1 00:00:00,160 --> 00:00:02,339 Hello, everyone, and welcome to Becker's Healthcare podcast. 2 00:00:02,560 --> 00:00:04,240 I'm Scott King. Thrilled today to be joined 3 00:00:04,240 --> 00:00:05,859 by two very special guests. 4 00:00:06,160 --> 00:00:08,720 First, we'll start with Rick Peng, digital ventures 5 00:00:08,720 --> 00:00:12,000 lead in Memorial Sloan Kettering Cancer Center's office 6 00:00:12,000 --> 00:00:12,740 of entrepreneurship 7 00:00:13,294 --> 00:00:14,035 and Commercialization. 8 00:00:14,574 --> 00:00:16,515 And we also have Natalia Somerville, 9 00:00:16,894 --> 00:00:20,035 director of decision intelligence at Memorial Sloan Kettering 10 00:00:20,094 --> 00:00:22,015 Cancer Center. Thank you both so much for 11 00:00:22,015 --> 00:00:23,375 joining. I know we have a a lot 12 00:00:23,375 --> 00:00:24,894 planned for a for a great, 13 00:00:25,535 --> 00:00:27,850 discussion on all the innovative things you're up 14 00:00:27,850 --> 00:00:29,609 to over at Sloan Kettering. How are you 15 00:00:29,609 --> 00:00:30,269 both doing? 16 00:00:30,810 --> 00:00:32,729 Doing good. Very excited to be here. Thank 17 00:00:32,729 --> 00:00:34,890 you so much for having us. And, yeah, 18 00:00:34,890 --> 00:00:35,390 definitely, 19 00:00:36,329 --> 00:00:38,510 very happy to share some of our experiences. 20 00:00:39,625 --> 00:00:41,145 Yes. No. Thanks so much to to both 21 00:00:41,145 --> 00:00:42,664 of you for joining, and and we'll go 22 00:00:42,664 --> 00:00:44,825 ahead and get started here with this, important 23 00:00:44,825 --> 00:00:45,325 discussion. 24 00:00:46,185 --> 00:00:47,864 Now now, Rick, I'll I'll get started with 25 00:00:47,864 --> 00:00:49,405 you. I know, you 26 00:00:49,784 --> 00:00:52,265 know, MSK is is often seen as a 27 00:00:52,265 --> 00:00:55,559 leader in applying AI in clinical settings. 28 00:00:56,020 --> 00:00:58,759 How is AI currently being used across 29 00:00:59,140 --> 00:01:01,059 the system, and where do you where are 30 00:01:01,059 --> 00:01:02,200 you seeing the greatest 31 00:01:02,899 --> 00:01:05,319 impact in regards to patients today? 32 00:01:06,355 --> 00:01:09,234 Yeah. So AI is definitely being applied across 33 00:01:09,234 --> 00:01:10,534 the board here at MSK. 34 00:01:11,075 --> 00:01:13,954 It's being applied in not only clinical care 35 00:01:13,954 --> 00:01:15,814 but also in the research space. 36 00:01:16,834 --> 00:01:17,334 It's, 37 00:01:17,715 --> 00:01:20,299 in the clinical spaces, it's it's used in, 38 00:01:21,000 --> 00:01:22,540 use cases such as assisting 39 00:01:23,079 --> 00:01:26,200 clinicians in their visits with patients. Key examples 40 00:01:26,200 --> 00:01:29,079 are partnership with companies like Abridge that help 41 00:01:29,079 --> 00:01:30,219 us, manage 42 00:01:30,680 --> 00:01:31,900 a transcription documentation 43 00:01:32,359 --> 00:01:34,140 of, patient visits. 44 00:01:34,984 --> 00:01:37,545 In the research space, we we do a 45 00:01:37,545 --> 00:01:39,085 lot of internal innovation, 46 00:01:39,944 --> 00:01:41,645 as far as, AI, 47 00:01:42,185 --> 00:01:45,245 development models that can predict responses to therapy, 48 00:01:45,944 --> 00:01:48,905 and that assist in the the development and 49 00:01:48,905 --> 00:01:50,040 discovery of new, 50 00:01:50,599 --> 00:01:51,819 therapeutics in oncology. 51 00:01:52,200 --> 00:01:54,439 We also partner with external companies in this 52 00:01:54,439 --> 00:01:57,159 space as well to co develop new AIs 53 00:01:57,159 --> 00:01:59,880 that can be applied in these situations. And, 54 00:01:59,880 --> 00:02:01,880 of course, we have a large network of 55 00:02:01,880 --> 00:02:05,265 pharma partnerships that we that, can leverage these 56 00:02:05,424 --> 00:02:08,305 technologies to accelerate the development and bringing to 57 00:02:08,305 --> 00:02:10,784 market new drugs that serve unmet needs in 58 00:02:10,784 --> 00:02:11,284 cancer. 59 00:02:11,664 --> 00:02:13,344 And of course, on top of that, we 60 00:02:13,344 --> 00:02:14,245 also apply 61 00:02:14,625 --> 00:02:16,564 AI as tools across the enterprise, 62 00:02:17,824 --> 00:02:19,764 to help, staff at MSK, 63 00:02:20,759 --> 00:02:23,159 outside of just clinicians and researchers to to 64 00:02:23,159 --> 00:02:25,240 do their jobs as effectively as possible as 65 00:02:25,240 --> 00:02:26,860 well. So really across the board. 