1 00:00:00,080 --> 00:00:02,720 Hello, everyone. This is Erica Spicer Mason with 2 00:00:02,720 --> 00:00:05,120 Becker's Healthcare. Thank you so much for tuning 3 00:00:05,120 --> 00:00:07,379 into the Becker's Healthcare podcast series. 4 00:00:08,000 --> 00:00:09,919 So today, we're going to talk about how 5 00:00:09,919 --> 00:00:13,460 small domain specific language models are transforming healthcare, 6 00:00:14,064 --> 00:00:16,085 unlocking new opportunities for innovation, 7 00:00:16,385 --> 00:00:18,885 operational efficiency, and data driven applications 8 00:00:19,265 --> 00:00:21,285 across the payer and provider ecosystem. 9 00:00:21,904 --> 00:00:23,824 And joining me for this conversation, we have 10 00:00:23,824 --> 00:00:26,884 with us Gene German, the chief technology officer 11 00:00:26,945 --> 00:00:27,684 at Lyric. 12 00:00:28,269 --> 00:00:29,870 Gene, welcome to the podcast. It's so great 13 00:00:29,870 --> 00:00:31,010 to have you with us today. 14 00:00:31,870 --> 00:00:34,750 Hi. Thanks for having me. I'm Gene. I'm 15 00:00:34,750 --> 00:00:37,570 the chief technology officer for Lyric. I've been 16 00:00:37,630 --> 00:00:40,530 with Lyric for approximately three years. 17 00:00:41,265 --> 00:00:42,085 Prior to, 18 00:00:42,704 --> 00:00:44,784 joining the company, it's been in health care 19 00:00:44,784 --> 00:00:47,505 for another, seven years focused on the consumer 20 00:00:47,505 --> 00:00:48,005 tech. 21 00:00:48,465 --> 00:00:51,925 When I joined Lyric, the company already 22 00:00:52,385 --> 00:00:53,365 had decades 23 00:00:53,745 --> 00:00:56,304 of service to our clients and has built 24 00:00:56,304 --> 00:00:56,729 trust, 25 00:00:57,130 --> 00:00:59,789 over these decades through delivering of value, 26 00:01:00,890 --> 00:01:01,869 to those clients, 27 00:01:02,649 --> 00:01:03,710 you know, through our products. 28 00:01:04,409 --> 00:01:06,890 And so, we already had, like, a really 29 00:01:06,890 --> 00:01:08,829 good place to start at, 30 00:01:09,204 --> 00:01:11,604 But there was a challenge, that we faced, 31 00:01:11,604 --> 00:01:14,405 and that challenge was the bulk of our 32 00:01:14,405 --> 00:01:14,905 software 33 00:01:15,525 --> 00:01:17,625 was hosted on prem, 34 00:01:18,004 --> 00:01:18,984 in the customer's, 35 00:01:19,604 --> 00:01:20,424 data centers. 36 00:01:20,884 --> 00:01:23,629 And not only is it difficult to deliver 37 00:01:23,689 --> 00:01:24,430 AI solutions, 38 00:01:24,969 --> 00:01:25,869 when your software 39 00:01:26,250 --> 00:01:29,789 is hosted on prem, but also any other 40 00:01:29,849 --> 00:01:30,349 capabilities, 41 00:01:31,450 --> 00:01:33,310 that we wanted to give to our customers, 42 00:01:33,530 --> 00:01:35,390 because, typically, upgrade cycles, 43 00:01:36,424 --> 00:01:38,424 happen every two to three years, for our 44 00:01:38,424 --> 00:01:39,884 software in the past. 45 00:01:40,424 --> 00:01:44,204 And so we've built our platform, Lyric 42, 46 00:01:44,504 --> 00:01:46,185 that gave us kind of the pipes and 47 00:01:46,185 --> 00:01:46,765 the capability, 48 00:01:47,864 --> 00:01:48,604 to deliver, 49 00:01:49,545 --> 00:01:50,524 multiple products, 50 00:01:50,825 --> 00:01:53,519 across a whole spectrum of payment integrity, 51 00:01:53,920 --> 00:01:55,859 solutions, not just, in primary, 52 00:01:56,640 --> 00:01:58,899 but also our secondary content. 