02:33,425 that's required in healthcare. 61 00:02:34,955 --> 00:02:36,455 But I think what was more interesting to us 62 00:02:36,515 --> 00:02:38,015 as technologists was that 63 00:02:38,665 --> 00:02:39,695 these conversations 64 00:02:40,005 --> 00:02:41,455 were actually healthcare source of truth. 65 00:02:42,315 --> 00:02:44,895 So every patient's story started from one 66 00:02:44,895 --> 00:02:46,735 of these interactions and no one was actually 67 00:02:46,735 --> 00:02:48,135 collecting this data. 68 00:02:48,715 --> 00:02:51,655 So, you know, we thought that if we could collect this data, 69 00:02:52,115 --> 00:02:53,855 not only could we solve documentation, 70 00:02:54,235 --> 00:02:55,695 but we could start to solve some 71 00:02:55,695 --> 00:02:56,855 of the larger problems in medicine. 72 00:02:57,235 --> 00:03:00,205 And so we started Deep Scribe to do exactly this. 73 00:03:00,665 --> 00:03:02,645 So we launched the product in 2020 74 00:03:02,985 --> 00:03:04,165 and we quickly became one 75 00:03:04,165 --> 00:03:05,845 of the market leaders in documentation. 76 00:03:06,385 --> 00:03:10,325 So today Deep Scribe is deployed in over a thousand provider 77 00:03:10,325 --> 00:03:12,885 organizations, everything from a private practice 78 00:03:13,145 --> 00:03:14,605 to a larger health system. 79 00:03:15,185 --> 00:03:18,045 And you know, on average we save about three hours a 80 00:03:18,045 --> 00:03:19,085 day for each of our doctors. 81 00:03:19,865 --> 00:03:22,405 But I think the most important thing to us is 82 00:03:22,405 --> 00:03:24,965 that we've been able to do this across a diversity 83 00:03:25,305 --> 00:03:26,645 of workflows and specialties. 84 00:03:27,065 --> 00:03:30,045 So when you see digital health products deployed, typically 85 00:03:30,605 --> 00:03:32,805 a lot of groups of clinicians are left out, you know, 86 00:03:32,805 --> 00:03:34,925 whether it's the specialists or whether it's the inpatient 87 00:03:35,035 --> 00:03:37,925 docs, because their workflows are highly nuanced 88 00:03:37,925 --> 00:03:41,045 and diverse to the point that the product can't fit them. 89 00:03:41,315 --> 00:03:44,925 However, with Deep Scribe, by deploying the primary care 90 00:03:45,025 --> 00:03:46,965 or private practices early on 91 00:03:47,305 --> 00:03:49,925 and seeing thousands of different workflow permutations, 92 00:03:50,345 --> 00:03:53,565 we were able to build what we call customization studio 93 00:03:53,795 --> 00:03:56,805 that allows any provider to fine tune deep scribe 94 00:03:56,945 --> 00:03:58,445 to fit whatever workload they had. 95 00:03:58,905 --> 00:04:01,685 So we've actually, by doing this, found 96 00:04:01,685 --> 00:04:04,885 that this is the single biggest key to adoption of any sort 97 00:04:04,885 --> 00:04:05,965 of AI in healthcare. 98 00:04:07,105 --> 00:04:09,565 - It sounds like this is a very personal passion for you, 99 00:04:09,745 --> 00:04:12,645 uh, on top of being obviously a professional one 100 00:04:12,645 --> 00:04:13,725 as well, Quiles? 101 00:04:14,535 --> 00:04:17,345 - Yeah, absolutely. I mean, I think one of my favorite parts 102 00:04:17,485 --> 00:04:19,585 of the whole Deep Scribe journey is 103 00:04:19,585 --> 00:04:22,465 that my dad actually uses Deep Scribe today, 104 00:04:22,685 --> 00:04:25,145 and then he's one of the clinicians that actually is able 105 00:04:25,145 --> 00:04:26,825 to save multiple hours in his 106 00:04:26,825 --> 00:04:28,025 day now because of Deep Scribe. 