66 00:02:27,719 --> 00:02:29,959 Yeah. Appreciate the details there on all the 67 00:02:29,959 --> 00:02:33,000 AI usages, Rick. Now, Natalia, I know you 68 00:02:33,000 --> 00:02:34,219 you co create MSK's 69 00:02:34,680 --> 00:02:35,500 AI governance 70 00:02:35,985 --> 00:02:38,084 operating model, which is obviously a huge accomplishment, 71 00:02:38,224 --> 00:02:40,884 but what problem was MSK trying to solve 72 00:02:41,504 --> 00:02:43,764 when you were developing the framework? 73 00:02:44,864 --> 00:02:45,264 Yeah. 74 00:02:45,905 --> 00:02:47,685 So a couple years ago, 75 00:02:48,305 --> 00:02:49,284 here at MSK, 76 00:02:49,664 --> 00:02:50,064 our, 77 00:02:50,625 --> 00:02:51,125 leadership 78 00:02:51,960 --> 00:02:54,939 created what was called the AI task force, 79 00:02:55,639 --> 00:02:59,080 where kinda given all the movement and, 80 00:02:59,560 --> 00:03:00,060 progress 81 00:03:00,360 --> 00:03:00,860 happening, 82 00:03:01,159 --> 00:03:02,300 AI and health care, 83 00:03:03,159 --> 00:03:05,479 our leadership wanted to understand what are the 84 00:03:05,479 --> 00:03:08,175 important aspects and what should we be looking 85 00:03:08,175 --> 00:03:11,055 for, what are the valuable use cases. So 86 00:03:11,055 --> 00:03:13,635 one of the work groups was governance. 87 00:03:14,335 --> 00:03:17,215 And, that work group was led by by 88 00:03:17,215 --> 00:03:17,715 our 89 00:03:18,974 --> 00:03:21,555 chief informatics officer, doctor Pitstetson, 90 00:03:22,094 --> 00:03:22,915 at the time. 91 00:03:23,400 --> 00:03:26,840 And, during that group, I was supporting the 92 00:03:26,840 --> 00:03:27,340 lead, 93 00:03:27,960 --> 00:03:30,060 and we had conversations with 94 00:03:30,599 --> 00:03:31,099 important, 95 00:03:32,520 --> 00:03:33,020 players, 96 00:03:33,719 --> 00:03:35,819 let's say within the hospital, like 97 00:03:36,155 --> 00:03:36,814 legal and, 98 00:03:37,754 --> 00:03:39,534 clinicians who have been working ethics 99 00:03:40,155 --> 00:03:40,635 and, 100 00:03:40,955 --> 00:03:41,455 technologists. 101 00:03:42,314 --> 00:03:45,375 And, that's where there was obviously 102 00:03:45,675 --> 00:03:48,655 kinda an understanding between everybody that 103 00:03:48,969 --> 00:03:51,610 we are starting to deploy more and more 104 00:03:51,610 --> 00:03:52,110 AI, 105 00:03:52,489 --> 00:03:54,189 and we need to be able 106 00:03:55,289 --> 00:03:57,229 to make sure that it's done safely, 107 00:03:58,250 --> 00:04:00,330 and also that we can, 108 00:04:00,729 --> 00:04:03,709 scale it up because we totally understand that 109 00:04:03,814 --> 00:04:05,814 if maybe a couple years there was a 110 00:04:05,814 --> 00:04:07,194 couple models deployed, 111 00:04:07,655 --> 00:04:08,314 this is 112 00:04:08,775 --> 00:04:11,335 growing exponentially now. So we wanna be able 113 00:04:11,335 --> 00:04:12,694 to do that, 114 00:04:13,254 --> 00:04:15,034 responsibly and support 115 00:04:15,414 --> 00:04:16,235 that innovation 116 00:04:16,535 --> 00:04:17,355 and that, 117 00:04:17,850 --> 00:04:19,709 scaling up of AI technologists. 118 00:04:20,329 --> 00:04:20,829 Absolutely. 119 00:04:21,129 --> 00:04:22,490 And, Rick, I want to ask you about 120 00:04:22,490 --> 00:04:24,009 the AI governance as well. You know, in 121 00:04:24,009 --> 00:04:26,490 in a high stakes environment like cancer care, 122 00:04:26,490 --> 00:04:27,310 why is governance 123 00:04:27,850 --> 00:04:29,689 it's not optional when it when it comes 124 00:04:29,689 --> 00:04:29,930 to, 125 00:04:30,810 --> 00:04:31,629 AI adoption, 126 00:04:32,009 --> 00:04:32,990 and why is that? 127 00:04:33,634 --> 00:04:35,074 Yeah. I mean, I I I think there 128 00:04:35,154 --> 00:04:36,055 there's obvious, 129 00:04:36,435 --> 00:04:37,415 you know, ramifications 130 00:04:37,875 --> 00:04:39,014 on, you know, 131 00:04:39,394 --> 00:04:42,855 how patients receive care, how research is conducted, 132 00:04:43,634 --> 00:04:45,394 not only in terms of, like, the clinical 133 00:04:45,394 --> 00:04:47,175 impact on patients, but also, 134 00:04:47,475 --> 00:04:48,870 you know, everything from, 135 