53 00:01:59,679 --> 00:02:01,599 We've acquired a company called, 54 00:02:02,000 --> 00:02:02,500 ClaimShark, 55 00:02:03,280 --> 00:02:04,899 with Replay and, 56 00:02:05,359 --> 00:02:05,859 Virtusa, 57 00:02:06,775 --> 00:02:09,094 replay platform and Virtusa products on top of 58 00:02:09,094 --> 00:02:09,594 it 59 00:02:09,974 --> 00:02:10,634 as well, 60 00:02:10,935 --> 00:02:13,735 which allowed us to deliver manual reviews. None 61 00:02:13,735 --> 00:02:15,814 of that will be possible with our without 62 00:02:15,814 --> 00:02:17,514 our Lyric 42 platform. 63 00:02:19,150 --> 00:02:20,750 Well, Jean, that's great to learn more about 64 00:02:20,750 --> 00:02:22,930 you and also Lyric's platform. 65 00:02:23,389 --> 00:02:26,270 Great context setting for this conversation because we 66 00:02:26,270 --> 00:02:27,090 are gonna go, 67 00:02:27,469 --> 00:02:28,909 a little bit deeper on some of the 68 00:02:28,909 --> 00:02:31,389 trends that we're seeing here with small language 69 00:02:31,389 --> 00:02:32,689 models across the industry. 70 00:02:33,664 --> 00:02:35,364 We know that SLMs are enabling 71 00:02:35,824 --> 00:02:38,784 really this new wave of purpose built domain 72 00:02:38,784 --> 00:02:39,764 specific applications 73 00:02:40,144 --> 00:02:43,104 across industries and including health care. So from 74 00:02:43,104 --> 00:02:45,905 your perspective, what makes health care uniquely positioned 75 00:02:45,905 --> 00:02:48,324 to apply these tools effectively right now? 76 00:02:49,099 --> 00:02:49,599 Well, 77 00:02:49,900 --> 00:02:50,879 AI generally 78 00:02:51,579 --> 00:02:53,599 can be very useful in working 79 00:02:53,979 --> 00:02:54,479 with, 80 00:02:55,739 --> 00:02:57,120 very large datasets. 81 00:02:57,900 --> 00:03:01,180 And AI can certainly drive innovation. And in 82 00:03:01,180 --> 00:03:01,919 health care, 83 00:03:02,534 --> 00:03:04,055 there's plenty of that. There's, 84 00:03:04,534 --> 00:03:05,034 datasets 85 00:03:05,334 --> 00:03:06,074 of claims, 86 00:03:06,534 --> 00:03:08,794 member data, provider data, eligibility, 87 00:03:09,495 --> 00:03:11,334 and so on and so forth. And that 88 00:03:11,334 --> 00:03:13,435 data comes from various different sources. 89 00:03:14,215 --> 00:03:17,010 It varies in formats. It varies in levels 90 00:03:17,010 --> 00:03:17,830 of completion 91 00:03:18,530 --> 00:03:19,990 and many other attributes. 92 00:03:20,530 --> 00:03:21,349 And so 93 00:03:21,650 --> 00:03:24,290 we see that LLMs are, very good at 94 00:03:24,290 --> 00:03:25,590 dealing with this, 95 00:03:25,969 --> 00:03:26,870 high textual 96 00:03:27,170 --> 00:03:27,670 complexity. 97 00:03:28,544 --> 00:03:30,784 We also see that the small language models, 98 00:03:30,784 --> 00:03:32,645 which Lyric tends to use, 99 00:03:33,264 --> 00:03:35,504 are at least as good or better than 100 00:03:35,504 --> 00:03:36,004 large, 101 00:03:36,305 --> 00:03:37,205 language models 102 00:03:37,504 --> 00:03:38,004 in, 103 00:03:38,465 --> 00:03:40,625 solving problems when those, 104 00:03:41,300 --> 00:03:44,599 small language models are grounded in specific domains. 