Developing predictive tools for gastric cancer screening – 2022 Gastric Cancer Summit

[00:00] Our next speaker, it's my pleasure to introduce Dr. Haejin In. She is a surgical oncologist at Rutgers and is very interested in better strategies for early detection. I'm very interested to hear what she has to tell us about developing predictive tools for gastric cancer screening. Can we hone in on higher risk populations?

[00:20] can we focus our endoscopies on higher yield situations? So very interested in what you'll have to say. Welcome. Thank you. Thank you so much for the kind introduction. All right, thank you, everybody. So I'm really excited to be here because

[00:40] As a surgeon, so I started off, just a little background about myself, I started off, my research was actually more in comparative effectiveness and quality for surgery. And when I started my faculty position, I kept on, I'm an upper GI surgical oncologist.

[01:00] I was dealing with a lot of stomach cancer. And at some point, it kind of dawned on me, like why am I not looking at stomach cancer? So I really started to delve into stomach cancer. And because I started a little bit late and without the benefit of lots of mentors and people who knew how to do stomach cancer.

[01:20] cancer, I kind of got a free hand in anything I wanted to do. And so my research is predominantly based on, you know, more epidemiology, population health and such, but also some modeling as well, but really just continuing to hone in on the

[01:40] subject of gastric cancer and in particular, as a surgeon I'm kind of tired of seeing people coming in really late, especially coming from Korea where I saw people getting screening, coming early. You have all sorts of options like ESD or even surgery. Works for you.

[02:00] so much better when they come in not at stage three or stage four. So that's really the background of how I got into everything. Let's see. Next slide. Next slide. Okay, thank you. So I think everybody here knows.

[02:20] that gastric cancer is considered relatively rare even though it's not that rare. And because of that, the imperative for, you know, that's one of the reasons why prediction modeling may be something that we can really utilize because we have this need to identify who is at high risk in the United States.

[02:40] States. Next slide. And as other speakers have already emphasized, you know, in the United States it's really tragic how we find everybody very late in their cancers, and so trying to find them earlier is important. Next slide.

[03:00] We also know that gastric cancer is a very, very problematic cancer in the sense of disparities. This recent AECR disparities report points out that gastric cancer is ranking number one to number two for all race ethnicities in terms of mortality.

[03:20] rank immortality. So there is a doubling in mortality for African Americans, for Hispanics, as well as Asian Pacific Islanders. So we are double in mortality for all of those non-white populations. Next slide. And even though we

[03:40] have somewhat society guidelines. This one was published in 2010, suggesting that EGD for gastric cancer in immigrants should be done, especially if there's family history. Really, we haven't had a lot of traction on this. Next slide. And the stand-in.

[04:00] The importance of the National Cancer Institute is that there's inadequate evidence of benefit associated with screening and that the evidence from East Asia is consistent with the reduction. However, studies from high risk regions may not be applicable to low risk regions. Next slide. They go on to say that

[04:20] There's considerable discussion about how much incidence would make the examination worthwhile, even though they recognize that there are some populations at higher risk. Next slide. So I think that this all put together goes to say we really need better data. And I think all of my talk really is going to end up with a conclusion.

[04:40] that we really need better data about the risk factors that are important for gastric cancer in particular because there are some unique risk factors that are not collected in the major cancer registries or even a lot of the cohort data that we have.

[05:00] So the questions really I think that's burning are how do we screen in this low risk region as the US? And so thinking about how much incidence is really worthwhile are really big questions that we have to answer. I just wanted to point out that when we're thinking about

[05:20] prediction tools and such. It took me a long time to finally kind of wrap my head around and maybe this is really basic for epidemiologists, but it wasn't for me. I kind of struggled with this for a long time. This notion between a risk prediction tool and a screening tool. And I say this because when we see a lot of the models that are out there,

[05:40] out there, it's very confusing to think about are you talking really about a risk prediction tool which is to see who is at risk when they're getting followed up. Not necessarily that they have it now, but when you're following somebody up. Versus do they have it now? And so the only way you can really get to a risk prediction tool is if you have longitude.

[06:00] no data, not case control study data. But as you'll see in my presentations, we don't have a lot of it. And so right now in the United States, we do not have the data to really create a risk prediction tool. Next slide. I'm not going to be talking about

[06:20] about the use of endoscopy. There's going to be a couple of things I'm not going to talk about when I talk about with this talk. Endoscopic findings, I left out of this talk because in my mind at least, I really wanted to hone in on the question of how do we detect people who are asymptomatic, don't have an EGG screening.

[06:40] you're trying to get them to there, right? So outcomes other than gastric cancer, I also left out. There's also a lot of models on prognosis for gastric cancer, lymphomatosys, surgical outcomes, which none of this talk will not cover. Next slide. So starting off with risk prediction tools. Next slide. So when we're talking about

[07:00] about risk prediction tools, there's a couple of different approaches. The big kind of buckets are the models that use basically survey-based or EMR-based data. And then there are some, and then the other big bucket is the ones that are using biomarkers, so blood. And then obviously you have a, you know, a mixture of the two.