00:04:49,730 --> 00:04:52,769 you know, ethics and, scientific integrity as far 136 00:04:52,769 --> 00:04:54,689 as all that work that we do. And 137 00:04:54,689 --> 00:04:57,110 I and I think, you know, broadly, like, 138 00:04:57,810 --> 00:05:00,194 the the need for governance in AI is 139 00:05:00,435 --> 00:05:00,935 particularly 140 00:05:01,235 --> 00:05:01,735 important, 141 00:05:02,194 --> 00:05:04,835 as we think about, you know, organizations like 142 00:05:04,835 --> 00:05:07,154 MSK are very large. There's a lot of 143 00:05:07,154 --> 00:05:08,615 different stakeholders involved, 144 00:05:09,714 --> 00:05:11,095 including external collaborators. 145 00:05:11,395 --> 00:05:12,694 And I think, you know, 146 00:05:13,314 --> 00:05:14,134 AI requires 147 00:05:14,850 --> 00:05:17,669 connectivity across all these different groups, whether they're, 148 00:05:17,810 --> 00:05:19,889 you know, at the points of, applying the 149 00:05:19,889 --> 00:05:21,729 AI, at the points of developing the AI, 150 00:05:21,729 --> 00:05:23,810 at the points of supplying the inputs to 151 00:05:23,810 --> 00:05:24,629 AI, including, 152 00:05:25,009 --> 00:05:25,509 data, 153 00:05:26,290 --> 00:05:28,529 where, you know, we really need all these 154 00:05:28,529 --> 00:05:31,204 pieces to be, you know, working in coordination 155 00:05:31,264 --> 00:05:32,944 with each other to make sure that what 156 00:05:32,944 --> 00:05:34,245 comes out the other end 157 00:05:34,704 --> 00:05:35,204 is, 158 00:05:35,904 --> 00:05:36,404 created, 159 00:05:36,785 --> 00:05:37,845 in a way that, 160 00:05:38,384 --> 00:05:41,904 optimizes its intended performance as well as as 161 00:05:41,904 --> 00:05:43,044 being applied responsibly, 162 00:05:44,379 --> 00:05:46,139 across all the different groups that need to 163 00:05:46,139 --> 00:05:47,120 be using it. 164 00:05:47,980 --> 00:05:50,860 Natalia, what qualify what what qualities I'm sorry. 165 00:05:50,860 --> 00:05:53,740 Should health systems look for an AI governance 166 00:05:53,740 --> 00:05:55,199 partner to ensure responsible 167 00:05:55,660 --> 00:05:57,360 adoption while still supporting 168 00:05:57,740 --> 00:05:58,480 some innovation? 169 00:05:59,514 --> 00:06:02,175 Yeah. I think it's definitely flexibility, 170 00:06:03,035 --> 00:06:03,535 because 171 00:06:04,154 --> 00:06:06,574 as we have learned and you just mentioned, 172 00:06:07,115 --> 00:06:09,595 we have to be able to balance those 173 00:06:09,595 --> 00:06:10,415 two goals, 174 00:06:11,689 --> 00:06:15,210 be able to deploy AI responsibly, but also 175 00:06:15,210 --> 00:06:16,430 making sure that 176 00:06:16,810 --> 00:06:19,230 we are not seen and we are not, 177 00:06:19,930 --> 00:06:20,430 blocking 178 00:06:20,810 --> 00:06:22,189 any any innovation 179 00:06:22,730 --> 00:06:25,629 performed by our clinicians. Our clinicians do amazing 180 00:06:25,689 --> 00:06:26,834 research, and that is 181 00:06:27,235 --> 00:06:30,774 partially why our institution is so well known 182 00:06:30,834 --> 00:06:32,675 because of the research they do. So we 183 00:06:32,675 --> 00:06:34,774 wanna make sure to encourage that research. 184 00:06:36,115 --> 00:06:38,595 So the way we we partner and the 185 00:06:38,595 --> 00:06:40,694 way we we see these processes 186 00:06:41,779 --> 00:06:44,899 is making sure that there are different paths, 187 00:06:44,899 --> 00:06:46,519 that if there is an AI 188 00:06:47,620 --> 00:06:49,399 model or AI system 189 00:06:49,860 --> 00:06:50,360 that, 190 00:06:50,740 --> 00:06:51,479 the governance 191 00:06:51,860 --> 00:06:53,860 committee is looking at a time that we 192 00:06:53,860 --> 00:06:55,959 have different paths, that we are flexible, 193 00:06:56,845 --> 00:06:59,824 understanding that the the research initiatives 194 00:07:00,764 --> 00:07:03,004 may have a faster path. They are working 195 00:07:03,004 --> 00:07:03,504 towards 196 00:07:03,964 --> 00:07:07,024 proof of concepts and, developing the science, 197 00:07:07,884 --> 00:07:11,185 versus tools that may go into workflow 198 00:07:11,564 --> 00:07:12,064 workflows, 199 00:07:12,839 --> 00:07:13,500 right away, 200 00:07:14,759 --> 00:07:16,539 let's say predicting mortality 201 00:07:16,919 --> 00:07:19,180 and we wanna deploy that within two months, 202 00:07:19,479 --> 00:07:22,199 then we definitely wanna have a different path, 203 00:07:22,680 --> 00:07:25,579 much more focused on reviewing the details. 