105 00:03:45,300 --> 00:03:47,000 We see them being cheaper, 106 00:03:47,620 --> 00:03:49,620 which is as important as being faster or 107 00:03:49,620 --> 00:03:52,099 more accurate when grounding them in the specific 108 00:03:52,099 --> 00:03:52,919 domain knowledge. 109 00:03:53,844 --> 00:03:55,605 Jean, thanks so much. It's been really helpful 110 00:03:55,605 --> 00:03:57,145 getting your perspective on 111 00:03:57,685 --> 00:03:59,605 what makes health care kind of unique in 112 00:03:59,605 --> 00:04:02,165 this space as they're apply as leaders are 113 00:04:02,165 --> 00:04:03,064 applying tools. 114 00:04:03,685 --> 00:04:06,325 At the same time, what unique challenges do 115 00:04:06,325 --> 00:04:09,145 these technologies introduce for health IT systems? 116 00:04:10,379 --> 00:04:10,860 Sure. 117 00:04:11,259 --> 00:04:13,500 Well, being in the health care industry for 118 00:04:13,500 --> 00:04:13,979 the past, 119 00:04:14,539 --> 00:04:15,599 decade or so, 120 00:04:15,900 --> 00:04:17,279 we typically see, 121 00:04:17,899 --> 00:04:19,180 health care IT comp 122 00:04:19,819 --> 00:04:20,319 groups, 123 00:04:21,180 --> 00:04:22,480 focus on deterministic, 124 00:04:23,354 --> 00:04:25,754 type of algorithms. So typically, for a given 125 00:04:25,754 --> 00:04:26,574 set of inputs, 126 00:04:27,354 --> 00:04:29,194 there's a very predictable way to see a 127 00:04:29,194 --> 00:04:30,334 given set of outputs. 128 00:04:31,035 --> 00:04:32,095 And LLMs, 129 00:04:32,954 --> 00:04:33,694 by definition, 130 00:04:34,154 --> 00:04:36,175 they're much more nondeterministic. 131 00:04:37,274 --> 00:04:38,014 And so 132 00:04:38,419 --> 00:04:40,819 it is very important as you think of 133 00:04:40,819 --> 00:04:42,500 the use cases to which you're gonna be 134 00:04:42,500 --> 00:04:44,819 applying AI that those use cases kind of 135 00:04:44,819 --> 00:04:48,040 combine this concept of determinism and, nondeterminism, 136 00:04:48,579 --> 00:04:50,500 and they deal with a very high cost 137 00:04:50,500 --> 00:04:52,040 in health care of failure. 138 00:04:52,894 --> 00:04:55,055 And, typically, that involves having humans in the 139 00:04:55,055 --> 00:04:55,555 middle, 140 00:04:56,014 --> 00:04:58,975 having well defined workflows in which the models 141 00:04:58,975 --> 00:05:01,055 are trained and given feedback as well as 142 00:05:01,055 --> 00:05:01,555 validated, 143 00:05:02,495 --> 00:05:03,875 and so on and so forth. 144 00:05:04,254 --> 00:05:04,754 Mhmm. 145 00:05:05,389 --> 00:05:08,269 Yeah. And I know another important piece of 146 00:05:08,269 --> 00:05:09,329 driving innovation 147 00:05:09,629 --> 00:05:12,449 or at least driving it successfully is data. 148 00:05:13,149 --> 00:05:15,949 So how should organizations think about leveraging their 149 00:05:15,949 --> 00:05:19,230 own data to build differentiated applications on top 150 00:05:19,230 --> 00:05:21,169 of emerging SLM platforms? 151 00:05:22,245 --> 00:05:24,884 That's a great question. Data is indeed very 152 00:05:24,884 --> 00:05:27,845 critical in driving innovation, and so it is 153 00:05:27,845 --> 00:05:29,545 very important, first and foremost, 154 00:05:30,564 --> 00:05:33,045 the the design tracking and usage of data 155 00:05:33,045 --> 00:05:34,264 is not an afterthought. 