107 00:04:28,125 --> 00:04:29,585 So it came full circle and, 108 00:04:29,645 --> 00:04:32,345 and you know, I know I built it for my dad initially, 109 00:04:32,445 --> 00:04:35,265 but it's great to have other doctors be able to have access 110 00:04:35,265 --> 00:04:36,305 to this technology as well. 111 00:04:36,615 --> 00:04:38,145 - Sure. Yeah. No, it's very cool to hear. 112 00:04:38,325 --> 00:04:41,145 And I do want to get deeper into what's, 113 00:04:41,145 --> 00:04:43,465 what's the latest on clinical note taking, but, 114 00:04:43,485 --> 00:04:45,825 but I do wanna stay, take a step back for a second 115 00:04:45,885 --> 00:04:49,185 and ask you a little bit about a trend we've been seeing 116 00:04:49,185 --> 00:04:52,825 over the last few years, uh, just this rise in hospital 117 00:04:52,825 --> 00:04:53,945 and health system leaders, 118 00:04:54,455 --> 00:04:57,545 investments in artificial intelligence technology across the 119 00:04:57,545 --> 00:04:59,185 board, not just in clinical note taking. 120 00:04:59,415 --> 00:05:02,145 Yeah. So based on your experience ish, 121 00:05:02,205 --> 00:05:06,665 how are you seeing organizations deploy AI the most, um, or, 122 00:05:06,665 --> 00:05:08,905 or to support clinical documentation? 123 00:05:09,445 --> 00:05:12,545 Uh, are you seeing key challenges that they're looking 124 00:05:12,545 --> 00:05:13,825 to address ultimately? 125 00:05:15,125 --> 00:05:18,455 - Yeah, absolutely. So it's been amazing from our vantage 126 00:05:18,455 --> 00:05:22,615 point because when we started the company in 2018, we got 127 00:05:22,615 --> 00:05:25,535 to see sort of this adoption curve go from 128 00:05:26,515 --> 00:05:27,615 no one is really looking 129 00:05:28,035 --> 00:05:29,815 and talking about clinical documentation 130 00:05:30,075 --> 00:05:33,095 to now everybody is talking about clinical documentation. 131 00:05:33,275 --> 00:05:35,775 So in the beginning, you know, we actually made the decision 132 00:05:36,315 --> 00:05:39,375 to serve deep scribe to only private practices 133 00:05:39,405 --> 00:05:40,895 because the sales cycles 134 00:05:40,895 --> 00:05:42,415 and the implementation cycles 135 00:05:42,415 --> 00:05:44,815 that systems were just unsustainable for us 136 00:05:44,835 --> 00:05:46,015 as a, as a startup. 137 00:05:46,635 --> 00:05:49,495 And, you know, the key things there are just accuracy and, 138 00:05:49,495 --> 00:05:51,975 and proof points that tech would actually, you know, 139 00:05:51,975 --> 00:05:54,055 solve the challenges that the health systems are looking at. 140 00:05:54,315 --> 00:05:56,815 But over time, as we've built up a name 141 00:05:56,815 --> 00:05:58,605 for ourselves in the private practice industry, 142 00:05:58,855 --> 00:06:03,045 we've started to see that, you know, as the AI kind 143 00:06:03,045 --> 00:06:05,045 of thought space has evolved, a lot 144 00:06:05,045 --> 00:06:08,765 of health systems are now finally looking at, uh, AI 145 00:06:08,905 --> 00:06:10,485 as something that's here to stay. 146 00:06:10,665 --> 00:06:12,605 And so I think if you look at the numbers, you know, 147 00:06:12,605 --> 00:06:15,925 earlier this year there's a study done, um, that share 148 00:06:15,925 --> 00:06:17,245 that about 30% 149 00:06:17,585 --> 00:06:20,045 of systems are currently looking at documentation. 150 00:06:20,545 --> 00:06:22,005 And now I think, you know, 151 00:06:22,005 --> 00:06:24,725 the latest studies show something closer to 60%. 152 00:06:24,745 --> 00:06:27,725 So it's been amazing to see health systems kind 153 00:06:27,725 --> 00:06:28,965 of embrace AI. 154 00:06:29,505 --> 00:06:32,045 And you know, at the center of it, I think is the challenge 155 00:06:32,505 --> 00:06:33,965 of physician burnout. 156 00:06:34,345 --> 00:06:37,125 Um, I think that's what a lot of these systems are, 157 00:06:37,145 --> 00:06:38,365 are looking to solve with, 158 00:06:38,365 --> 00:06:40,125 with documentation tools like Deep Scribe. 