204 00:07:26,144 --> 00:07:28,245 So, yeah, I believe that that flexibility 205 00:07:28,785 --> 00:07:29,764 on adjusting 206 00:07:30,545 --> 00:07:31,285 the processes 207 00:07:31,904 --> 00:07:32,384 and, 208 00:07:33,024 --> 00:07:34,884 our AI life cycles 209 00:07:35,425 --> 00:07:38,004 across the different types of implementation 210 00:07:38,944 --> 00:07:40,084 is is key. 211 00:07:41,329 --> 00:07:42,870 Natalia, how does MSK's 212 00:07:43,329 --> 00:07:44,069 AI governance 213 00:07:44,529 --> 00:07:47,730 operating model help ensure that AI tools are 214 00:07:47,730 --> 00:07:48,550 safe and 215 00:07:48,850 --> 00:07:51,509 accountable and aligned with clinical standards? 216 00:07:52,689 --> 00:07:54,550 Mhmm. Yeah. That's that's a great question. 217 00:07:55,095 --> 00:07:56,395 So over the last, 218 00:07:56,935 --> 00:07:58,555 I wanna say two years, 219 00:07:59,095 --> 00:07:59,835 we developed, 220 00:08:00,775 --> 00:08:01,754 these processes 221 00:08:02,375 --> 00:08:02,855 and, 222 00:08:03,175 --> 00:08:05,735 an AI life cycle. So this is a 223 00:08:05,735 --> 00:08:06,694 tool that, 224 00:08:07,254 --> 00:08:07,754 kinda 225 00:08:08,055 --> 00:08:09,675 ground the grounds 226 00:08:10,110 --> 00:08:10,610 the 227 00:08:10,990 --> 00:08:11,490 AI 228 00:08:12,029 --> 00:08:13,089 research and deployment 229 00:08:13,470 --> 00:08:14,209 that happens. 230 00:08:14,669 --> 00:08:17,410 An AI life cycle is the the, 231 00:08:18,350 --> 00:08:21,410 the the sequence of steps that you follow 232 00:08:21,709 --> 00:08:22,529 such as, 233 00:08:23,335 --> 00:08:25,735 you have an idea, you have to validate 234 00:08:25,735 --> 00:08:26,634 that the idea, 235 00:08:27,895 --> 00:08:28,395 has, 236 00:08:28,774 --> 00:08:30,634 you know, value within the organization, 237 00:08:31,095 --> 00:08:34,215 and then design, deployment, testing, all all those 238 00:08:34,215 --> 00:08:35,115 different steps. 239 00:08:36,080 --> 00:08:38,720 I do wanna call out that, actually, two 240 00:08:38,720 --> 00:08:41,860 years ago, our our life cycle was, 241 00:08:42,320 --> 00:08:44,720 our original life cycle. We did use, 242 00:08:45,120 --> 00:08:48,205 Duke Health Systems life cycle. They are very 243 00:08:48,205 --> 00:08:50,544 advanced in this area, and then we modified 244 00:08:50,684 --> 00:08:53,085 it what makes sense for us. But so 245 00:08:53,085 --> 00:08:55,825 what I was mentioning is this life cycle 246 00:08:56,125 --> 00:08:58,065 allows us to create steps 247 00:08:58,524 --> 00:08:59,024 depending 248 00:08:59,725 --> 00:09:02,945 on what stage the AI system is. 249 00:09:03,259 --> 00:09:04,159 So for example, 250 00:09:04,460 --> 00:09:06,080 if the AI system 251 00:09:06,620 --> 00:09:07,600 that we are, 252 00:09:08,299 --> 00:09:10,080 looking into or, 253 00:09:11,419 --> 00:09:13,279 governing at the moment is 254 00:09:13,740 --> 00:09:17,360 in, pilot mode, then there are certain requirements 255 00:09:17,659 --> 00:09:19,595 that we ask. We ask for, 256 00:09:20,214 --> 00:09:20,875 the pilot, 257 00:09:22,134 --> 00:09:24,774 plan to be very explicit and presented to 258 00:09:24,774 --> 00:09:28,294 the committee. We're asking for a monitoring plan 259 00:09:28,294 --> 00:09:30,394 in advance of the deployment 260 00:09:31,014 --> 00:09:31,514 versus 261 00:09:31,970 --> 00:09:35,250 if, the AI system is earlier. Let's say 262 00:09:35,250 --> 00:09:37,410 they're just kind of finished design and going 263 00:09:37,410 --> 00:09:38,309 into development. 264 00:09:38,690 --> 00:09:41,570 There's other sorts of questions that we we 265 00:09:41,570 --> 00:09:44,230 ask. So depending on the life cycle, 266 00:09:44,875 --> 00:09:47,454 we have these different touch points 267 00:09:47,995 --> 00:09:50,894 with the model owners and model sponsors 268 00:09:51,274 --> 00:09:54,334 to make sure that it goes through, 269 00:09:55,115 --> 00:09:57,355 in a in a safe way, and we're 270 00:09:57,355 --> 00:09:58,334 able to 271 00:09:58,769 --> 00:10:00,389 suggest any modifications 272 00:10:00,850 --> 00:10:02,710 as early as possible if needed. 