156 00:05:34,725 --> 00:05:37,020 There's many different kinds of data in health 157 00:05:37,020 --> 00:05:38,080 care as I mentioned, 158 00:05:38,460 --> 00:05:40,540 but also, you know, some data, like you 159 00:05:40,540 --> 00:05:41,920 said, is generated 160 00:05:42,460 --> 00:05:44,240 by the health health care organization 161 00:05:44,620 --> 00:05:45,120 itself, 162 00:05:45,660 --> 00:05:48,080 but some data is received from, 163 00:05:48,620 --> 00:05:50,879 members or from providers. 164 00:05:51,524 --> 00:05:53,125 And so all of that data can be 165 00:05:53,125 --> 00:05:54,024 used to drive, 166 00:05:55,125 --> 00:05:57,625 improve areas such as efficiency, quality, 167 00:05:58,165 --> 00:05:59,625 outcomes, and speed. 168 00:06:00,805 --> 00:06:02,725 The key here is to make sure that 169 00:06:02,725 --> 00:06:04,024 you focus on the goals, 170 00:06:04,564 --> 00:06:06,745 upfront, and you define the use cases 171 00:06:07,139 --> 00:06:09,220 around those goals to make sure that everything 172 00:06:09,220 --> 00:06:10,360 from data rights 173 00:06:10,660 --> 00:06:11,639 to data pipelines, 174 00:06:12,300 --> 00:06:15,220 is, thought through clearly in the beginning. And 175 00:06:15,220 --> 00:06:17,479 you bring in your, you know, clients, 176 00:06:18,180 --> 00:06:20,654 along with you around the value that is 177 00:06:20,654 --> 00:06:22,194 being driven and the usage, 178 00:06:22,654 --> 00:06:23,474 of data. 179 00:06:24,254 --> 00:06:26,495 And it's true for your own data as 180 00:06:26,495 --> 00:06:26,995 well. 181 00:06:27,615 --> 00:06:30,754 When you combine AI technology and people together, 182 00:06:30,974 --> 00:06:33,454 you can achieve very many different goals leveraging 183 00:06:33,454 --> 00:06:34,354 this new tech. 184 00:06:35,120 --> 00:06:36,720 Yeah. And, Gina, I've heard you emphasize a 185 00:06:36,720 --> 00:06:39,120 few times now the importance of getting really 186 00:06:39,120 --> 00:06:41,839 clear on those use cases. So I appreciate 187 00:06:41,839 --> 00:06:42,240 you, 188 00:06:42,720 --> 00:06:45,139 underscoring that as a consideration for our listeners. 189 00:06:46,399 --> 00:06:48,480 And I know at Lyric, you focused on 190 00:06:48,480 --> 00:06:51,655 payment integrity and claims analytics, which you mentioned 191 00:06:51,655 --> 00:06:53,435 in some of your introductory remarks. 192 00:06:53,895 --> 00:06:56,134 Where are you seeing the most meaningful efficiency 193 00:06:56,134 --> 00:06:59,254 gains today across the payer provider ecosystem in 194 00:06:59,254 --> 00:06:59,915 this realm? 195 00:07:01,095 --> 00:07:01,574 Yes. 196 00:07:01,895 --> 00:07:02,634 For us, 197 00:07:03,014 --> 00:07:03,959 because we've 198 00:07:04,839 --> 00:07:07,240 deployed Lyric 42, and there's a very high 199 00:07:07,240 --> 00:07:09,740 adoption from our client base, 200 00:07:10,759 --> 00:07:13,099 across the client base of our platform, 201 00:07:13,479 --> 00:07:15,079 it opens up a whole new set of 202 00:07:15,079 --> 00:07:16,680 avenues. So we could have done a lot 203 00:07:16,680 --> 00:07:18,779 of different things, but we've decided 204 00:07:19,324 --> 00:07:21,584 to focus on four key areas, 205 00:07:22,044 --> 00:07:24,225 for us. So one is the 206 00:07:24,685 --> 00:07:28,384 ideation building and validate validating the payment integrity, 207 00:07:29,084 --> 00:07:31,165 content. So this is our bread and butter, 208 00:07:31,165 --> 00:07:33,104 both primary and secondary content. 