159 00:06:40,385 --> 00:06:42,085 And I think that sits at the center 160 00:06:42,665 --> 00:06:45,485 of generative AI in healthcare for a couple reasons. 161 00:06:45,665 --> 00:06:48,405 You know, the first, I think there have been companies like 162 00:06:48,405 --> 00:06:50,285 Deep Scribe that have been building solutions 163 00:06:50,305 --> 00:06:51,845 and building the workflow enhancements 164 00:06:51,845 --> 00:06:55,085 and integrations, so that come time to implement, 165 00:06:55,195 --> 00:06:57,245 they're ready to go with all 166 00:06:57,245 --> 00:06:58,525 of the surrounding functionality. 167 00:06:58,585 --> 00:07:01,405 And now you, you have Cort Tech that's finally mature, 168 00:07:01,625 --> 00:07:04,325 so now you can actually deploy these mature solutions. 169 00:07:04,865 --> 00:07:07,645 But I think the second thing is that, um, it's really easy 170 00:07:07,645 --> 00:07:10,045 to see the value from these solutions. 171 00:07:10,145 --> 00:07:13,005 So, you know, with Deep Scribe, for example, as soon 172 00:07:13,005 --> 00:07:15,165 as it's implemented within a few hours, 173 00:07:15,505 --> 00:07:17,925 you're seeing clinicians actually start to save time 174 00:07:18,345 --> 00:07:21,605 and send, you know, little notes to administration that, 175 00:07:21,785 --> 00:07:25,325 you know, software IE deep scribe has been amazing 176 00:07:25,325 --> 00:07:27,205 because now they don't have to focus on notes anymore. 177 00:07:27,205 --> 00:07:28,605 They can now focus on their patient. 178 00:07:29,115 --> 00:07:32,445 Some physicians are that have been scaled down to just two 179 00:07:32,445 --> 00:07:35,485 to three days of, of practice from the full four 180 00:07:35,485 --> 00:07:37,525 or five are now going back to the four or five 181 00:07:37,525 --> 00:07:38,845 because of these documentation tools. 182 00:07:39,265 --> 00:07:42,445 So I think, you know, you're able to see instant value 183 00:07:42,825 --> 00:07:45,325 and, you know, I think the tech is mature enough for, 184 00:07:45,545 --> 00:07:46,605 for this specific space, 185 00:07:46,605 --> 00:07:49,565 which is why I think organizations see this as the, 186 00:07:49,745 --> 00:07:53,125 the most obvious first step in their generative AI journey. 187 00:07:54,355 --> 00:07:57,375 - And you mentioned a, a jump from 30 to 60% 188 00:07:57,675 --> 00:08:00,775 or so of systems now looking into AI 189 00:08:00,775 --> 00:08:02,375 powered clinical documentation. 190 00:08:03,045 --> 00:08:06,215 What advancements do you attribute that to in terms of 191 00:08:07,025 --> 00:08:08,575 where clinical note taking has 192 00:08:08,775 --> 00:08:09,975 improved over the last few years? 193 00:08:10,075 --> 00:08:13,495 And are, are you finding that customization fits in here? 194 00:08:13,755 --> 00:08:16,095 Um, why is all this significant, I guess is 195 00:08:16,095 --> 00:08:17,255 what I'm asking, Ailish? 196 00:08:18,465 --> 00:08:19,875 - Yeah, I think that's a good question. 197 00:08:20,215 --> 00:08:21,915 So, you know, since the beginning 198 00:08:22,255 --> 00:08:24,315 and still now, I think the challenges 199 00:08:24,315 --> 00:08:27,075 to adopting a technology like this have been the same. 200 00:08:27,335 --> 00:08:31,355 The two biggest are trust and customization 201 00:08:31,535 --> 00:08:34,435 and trust in our world is akin to accuracy. 202 00:08:34,735 --> 00:08:36,835 So if those are the problems, I think in terms 203 00:08:36,835 --> 00:08:40,795 of advancements, a lot of people see that the advancements 204 00:08:41,135 --> 00:08:45,595 for accuracy have been primarily tailored around GPD four 205 00:08:45,735 --> 00:08:47,435 and the advancements in foundation models. 