273 00:10:03,810 --> 00:10:06,129 And and, Rick, what risks do health systems 274 00:10:06,129 --> 00:10:08,690 face if they deploy AI tools, you know, 275 00:10:08,690 --> 00:10:11,269 without a clear strategy or framework in place? 276 00:10:12,024 --> 00:10:14,504 Yeah. So I think, like, for for health 277 00:10:14,504 --> 00:10:16,365 systems, if AI isn't 278 00:10:17,464 --> 00:10:21,004 isn't deployed in a way that's strategically directed, 279 00:10:21,544 --> 00:10:23,784 I think there's there's a couple levels of 280 00:10:23,784 --> 00:10:24,284 risks. 281 00:10:24,985 --> 00:10:25,884 One is, 282 00:10:26,690 --> 00:10:27,269 an operational 283 00:10:27,570 --> 00:10:28,450 risk of, 284 00:10:28,850 --> 00:10:31,990 have impacting the adoption of these technologies. 285 00:10:32,850 --> 00:10:35,829 If there isn't a clear idea of what 286 00:10:35,889 --> 00:10:37,350 workflows or what stakeholders, 287 00:10:38,450 --> 00:10:40,049 need to pick this up and, you know, 288 00:10:40,049 --> 00:10:41,269 build it into practice, 289 00:10:41,924 --> 00:10:43,225 it can really limit, 290 00:10:43,845 --> 00:10:46,004 the extent to which the the AI is 291 00:10:46,004 --> 00:10:47,465 being applied in the first place. 292 00:10:48,004 --> 00:10:51,044 I think, ultimately, longer term, it can impact 293 00:10:51,044 --> 00:10:53,625 the ROI of these types of solutions, especially 294 00:10:53,684 --> 00:10:56,250 if there, you know, if there's significant investment 295 00:10:56,250 --> 00:10:58,750 in them, whether it's through purchasing external 296 00:11:00,090 --> 00:11:01,230 solutions, or, 297 00:11:01,929 --> 00:11:04,750 you know, investing in internal development of capabilities 298 00:11:05,129 --> 00:11:07,129 if if there isn't the right adoption or 299 00:11:07,129 --> 00:11:08,649 if it's not being applied in the most 300 00:11:08,649 --> 00:11:09,950 impactful use cases 301 00:11:10,415 --> 00:11:12,575 and reduce your ROI, which I think, you 302 00:11:12,575 --> 00:11:13,215 know, can, 303 00:11:13,934 --> 00:11:14,675 not only, 304 00:11:15,455 --> 00:11:18,995 hamper the the the impact of the particular 305 00:11:19,535 --> 00:11:22,415 AI solution in question, but also just AI 306 00:11:22,415 --> 00:11:23,955 across the board as well. 307 00:11:24,389 --> 00:11:27,370 And and one special consideration for for organizations 308 00:11:27,430 --> 00:11:30,389 like MSK as well is, like, MSK definitely, 309 00:11:30,389 --> 00:11:30,970 you know, 310 00:11:31,509 --> 00:11:33,029 it's it's point of pride that we are 311 00:11:33,029 --> 00:11:35,110 one of the leading cancer centers in The 312 00:11:35,110 --> 00:11:37,575 US and in the world. And so I 313 00:11:37,575 --> 00:11:39,975 I think, you know, application of AI also 314 00:11:39,975 --> 00:11:42,615 needs to be done in a way that 315 00:11:42,615 --> 00:11:45,754 fully takes advantage of our the differentiated 316 00:11:46,375 --> 00:11:48,394 expertise that MSK has. 317 00:11:48,774 --> 00:11:50,375 I think a lot of times where that 318 00:11:50,375 --> 00:11:52,394 comes into play especially is, 319 00:11:53,080 --> 00:11:55,580 you know, in terms of using AI solutions, 320 00:11:56,200 --> 00:11:57,659 ensuring that there's, 321 00:11:58,120 --> 00:12:00,839 the appropriate level of governance, but also input 322 00:12:00,839 --> 00:12:01,580 the development 323 00:12:02,440 --> 00:12:04,860 of those AI tools, whether it's through leveraging 324 00:12:05,079 --> 00:12:05,579 a, 325 00:12:05,960 --> 00:12:06,460 MSK's, 326 00:12:07,159 --> 00:12:09,225 unique know how expertise of its 327 00:12:09,925 --> 00:12:11,924 clinical and research experts, to really to really 328 00:12:11,924 --> 00:12:14,485 to really, you know, infuse the AI solutions 329 00:12:14,485 --> 00:12:15,544 that we do apply, 330 00:12:16,485 --> 00:12:18,565 with the with the sort of secret sauce 331 00:12:18,565 --> 00:12:19,325 at MSK so that, you know, when applied, 332 00:12:19,325 --> 00:12:19,924 it really reflects the level of of quality 333 00:12:19,924 --> 00:12:20,174 that, 334 00:12:21,480 --> 00:12:23,799 applied, it really reflects the level of of 335 00:12:23,799 --> 00:12:27,079 quality that, MSK has come to be known 336 00:12:27,079 --> 00:12:29,159 for in the world of cancer care and 337 00:12:29,159 --> 00:12:29,659 research. 