209 00:07:33,740 --> 00:07:35,900 Second is the adoption of that content. Once 210 00:07:35,900 --> 00:07:38,000 that content is built and we've seen significant 211 00:07:38,139 --> 00:07:40,240 increases in how quickly and accurately, 212 00:07:40,860 --> 00:07:42,639 we can deliver that content leveraging, 213 00:07:43,340 --> 00:07:44,080 AI technology, 214 00:07:44,939 --> 00:07:47,279 it's the adoption and deciding how to prioritize 215 00:07:47,500 --> 00:07:48,000 it 216 00:07:49,454 --> 00:07:52,094 with our clients and figuring out which content 217 00:07:52,094 --> 00:07:54,495 will cause more abrasion, less abrasion, and, you 218 00:07:54,495 --> 00:07:56,175 know, focusing on the holistic, 219 00:07:56,495 --> 00:07:57,474 health care system. 220 00:07:58,094 --> 00:08:00,194 Three, it is, like I mentioned, 221 00:08:01,134 --> 00:08:02,370 a replay product, 222 00:08:02,769 --> 00:08:05,569 which is a manual review product integrated with 223 00:08:05,569 --> 00:08:07,430 Lyric 42, our platform. 224 00:08:08,529 --> 00:08:10,149 Now that we're able to, 225 00:08:10,850 --> 00:08:13,029 roll it out much quicker to our customers, 226 00:08:13,329 --> 00:08:14,229 this new product, 227 00:08:14,915 --> 00:08:16,855 We are leveraging AI, 228 00:08:17,394 --> 00:08:19,095 for prioritization automation 229 00:08:19,555 --> 00:08:20,055 and 230 00:08:20,675 --> 00:08:21,254 the summary 231 00:08:21,634 --> 00:08:22,134 and, 232 00:08:23,074 --> 00:08:25,654 search of very complex information that's required, 233 00:08:26,035 --> 00:08:27,735 for those reviews, whether it's 234 00:08:28,209 --> 00:08:30,769 information related to eligibility, like, for coordination of 235 00:08:30,769 --> 00:08:34,209 benefits or itemized bill review or pulling a 236 00:08:34,209 --> 00:08:34,709 chart. 237 00:08:35,409 --> 00:08:37,970 AI is definitely helpful in putting all those 238 00:08:37,970 --> 00:08:39,330 tools into the fingerprint 239 00:08:39,889 --> 00:08:40,389 fingertips 240 00:08:41,089 --> 00:08:43,830 of people that are leveraging our software. 241 00:08:44,184 --> 00:08:45,485 And then fourth, 242 00:08:46,105 --> 00:08:47,725 last but not least is, 243 00:08:48,425 --> 00:08:51,305 leveraging our Lyric 42 pipes, and the learnings 244 00:08:51,305 --> 00:08:52,764 that we get, to 245 00:08:53,225 --> 00:08:55,465 create new products and test those products for 246 00:08:55,465 --> 00:08:57,945 our customers because our our clients are constantly 247 00:08:57,945 --> 00:08:58,764 asking us, 248 00:08:59,250 --> 00:09:01,590 for new, products that improved, 249 00:09:02,769 --> 00:09:03,910 their claims flow. 250 00:09:05,490 --> 00:09:07,970 Yeah, Jean. Again, thank you for for that 251 00:09:07,970 --> 00:09:10,769 helpful overview. And I'm curious as you're helping 252 00:09:10,769 --> 00:09:11,269 organizations 253 00:09:11,730 --> 00:09:13,350 roll out some of these tools 254 00:09:13,715 --> 00:09:16,514 in the payment integrity and claims space, have 255 00:09:16,514 --> 00:09:18,754 there been any outcomes that have surprised you 256 00:09:18,754 --> 00:09:20,915 or maybe even key lessons that you think 257 00:09:20,915 --> 00:09:22,615 leaders should take away from that work? 258 00:09:23,634 --> 00:09:26,355 Yes. You know, there are certain areas we 259 00:09:26,355 --> 00:09:26,855 expected, 260 00:09:27,715 --> 00:09:29,254 to improve such as 261 00:09:29,870 --> 00:09:32,669 speed, accuracy, and the outcomes and the value 262 00:09:32,669 --> 00:09:34,610 prop our technology is driving. 