206 00:08:47,775 --> 00:08:51,115 But we actually see that the biggest advancements were in 207 00:08:51,345 --> 00:08:52,555 open source lms. 208 00:08:52,895 --> 00:08:54,995 So, you know, while GPD four is incredible, 209 00:08:55,395 --> 00:08:56,405 without the right checks 210 00:08:56,425 --> 00:08:58,565 and balances, that mislabel a medication 211 00:08:58,825 --> 00:09:00,405 almost every single time. 212 00:09:00,905 --> 00:09:02,965 So this open source of movement has been amazing 213 00:09:02,965 --> 00:09:05,925 because as you collect a small amount 214 00:09:05,925 --> 00:09:06,965 of domain specific data 215 00:09:07,025 --> 00:09:08,245 and train these alums on it, 216 00:09:08,465 --> 00:09:10,805 you can quickly become on par with GPD four. 217 00:09:10,985 --> 00:09:13,725 But with large amounts like we have, you can actually 218 00:09:14,255 --> 00:09:17,125 solve problems like hallucinations or misinterpretation 219 00:09:17,225 --> 00:09:19,845 or other common mistakes fairly effectively. 220 00:09:20,185 --> 00:09:22,725 So I think that's the first big advancement is 221 00:09:22,725 --> 00:09:24,765 that trust has just come a long way 222 00:09:24,785 --> 00:09:26,445 and accuracy has come a long way to, 223 00:09:26,545 --> 00:09:27,565 to finally meet that bar. 224 00:09:28,065 --> 00:09:29,925 But the second thing that I think you mentioned was 225 00:09:30,445 --> 00:09:32,925 customizations, and I think that's probably 226 00:09:33,585 --> 00:09:35,005 the other large advancement. 227 00:09:35,185 --> 00:09:37,245 And so the product, in order 228 00:09:37,305 --> 00:09:40,965 to get full widespread adoption, can't just work for 229 00:09:41,195 --> 00:09:43,965 that one clinician that gets all their, uh, 230 00:09:43,965 --> 00:09:45,245 information from the conversation. 231 00:09:45,245 --> 00:09:46,965 They're okay with unstructured notes. 232 00:09:47,185 --> 00:09:49,365 It also has to work for, you know, the oncologist 233 00:09:49,435 --> 00:09:51,685 that gets most of their node from previous visits, 234 00:09:52,105 --> 00:09:56,045 or the highly specific geriatricians that want the node 235 00:09:56,045 --> 00:09:58,725 to follow exact structure that they've been using 236 00:09:58,785 --> 00:10:00,085 for the last like 10 years. 237 00:10:00,585 --> 00:10:03,805 So, um, I think the, the second big advancement 238 00:10:03,875 --> 00:10:05,205 that we've seen is, you know, 239 00:10:05,205 --> 00:10:07,685 given the advancements in foundation models 240 00:10:07,865 --> 00:10:11,165 and open source lms, we can now evolve that 241 00:10:11,505 --> 00:10:12,765 to customizations. 242 00:10:12,825 --> 00:10:14,365 And Deep Scribe has embraced that 243 00:10:14,385 --> 00:10:16,765 and released something called Customization Studio 244 00:10:17,195 --> 00:10:19,285 that is the culmination of all these little things 245 00:10:19,285 --> 00:10:20,605 that clinicians have asked us 246 00:10:20,625 --> 00:10:23,045 for over the last five years wrapped into this, 247 00:10:23,405 --> 00:10:24,565 I guess, seamless interface. 248 00:10:24,745 --> 00:10:27,805 And you can kind of think of it as clinicians now being able 249 00:10:27,805 --> 00:10:30,965 to build their own custom model in a few minutes. 250 00:10:31,265 --> 00:10:34,365 So when it comes to adoption, now you have trust 251 00:10:34,505 --> 00:10:37,245 that's relatively solved as well as customizations 252 00:10:37,265 --> 00:10:39,605 that's also solved that are the two keys 253 00:10:39,625 --> 00:10:41,685 to just rolling out something like deep scribe 254 00:10:41,685 --> 00:10:42,925 system wide. I 255 00:10:43,205 --> 00:10:45,605 - Wanna follow up with you on, on the trust aspect of this 256 00:10:45,605 --> 00:10:47,925 because I think when we're talking about, uh, 257 00:10:47,925 --> 00:10:50,765 implementing AI tools within healthcare, the other half 258 00:10:50,765 --> 00:10:52,925 of this conversation has to be around, uh, 259 00:10:52,925 --> 00:10:55,725 the responsibilities around using this technology. 