338 00:12:31,480 --> 00:12:33,000 Yeah. Thanks so much, Ray. And maybe diving 339 00:12:33,000 --> 00:12:35,004 into that that secret sauce, 340 00:12:35,384 --> 00:12:38,345 Natalia, that the process with AI over at 341 00:12:38,345 --> 00:12:40,665 MSK, you know, you know, even after tools 342 00:12:40,665 --> 00:12:43,225 are approved and deployed, the work's not done. 343 00:12:43,225 --> 00:12:45,644 Right? So why is continuous monitoring 344 00:12:46,345 --> 00:12:47,725 of AI systems essential? 345 00:12:48,750 --> 00:12:51,230 Yeah. That is it's it's very important as 346 00:12:51,230 --> 00:12:53,950 you mentioned, and, I the way I talk 347 00:12:53,950 --> 00:12:56,129 about monitoring is actually twofold. 348 00:12:57,310 --> 00:12:59,170 One is monitoring, 349 00:13:00,029 --> 00:13:02,335 the the the safety. You know? Has there 350 00:13:02,335 --> 00:13:04,754 been any adverse events, for example? Because, 351 00:13:05,534 --> 00:13:06,034 as 352 00:13:06,335 --> 00:13:07,695 as Rick was mentioning, there's, 353 00:13:08,335 --> 00:13:09,154 kinda several 354 00:13:09,615 --> 00:13:12,254 aspects that could potentially go wrong. So we 355 00:13:12,254 --> 00:13:13,634 wanna be able to monitoring 356 00:13:14,330 --> 00:13:16,029 any, AI adverse, 357 00:13:16,570 --> 00:13:17,070 adverse, 358 00:13:17,929 --> 00:13:18,429 events, 359 00:13:19,290 --> 00:13:21,549 as well as kinda the the technical 360 00:13:21,929 --> 00:13:24,429 model itself. Is there any drift? 361 00:13:25,290 --> 00:13:27,690 For example, if there is a pattern change 362 00:13:27,690 --> 00:13:28,590 in the data, 363 00:13:29,004 --> 00:13:31,584 then we need to be able to acknowledge 364 00:13:31,725 --> 00:13:33,485 that in the model if it doesn't do 365 00:13:33,485 --> 00:13:34,225 it automatically. 366 00:13:34,924 --> 00:13:36,845 So depending on the type of model, we 367 00:13:36,845 --> 00:13:38,384 need to be able to, 368 00:13:39,485 --> 00:13:40,625 see that every 369 00:13:41,019 --> 00:13:43,419 quarter, every month. It it really depends on 370 00:13:43,419 --> 00:13:45,039 how the model is is structured. 371 00:13:46,059 --> 00:13:48,459 At the same time, the second part that 372 00:13:48,459 --> 00:13:50,799 is also very important that sometimes 373 00:13:51,419 --> 00:13:54,879 it's talked about less actually than the technical 374 00:13:55,500 --> 00:13:56,000 monitoring 375 00:13:56,605 --> 00:13:59,325 is precisely what Rick was mentioning. We also 376 00:13:59,325 --> 00:14:01,164 want to make sure that the tool is 377 00:14:01,164 --> 00:14:02,304 actually being used. 378 00:14:02,845 --> 00:14:05,965 And beyond being used, that the tool is 379 00:14:05,965 --> 00:14:07,985 bringing the expected impact. 380 00:14:08,764 --> 00:14:09,825 If we are deploying, 381 00:14:10,684 --> 00:14:11,825 an AI tool 382 00:14:12,389 --> 00:14:12,970 to support, 383 00:14:13,509 --> 00:14:14,970 for example, discharges, 384 00:14:15,990 --> 00:14:18,569 the the nurses with their discharge efforts, 385 00:14:19,029 --> 00:14:22,149 and the original goal was to reduce the 386 00:14:22,149 --> 00:14:24,389 patient length of stay such that they can 387 00:14:24,389 --> 00:14:26,409 go home as soon as possible, 388 00:14:27,075 --> 00:14:29,894 and enjoy kinda more, comfort settings, 389 00:14:30,434 --> 00:14:32,754 then we wanna see that the length of 390 00:14:32,754 --> 00:14:34,375 stay is actually decreasing. 391 00:14:34,675 --> 00:14:36,995 So we wanna be able to see those 392 00:14:36,995 --> 00:14:39,495 metrics that we are targeting, 393 00:14:40,529 --> 00:14:42,529 as well as usage. We don't wanna be 394 00:14:42,529 --> 00:14:43,029 maintaining 395 00:14:43,570 --> 00:14:46,690 tools that nobody's using. So those are the 396 00:14:46,690 --> 00:14:49,750 two aspects, and they are typically monitored, 397 00:14:50,850 --> 00:14:52,709 differently and in kinda different 398 00:14:53,485 --> 00:14:56,304 platform and visualizations, but both are 399 00:14:56,605 --> 00:14:59,245 key to be able to say that the 400 00:14:59,245 --> 00:15:01,904 AI is being used properly. 