263 00:09:35,309 --> 00:09:37,549 What one area that we're kind of surprised 264 00:09:37,549 --> 00:09:39,409 is how much how much quality, 265 00:09:40,429 --> 00:09:42,049 we can drive with AI, 266 00:09:42,589 --> 00:09:45,095 even on top of already, like, high quality 267 00:09:45,095 --> 00:09:47,254 products. So not only are we accelerating the 268 00:09:47,254 --> 00:09:47,754 throughput, 269 00:09:48,054 --> 00:09:49,034 but we're also 270 00:09:49,654 --> 00:09:52,375 accelerating the way that we deliver quality at 271 00:09:52,375 --> 00:09:54,954 a level that would be just not possible 272 00:09:55,254 --> 00:09:56,394 without this technology. 273 00:09:57,899 --> 00:10:00,080 Yeah. So good to know, Jean. Thank you. 274 00:10:00,620 --> 00:10:02,879 And as we kind of close our conversations 275 00:10:02,940 --> 00:10:04,940 today, something I wanted to make sure we 276 00:10:04,940 --> 00:10:06,480 touched on was sustainability 277 00:10:07,580 --> 00:10:09,884 with some of these tools. So at Becker 278 00:10:09,884 --> 00:10:11,345 as we hear about leaders 279 00:10:11,725 --> 00:10:12,225 implementing 280 00:10:12,764 --> 00:10:13,504 new tools 281 00:10:13,884 --> 00:10:15,325 such as what we've touched on today with 282 00:10:15,325 --> 00:10:15,825 SLMs, 283 00:10:16,605 --> 00:10:19,264 but I think the struggle is in maintaining 284 00:10:19,644 --> 00:10:22,720 those those tools, making sure that they're driving 285 00:10:22,720 --> 00:10:25,759 continuous improvements and continuous outcomes that they're aiming 286 00:10:25,759 --> 00:10:28,320 for and really ensuring that they're going from 287 00:10:28,320 --> 00:10:30,019 pilot to real scale. 288 00:10:30,399 --> 00:10:33,220 So as more developer tools come to market, 289 00:10:33,519 --> 00:10:35,360 for you, what does a practical road map 290 00:10:35,360 --> 00:10:37,154 look like for for leaders who are aiming 291 00:10:37,154 --> 00:10:40,375 to balance innovation, trust, and measurable outcomes? 292 00:10:41,394 --> 00:10:43,414 Yes. That's a great question, 293 00:10:43,955 --> 00:10:45,634 and I I think it's a challenge for 294 00:10:45,634 --> 00:10:46,134 everyone. 295 00:10:46,595 --> 00:10:49,154 It's really important not to get distracted on 296 00:10:49,154 --> 00:10:50,995 every shiny thing that comes to the market. 297 00:10:50,995 --> 00:10:51,654 As engineers, 298 00:10:52,429 --> 00:10:53,649 We tend to do that, 299 00:10:54,110 --> 00:10:56,429 more often than not. So, it is also 300 00:10:56,429 --> 00:10:59,389 important to understand that AI is not necessarily 301 00:10:59,389 --> 00:11:01,789 an answer for everything. It is a very 302 00:11:01,789 --> 00:11:03,490 powerful set of tooling. 303 00:11:03,870 --> 00:11:06,029 But whether it's new models or a new 304 00:11:06,029 --> 00:11:08,465 set of tools, the key here is to 305 00:11:08,465 --> 00:11:11,585 focus on the problem you're solving versus the 306 00:11:11,585 --> 00:11:12,485 tool itself. 