260 00:10:56,185 --> 00:10:58,285 And we know the healthcare industry is working 261 00:10:58,305 --> 00:10:59,925 to set guidelines right now, 262 00:10:59,955 --> 00:11:01,885 this year they're being released around the use 263 00:11:01,945 --> 00:11:05,645 and development of these technologies given the, the, 264 00:11:05,745 --> 00:11:09,245 the range of concerns around accuracy, privacy, uh, 265 00:11:09,245 --> 00:11:12,285 potential biases and discriminatory concerns. 266 00:11:12,385 --> 00:11:15,765 And so I would ask you what considerations, 267 00:11:15,955 --> 00:11:18,845 what best practices do you recommend for, 268 00:11:18,905 --> 00:11:22,365 for leaders in charge of these tools as they explore, 269 00:11:22,625 --> 00:11:25,085 as they begin to implement these tools 270 00:11:25,265 --> 00:11:26,725 and like what you mentioned as they begin 271 00:11:26,725 --> 00:11:28,725 to customize these note taking 272 00:11:28,775 --> 00:11:30,125 tools that we're talking about? 273 00:11:31,565 --> 00:11:34,215 - Yeah, absolutely. I mean, I think there, 274 00:11:34,505 --> 00:11:37,055 there are three things that I would highlight 275 00:11:37,355 --> 00:11:40,895 as very important to, to look into and, and get right 276 00:11:40,895 --> 00:11:43,855 before adopting generative ai. 277 00:11:44,295 --> 00:11:47,175 I think the first thing is everybody should do their own 278 00:11:47,175 --> 00:11:49,815 benchmarks or at least develop their own 279 00:11:49,815 --> 00:11:50,935 benchmarking methodology. 280 00:11:51,355 --> 00:11:54,325 And in the, you know, AI documentation space, uh, 281 00:11:54,335 --> 00:11:55,725 especially, you know, 282 00:11:55,725 --> 00:11:58,765 you can plug in a conversation into GPD four 283 00:11:59,105 --> 00:12:00,485 and have a note that comes out 284 00:12:00,485 --> 00:12:02,125 that looks like it would work. 285 00:12:02,385 --> 00:12:03,685 But when you examine it closely, 286 00:12:04,025 --> 00:12:05,925 you'll see hallucinations happen. 287 00:12:06,225 --> 00:12:08,805 You'll see, uh, things like misinterpretation 288 00:12:08,905 --> 00:12:10,085 or miscategorizing 289 00:12:10,085 --> 00:12:12,325 or mislabel that can have, um, 290 00:12:12,325 --> 00:12:13,885 pretty significant downstream effects. 291 00:12:14,145 --> 00:12:17,485 So having a di having a benchmarking methodology would help 292 00:12:17,565 --> 00:12:19,085 a system, you know, prevent against that. 293 00:12:19,085 --> 00:12:21,165 And, and we have our own a deep describe that we're happy 294 00:12:21,165 --> 00:12:23,085 to share with anyone that's, that's looking at this. 295 00:12:23,145 --> 00:12:24,805 But I think that's the, the first thing 296 00:12:24,945 --> 00:12:26,045 to, to catch mistakes early. 297 00:12:26,425 --> 00:12:29,245 The second thing is human oversight. 298 00:12:29,505 --> 00:12:31,485 So, you know, and you can even do, you, 299 00:12:31,485 --> 00:12:33,525 you can do this even if the solution is fully automated. 300 00:12:33,665 --> 00:12:34,765 So deep describe, for example, 301 00:12:35,025 --> 00:12:36,365 our solution's fully automated, 302 00:12:36,545 --> 00:12:39,085 but we have a team of folks internally 303 00:12:39,555 --> 00:12:43,005 that review at random different notes to ensure that 304 00:12:43,535 --> 00:12:46,325 there are no hallucinations that are currently being missed 305 00:12:46,325 --> 00:12:49,365 by, uh, clinicians or misinterpretations 306 00:12:49,385 --> 00:12:50,405 or, or other errors. 307 00:12:50,665 --> 00:12:52,725 So that's, I think the second thing that's important. 