401 00:15:03,245 --> 00:15:03,904 And, Rick, 402 00:15:04,285 --> 00:15:07,745 government regulation is always obviously a big topic 403 00:15:08,199 --> 00:15:10,779 in health care, and it doesn't escape AI. 404 00:15:11,159 --> 00:15:12,459 So with AI regulations 405 00:15:12,839 --> 00:15:14,679 now evolving at the state level, how does 406 00:15:14,679 --> 00:15:16,059 an organization like MSK 407 00:15:16,600 --> 00:15:20,220 streamline compliance without slowing progress or innovation? 408 00:15:21,264 --> 00:15:23,105 So I I think a big part of 409 00:15:23,105 --> 00:15:23,924 that is, 410 00:15:24,705 --> 00:15:28,144 creating visibility across all different stakeholder groups. Part 411 00:15:28,144 --> 00:15:31,024 of that is, on an operational level, the 412 00:15:31,024 --> 00:15:34,725 creation of the appropriate governance committees with, multidisciplinary, 413 00:15:35,970 --> 00:15:36,470 multifunctional 414 00:15:36,850 --> 00:15:37,350 representation, 415 00:15:38,129 --> 00:15:41,089 which, Natalia and her colleagues have really spearheaded 416 00:15:41,089 --> 00:15:43,029 over the past few years here at MSK. 417 00:15:44,209 --> 00:15:45,570 A a part of that as well is 418 00:15:45,570 --> 00:15:46,709 just facilitated 419 00:15:47,009 --> 00:15:50,929 through through technology solutions that, enable us to 420 00:15:50,929 --> 00:15:53,794 do this on an efficient basis. And so, 421 00:15:54,174 --> 00:15:55,455 having the tools for these, 422 00:15:55,934 --> 00:15:56,995 these these individuals 423 00:15:57,375 --> 00:16:00,414 within these committees to effectively communicate, have the 424 00:16:00,414 --> 00:16:02,914 right information at their fingertips so that when, 425 00:16:03,375 --> 00:16:05,955 there need to be conversations or decisions around, 426 00:16:06,620 --> 00:16:09,600 the the these AI solutions that that's handled. 427 00:16:09,899 --> 00:16:12,220 And I think, like, at MSK, I'm not 428 00:16:12,220 --> 00:16:14,779 directly embedded in these functions, but we also 429 00:16:14,779 --> 00:16:16,960 have very robust functions around 430 00:16:17,340 --> 00:16:19,899 regulatory compliance. And I think a big part 431 00:16:19,899 --> 00:16:21,120 of it is MSK 432 00:16:22,085 --> 00:16:24,565 overall, you know, prides ourselves in, you know, 433 00:16:24,565 --> 00:16:27,365 staying abreast of all the latest developments, not, 434 00:16:27,605 --> 00:16:30,404 in in all the regulatory and legal spaces 435 00:16:30,404 --> 00:16:31,924 that impact the work that we do. And 436 00:16:31,924 --> 00:16:32,424 so, 437 00:16:33,125 --> 00:16:36,184 our our processes are, and our our stakeholders 438 00:16:36,245 --> 00:16:38,889 in those functions are, ever evolving to to 439 00:16:38,889 --> 00:16:39,389 meet 440 00:16:39,929 --> 00:16:41,629 the the the the latest requirements, 441 00:16:42,490 --> 00:16:44,410 and make sure that we we stay not 442 00:16:44,410 --> 00:16:46,809 only up to speed, but also think ahead 443 00:16:46,809 --> 00:16:48,490 in terms of where we need to be, 444 00:16:48,809 --> 00:16:50,894 not just this year, but also next year. 445 00:16:51,455 --> 00:16:52,815 Yeah. And if I if if that's okay, 446 00:16:52,815 --> 00:16:55,054 I can, jump in as well as as 447 00:16:55,134 --> 00:16:57,554 definitely, as Rick is mentioning, the the 448 00:16:58,014 --> 00:17:00,115 the other aspect is also kinda 449 00:17:00,415 --> 00:17:00,915 being 450 00:17:01,295 --> 00:17:04,595 involved with and in communication with other organizations 451 00:17:05,055 --> 00:17:05,714 like us 452 00:17:06,049 --> 00:17:09,349 that are also working and implementing their governance 453 00:17:09,490 --> 00:17:09,990 processes, 454 00:17:11,009 --> 00:17:13,890 and part of CHI. For example, MSK is 455 00:17:13,890 --> 00:17:14,869 part of CHI, 456 00:17:15,250 --> 00:17:17,169 and that is a way that we try 457 00:17:17,169 --> 00:17:17,669 to, 458 00:17:18,210 --> 00:17:18,710 keep, 459 00:17:19,544 --> 00:17:22,044 keep present in kinda the latest developments 460 00:17:22,585 --> 00:17:23,085 and, 461 00:17:23,625 --> 00:17:25,005 as well from the 462 00:17:25,384 --> 00:17:27,404 government or legal aspects, 463 00:17:27,704 --> 00:17:28,204 any 464 00:17:28,505 --> 00:17:29,884 new technology developments. 