307 00:11:12,945 --> 00:11:16,384 So start by cleanly outlining the objectives and 308 00:11:16,384 --> 00:11:18,945 important KPIs. You know, for us, those were 309 00:11:18,945 --> 00:11:19,845 the four areas, 310 00:11:20,389 --> 00:11:22,389 that I mentioned. But for a different company, 311 00:11:22,389 --> 00:11:24,490 maybe a different set of priorities. And, 312 00:11:25,590 --> 00:11:27,990 these KPIs will help you measure along the 313 00:11:27,990 --> 00:11:30,410 way and not, you know, get you distracted 314 00:11:30,470 --> 00:11:33,134 from what success looks like. And a good 315 00:11:33,134 --> 00:11:35,075 road map would look like, 316 00:11:35,774 --> 00:11:36,915 especially in AI, 317 00:11:37,455 --> 00:11:39,935 where you laser focus on your key use 318 00:11:39,935 --> 00:11:42,654 cases, where you believe that AI is the 319 00:11:42,654 --> 00:11:44,970 right solution based on the attributes of those, 320 00:11:45,529 --> 00:11:48,329 use cases. And road maps we typically see 321 00:11:48,329 --> 00:11:49,389 is that are successful 322 00:11:49,929 --> 00:11:51,389 is where you start small. 323 00:11:51,929 --> 00:11:54,110 You prove it out. We started, 324 00:11:54,569 --> 00:11:56,490 in kind of a single space state Medicaid 325 00:11:56,490 --> 00:11:57,789 that was extremely difficult, 326 00:11:58,274 --> 00:11:59,574 to run-in quality, 327 00:12:00,115 --> 00:12:02,694 without these tools, and we were extremely successful 328 00:12:03,154 --> 00:12:05,654 in that. And we've expanded those results, 329 00:12:06,754 --> 00:12:09,154 to other areas. And then sometimes, you know, 330 00:12:09,154 --> 00:12:11,074 you may fail. You wanna abandon it and 331 00:12:11,074 --> 00:12:12,214 not kind of continue 332 00:12:13,079 --> 00:12:14,220 doing the same thing, 333 00:12:14,759 --> 00:12:17,259 if you're not seeing a progress there. And 334 00:12:17,319 --> 00:12:20,120 if it works, then the idea is you 335 00:12:20,120 --> 00:12:22,940 scale it by kind of building the foundations 336 00:12:23,319 --> 00:12:25,159 that are needed, then you expand either to 337 00:12:25,159 --> 00:12:27,495 other populations or to larger sizes of data 338 00:12:27,654 --> 00:12:29,115 or to kind of adjacent, 339 00:12:30,054 --> 00:12:30,875 use cases. 340 00:12:31,335 --> 00:12:32,154 But if you're 341 00:12:32,455 --> 00:12:34,934 kind of responsible looking at the roadmap that 342 00:12:34,934 --> 00:12:37,575 way, then more often than not, you're gonna 343 00:12:37,575 --> 00:12:38,235 be successful. 344 00:12:39,894 --> 00:12:42,440 Such sound advice, Jean. Thank you. And I 345 00:12:42,440 --> 00:12:43,960 know we've covered a lot of ground today, 346 00:12:43,960 --> 00:12:45,720 but wanted to just check-in to see if 347 00:12:45,720 --> 00:12:47,879 there was anything that we haven't covered that 348 00:12:47,879 --> 00:12:49,879 you think is important for health IT leaders 349 00:12:49,879 --> 00:12:51,179 to understand right now. 350 00:12:51,879 --> 00:12:55,339 Yes. We talk a lot about the value 351 00:12:55,995 --> 00:12:58,815 that AI can drive in certain circumstances. 