308 00:12:53,105 --> 00:12:55,405 The third thing is automated oversight, 309 00:12:55,665 --> 00:12:56,925 and we've done some work here as well, 310 00:12:57,025 --> 00:13:00,285 but you can actually configure LLMs to be able 311 00:13:00,285 --> 00:13:02,805 to catch their own mistakes by having one 312 00:13:04,045 --> 00:13:07,845 specific dedicated reviewer, LLM to to review the notes, 313 00:13:07,845 --> 00:13:08,925 compare it with the transcript, 314 00:13:08,925 --> 00:13:10,805 and see if there's anything, uh, that's been missed. 315 00:13:10,825 --> 00:13:13,645 So I think those three things I I would say are, 316 00:13:13,785 --> 00:13:16,445 are probably the most important as it pertains to, 317 00:13:16,445 --> 00:13:18,765 to guidelines and, and, and safe AI development. 318 00:13:19,335 --> 00:13:21,065 - Yeah. Well, I appreciate you taking us through that 319 00:13:21,125 --> 00:13:22,225 and offering all those details. 320 00:13:22,845 --> 00:13:25,185 The last thing I wanted to ask you about achiles 321 00:13:25,485 --> 00:13:26,785 is on the ground. 322 00:13:27,265 --> 00:13:29,385 I know you, you mentioned your father 323 00:13:29,445 --> 00:13:32,465 and how he's been using this tool in oncology 324 00:13:32,605 --> 00:13:34,665 and some other physicians you know, as well. 325 00:13:34,805 --> 00:13:37,305 But can you share some, some on the ground examples 326 00:13:37,445 --> 00:13:40,265 or some case studies of where we are seeing hospitals 327 00:13:40,265 --> 00:13:44,825 and health systems deploy AI for customized medical notes 328 00:13:45,005 --> 00:13:47,225 and ultimately the outcomes that you, that you're seeing 329 00:13:47,225 --> 00:13:48,505 and that they are seeing from that? 330 00:13:50,365 --> 00:13:51,415 - Yeah, absolutely. 331 00:13:51,595 --> 00:13:55,335 So one of my favorite case studies 332 00:13:55,875 --> 00:13:57,535 is Covenant Healthcare. 333 00:13:57,715 --> 00:14:00,095 So Covenant Healthcare has used Deep Scribe 334 00:14:00,555 --> 00:14:02,975 for actually a little bit less than a year. 335 00:14:03,365 --> 00:14:04,375 They're on Epic. 336 00:14:05,035 --> 00:14:08,855 And, um, within six months, they were able 337 00:14:08,855 --> 00:14:11,215 to not just reduce documentation time 338 00:14:11,595 --> 00:14:15,295 by roughly 53% across all their providers, 339 00:14:15,795 --> 00:14:20,455 but they were able to see a 72% reduction in pajama time. 340 00:14:20,755 --> 00:14:21,855 And that's my favorite stat 341 00:14:21,855 --> 00:14:23,735 because, um, a lot of the work 342 00:14:23,735 --> 00:14:26,015 that clinicians do ends up being carried home 343 00:14:26,275 --> 00:14:28,415 and being able to reduce that by three fourths. 344 00:14:28,515 --> 00:14:30,415 So, you know, if they're spending an hour 345 00:14:30,515 --> 00:14:33,055 or a day on it in the evening, now they only have 346 00:14:33,055 --> 00:14:35,535 to spend 15 minutes, I think is a big, big win. 347 00:14:35,595 --> 00:14:38,415 Second big win is that, you know, even though we like 348 00:14:38,415 --> 00:14:41,815 to say we're able to save clinicians' documentation 80, uh, 349 00:14:41,815 --> 00:14:44,415 reduced documentation by 80% in practice 350 00:14:44,415 --> 00:14:46,295 because of the diversity of clinicians 351 00:14:46,295 --> 00:14:48,215 and their workflows, um, we've been able to show 352 00:14:48,215 --> 00:14:50,175 that we were able to do that by 53% 353 00:14:50,175 --> 00:14:51,600 consistently across everybody. 354 00:14:51,625 --> 00:14:54,845 So regardless of if you're, you know, the hyper picky doc 355 00:14:54,845 --> 00:14:56,685 that wants their documentation a certain way 356 00:14:57,025 --> 00:14:59,565 or specialist that has a highly configured workflow, 357 00:14:59,855 --> 00:15:02,285 we're able to make an impact, um, for, for everyone. 