465 00:17:30,585 --> 00:17:33,005 I do believe that it's kinda important that 466 00:17:33,109 --> 00:17:35,269 all these organizations that we are, 467 00:17:35,750 --> 00:17:38,869 sharing with each other and kinda progressing together 468 00:17:38,869 --> 00:17:39,690 in this way. 469 00:17:40,549 --> 00:17:42,309 Yeah. Yeah. As you as you share what 470 00:17:42,309 --> 00:17:44,869 you've learned through this experience, Natalia, like, what 471 00:17:44,869 --> 00:17:47,474 lessons do you do you think from MSK's 472 00:17:47,535 --> 00:17:49,934 approach to AI governance do you think can 473 00:17:49,934 --> 00:17:51,154 be applied to other 474 00:17:51,535 --> 00:17:53,234 hospitals or or health care organizations? 475 00:17:54,494 --> 00:17:56,674 Yeah. Definitely. I love this question because 476 00:17:56,974 --> 00:17:57,855 one thing that, 477 00:17:58,429 --> 00:17:59,329 is often 478 00:17:59,710 --> 00:18:00,210 also 479 00:18:00,509 --> 00:18:02,369 not talked about is actually how 480 00:18:02,750 --> 00:18:03,250 operationally 481 00:18:03,950 --> 00:18:04,450 heavy 482 00:18:04,789 --> 00:18:05,289 is 483 00:18:05,630 --> 00:18:06,130 to 484 00:18:07,309 --> 00:18:09,730 manage the govern the the governance process. 485 00:18:10,190 --> 00:18:12,769 Because as I mentioned at the beginning, the 486 00:18:13,795 --> 00:18:15,654 AI, kind of the the AI 487 00:18:15,955 --> 00:18:18,615 implementations and systems and development is, 488 00:18:19,154 --> 00:18:20,134 growing exponentially. 489 00:18:20,434 --> 00:18:20,934 But 490 00:18:21,315 --> 00:18:23,394 if we want to have a touch base 491 00:18:23,394 --> 00:18:26,615 with every model, every system that is medium 492 00:18:26,674 --> 00:18:29,220 to high risk, there's a lot of operational 493 00:18:30,400 --> 00:18:32,400 tasks that need to be done, you know, 494 00:18:32,960 --> 00:18:35,619 kinda check with the model owner, have them 495 00:18:35,759 --> 00:18:37,059 register their systems 496 00:18:37,440 --> 00:18:41,015 and verify they are pilots and perform risk 497 00:18:41,015 --> 00:18:43,595 assessments. And so it's actually operationally 498 00:18:43,894 --> 00:18:44,954 very, very heavy. 499 00:18:45,974 --> 00:18:46,454 And, 500 00:18:47,494 --> 00:18:48,714 so partnering 501 00:18:49,575 --> 00:18:52,500 with somebody who can help with that, 502 00:18:53,059 --> 00:18:54,519 those sort of operations 503 00:18:55,059 --> 00:18:56,759 can be very, very helpful. 504 00:18:57,859 --> 00:19:00,920 So for organizations that are beginning to look 505 00:19:01,059 --> 00:19:04,759 into this AI governance systems, I will definitely 506 00:19:04,819 --> 00:19:07,755 recommend two things. One is take a look 507 00:19:07,755 --> 00:19:09,215 at what has been done already. 508 00:19:09,835 --> 00:19:12,095 Many many of us has have published 509 00:19:12,474 --> 00:19:14,654 some of our systems. We are also always 510 00:19:14,715 --> 00:19:17,994 happy to talk about that, kinda implement some 511 00:19:17,994 --> 00:19:19,455 of the processes, 512 00:19:20,474 --> 00:19:23,289 and guidance that other organizations have done already 513 00:19:23,509 --> 00:19:25,910 and also look at platforms that can help 514 00:19:25,910 --> 00:19:27,049 with that operationalization 515 00:19:28,470 --> 00:19:29,450 of the governance. 516 00:19:30,549 --> 00:19:31,529 That is typically 517 00:19:32,470 --> 00:19:32,970 bypassed 518 00:19:33,349 --> 00:19:34,869 at the beginning and, 519 00:19:35,190 --> 00:19:36,970 that can really delay 520 00:19:37,644 --> 00:19:39,105 any sort of meaningful 521 00:19:40,045 --> 00:19:42,545 scaling that you can do with your governance 522 00:19:42,765 --> 00:19:43,265 process. 523 00:19:43,884 --> 00:19:45,884 Well, Rick and Natalia, thank you so much 524 00:19:45,884 --> 00:19:47,724 for joining the podcast and just for a 525 00:19:47,724 --> 00:19:48,545 great discussion 526 00:19:49,404 --> 00:19:51,960 on MSK's, you know, innovative use of of 527 00:19:51,960 --> 00:19:53,400 AI. I look forward to working with you 528 00:19:53,400 --> 00:19:54,380 both again soon. 529 00:19:55,000 --> 00:19:56,700 Thank you so much for having us. 530 00:19:57,000 --> 00:19:58,279 Yeah. Thank you so much, Scott. It was 531 00:19:58,279 --> 00:19:59,179 great to be here.