352 00:13:00,075 --> 00:13:01,294 We haven't talked, 353 00:13:02,075 --> 00:13:03,855 much around the foundation 354 00:13:04,315 --> 00:13:05,054 that's needed, 355 00:13:05,514 --> 00:13:06,815 to be able to build 356 00:13:07,274 --> 00:13:07,980 AI successfully. 357 00:13:16,059 --> 00:13:16,559 Unfortunately, 358 00:13:17,100 --> 00:13:19,179 AI is expensive, and it's kind of very 359 00:13:19,179 --> 00:13:20,459 hard to do it at the side of 360 00:13:20,459 --> 00:13:22,139 your desk to do it at a scale, 361 00:13:22,139 --> 00:13:24,235 especially in health care. You need to make 362 00:13:24,235 --> 00:13:26,334 investments into your governance programs. 363 00:13:26,714 --> 00:13:28,794 You need to be able to make investments 364 00:13:28,794 --> 00:13:29,534 into technology, 365 00:13:30,634 --> 00:13:32,975 tracking both data and model drifts, 366 00:13:33,355 --> 00:13:33,855 automatically 367 00:13:34,475 --> 00:13:36,714 and be able to monitor and alert on 368 00:13:36,714 --> 00:13:39,089 those, when they happen to make sure that 369 00:13:39,089 --> 00:13:39,750 your models 370 00:13:40,209 --> 00:13:41,829 change successfully over time. 371 00:13:42,289 --> 00:13:45,009 And, setting the AI controls to prevent things 372 00:13:45,009 --> 00:13:47,350 like data poisoning, dealing with edge cases, 373 00:13:48,209 --> 00:13:49,669 building, like, the right 374 00:13:49,995 --> 00:13:52,315 user experience to be able to provide, 375 00:13:52,794 --> 00:13:55,835 feedback and train your AI algorithms and then 376 00:13:55,835 --> 00:13:57,455 have kind of human in the middle. 377 00:13:58,475 --> 00:14:01,914 And, you know, that differentiates going from a 378 00:14:01,914 --> 00:14:04,414 POC to production ready 379 00:14:04,930 --> 00:14:07,250 solutions. It's somewhat maybe kind of some of 380 00:14:07,250 --> 00:14:09,490 the boring stuff, not as exciting as the 381 00:14:09,490 --> 00:14:10,790 ROI. But without 382 00:14:11,170 --> 00:14:13,170 that boring stuff, you're not gonna be able 383 00:14:13,170 --> 00:14:13,830 to scale. 384 00:14:15,009 --> 00:14:17,894 Yeah, Gina. It's a really great point. I've 385 00:14:17,894 --> 00:14:20,075 I've noticed this through line throughout our conversation 386 00:14:20,134 --> 00:14:22,615 that going back to the drawing board, really 387 00:14:22,615 --> 00:14:25,014 establishing the use cases you're looking for, getting 388 00:14:25,014 --> 00:14:27,095 clear on those, getting clear on KPIs, but 389 00:14:27,095 --> 00:14:29,414 also going back to the foundational systems that 390 00:14:29,414 --> 00:14:30,154 are in place. 391 00:14:30,649 --> 00:14:33,129 Those steps have to be done before success 392 00:14:33,129 --> 00:14:33,950 can be expected. 393 00:14:34,730 --> 00:14:35,230 Absolutely. 394 00:14:35,529 --> 00:14:37,210 Well, Jean, it's been great having you on 395 00:14:37,210 --> 00:14:39,129 the podcast today. I wanna thank you again 396 00:14:39,129 --> 00:14:40,509 for your time and your insights, 397 00:14:41,049 --> 00:14:42,990 and we'd also like to thank our podcast 398 00:14:43,049 --> 00:14:44,830 sponsor for today, Lyric. 399 00:14:45,554 --> 00:14:47,235 Listeners, be sure to tune in to more 400 00:14:47,235 --> 00:14:50,034 podcasts from Becker's Healthcare by visiting our podcast 401 00:14:50,034 --> 00:14:53,095 page at beckershospitalreview.com.