358 00:15:02,345 --> 00:15:04,325 So I think that's awesome to see in, 359 00:15:04,345 --> 00:15:07,005 in the last great tidbit that I wanted to insert there from 360 00:15:07,005 --> 00:15:10,485 that study is that they actually just in six months made the 361 00:15:10,485 --> 00:15:13,565 call to replace their virtual scribe program just months 362 00:15:13,565 --> 00:15:14,645 after implementing the scribe. 363 00:15:14,785 --> 00:15:17,485 So I think that's a great one that shows some 364 00:15:17,485 --> 00:15:18,885 of the value you can get from 365 00:15:18,975 --> 00:15:20,645 generative AI and documentation. 366 00:15:21,415 --> 00:15:23,465 - Yeah. And those are some incredible metrics you shared, 367 00:15:23,465 --> 00:15:26,625 especially that 53% reduction in documentation across 368 00:15:27,015 --> 00:15:29,105 specialties across all kinds of physicians. 369 00:15:29,105 --> 00:15:30,905 That's, that's amazing. Um, 370 00:15:31,325 --> 00:15:35,185 before we go though, Aish, any other final thoughts or, 371 00:15:35,185 --> 00:15:38,305 or key takeaways you would like to share with all 372 00:15:38,305 --> 00:15:41,905 of our health system listeners, uh, listening in today? 373 00:15:42,935 --> 00:15:46,305 - Yeah, absolutely. I mean, I think we're about to roll in 374 00:15:46,685 --> 00:15:50,145 to the biggest year for Gen AI in healthcare yet, 375 00:15:50,405 --> 00:15:53,265 and I think everybody is extremely excited, um, 376 00:15:53,265 --> 00:15:54,545 whether you're a startup 377 00:15:54,725 --> 00:15:56,985 or, uh, that's, that's building a solution for this 378 00:15:57,285 --> 00:15:59,665 or health system that that's, that's integrating it. 379 00:15:59,925 --> 00:16:03,025 But I think the two things to really look out for our trust, 380 00:16:03,485 --> 00:16:05,465 so, you know, how accurate is the ai, 381 00:16:05,935 --> 00:16:08,385 what benchmarking methodology does this company use in the 382 00:16:08,385 --> 00:16:10,025 backend to verify that, you know, 383 00:16:10,045 --> 00:16:12,625 in production it'll be reliable and and trustworthy. 384 00:16:12,885 --> 00:16:14,625 And the second thing is customization. 385 00:16:14,805 --> 00:16:18,585 So how can the AI adapt to all the, the big, I guess, 386 00:16:19,055 --> 00:16:21,985 diverse amount of workflows that your clinicians have 387 00:16:22,445 --> 00:16:24,665 and how does the product account for that? 388 00:16:24,765 --> 00:16:27,705 And I think those are gonna be the two most important things 389 00:16:28,045 --> 00:16:30,505 to get the adoption that everybody here is looking for. 390 00:16:30,805 --> 00:16:32,665 So, you know, we've done a lot of work on this and, 391 00:16:32,665 --> 00:16:34,705 and happy to share notes with anyone interested, 392 00:16:34,805 --> 00:16:36,585 but I think those are the two things 393 00:16:36,585 --> 00:16:38,505 that we predict are gonna be very important. 394 00:16:38,565 --> 00:16:42,865 And so I'm super excited to, to see this year play out and, 395 00:16:42,925 --> 00:16:47,305 and see just how much AI is gonna really impact and, 396 00:16:47,325 --> 00:16:49,225 and improve healthcare outcomes. 397 00:16:50,385 --> 00:16:52,595 - Fantastic. Well, Aish, thank you so much 398 00:16:52,695 --> 00:16:54,355 for taking the time to be with us 399 00:16:54,455 --> 00:16:55,875 and for sharing your insights 400 00:16:55,945 --> 00:16:57,515 with us. We really appreciate it. 401 00:16:58,065 --> 00:16:59,475 - Yeah, thanks for having me. Jacob. 402 00:17:00,305 --> 00:17:03,555 - We'd also like to thank our podcast sponsor, deep scribe, 403 00:17:03,815 --> 00:17:05,955 you continue to more podcasts from Becker's Healthcare 404 00:17:06,295 --> 00:17:09,195 by visiting our podcast page at becker's hospital 405 00:17:09,215 --> 00:17:10,195 review.com.