[07:20] Next slide. So starting off with, you know, EMR and clinical-based models. So in there have been this study out of Korea looked at the national health insurance data with a median of 11 years follow-up, looked at 1.3 million men and 800,000

[07:40] women and it was a 10 variable model that looked at age, BMI, family history, meal regularity, salt preference, alcohol, smoking, physical activity and ended up with a C statistic of 0.76 which in modeling world is about moderate. Then there was

[08:00] another one, also out of Korea, this was more a case control study. Once again, you know, as I said, you really need longitudinal data for this. But this case control study utilized 182 cases and 199 controls and utilized a lot more variables in their model and results

[08:20] resulted in an ROC of 0.9, which probably is a result of overfitting. Next slide. In the United States, we really don't have anything. So one of the first studies that I did with some grant support from the Alliance for Clinical Trials and Oncology was to think about how to collect and think about the variables that we should

[08:40] collecting in the United States. So this was one of my first projects and so I was really naive about it. I threw in 227 questions and went through the whole literature search, did focus groups, cognitive interviews, translated into four different languages, and finally

[09:00] Finally got a pilot study up and going of 30 cases and 60 controls in two different hospital systems and did this and then this survey. Next slide. So in this case control study model, you know,

[09:20] I, one of the key things that we asked about was, thank you, was different variables. How do you capture these variables? And so once again, because of my nigh activity in a way, I was very expansive on the type of variables I asked. And in particular, I asked about

[09:40] But not just whether or not you are an immigrant, but what country were you coming from? And then stratify that by the incidence that that country actually, the incidence of the country that they were coming from, as well as cultural foods. And looking at, we collect

[10:00] of course, what they were eating at every age, but found that what they were, if they were eating cultural foods in their teenage age was really what was most predictive. And so bringing that together with a five variable model, including age, cultural foods around your teenage years, education, what generation you are in acculturation.

[10:20] And we were able to, I was able to get this model to make a positive predictive value to be akin to the incidence rate of Korea and Japan, basically saying that we can identify, isolate a group of people who have the same risks as the population in Korea. If they're screening, why can't we

[10:40] screen these people. Next slide. I also threw in these models using what used you know added on these variables to an existing model called the Harvard Cancer Risk Index. This is a risk assessment tool for asymptomatic patients that was developed for primary prevention.

[11:00] It had the traditional variables that we know about age, gender, family history, body mass index, excessive salt intake, alcohol, smoking, blood type, and helicobacter pylori. And using the data I collected, we added on the ethnic and cultural variables that we found to be predictive. Next slide.

[11:20] Next slide. Thank you. And with that, we found that with the Harvard model alone, the positive predictive value was 28 per 100,000. Once you started adding in these variables and then got an abbreviated model of 10 items.

[11:40] We got the positive predictive value to be 81 per 100,000, once again, akin to Korea and Japan. So emphasizing how important it is to understand, fully understand immigration patterns and acculturation. Next slide.

[12:00] The biomarkers that have been examined mostly revolve around helicobacter pylori and pepcinogen testing. As we know, most gastric cancer, especially our intestinal types, is associated with helicobacter pylori. And so being exposed to helicobacter pylori is a marker of risk. And the

[12:20] And then, pepsynogen I, II, the inactive proenzyme of pepsyn, is what's really that the decrease in the pepsynogen level is reflecting the mucosal changes. And so, the decreased levels of pepsynogen I and II indicate atrophic changes of the mucosa.

[12:40] and numerous case control studies have shown that the combination of helicobacter, pylori, antibody, and pepsynogen work very well, and with a pooled odds ratio of four in more than 27 different studies, but most of them have been outside of the United States. There's a lot of evidence that the case control is not being used in the case control. It's not being used in the case control. It's not being used in the case control.

[13:00] also some efforts on other biomarkers such as SNPs or single nucleotide polymorphisms as well as antibody response to helicobacter pylori. And before we change slides, I just wanted to, I'm going to be talking about the ABCD groupings. So for people

[13:20] who've not heard of this, the Japanese have worked very diligently to examine how to combine the helicobacter pylori and the pepcinogen to risk-stratify the groups. And depending on the status of the helicobacter pylori and the pepcinogen, have developed four different

[13:40] groups for risk ABCD. Next slide. So in this study, in this very large study of a population-based prospective study with a 20-year follow-up in a county in Japan, there were

[14:00] 2,742 participants and helicobacter pylori and pipsinogen was used in this ABCD group pattern. And they found that the hazard ratio for group B was upwards of 4 and for group C and D had a hazard ratio of 1.

[14:20] And you can see in that, in that, in the curves there, the splitting of groups A, B, C, and D based on what their biomarkers were. They combined this with a basic model, you know, incorporating the demographics that we know.

[14:40] So in a study that used the Japanese public health cancer cohort too, there was 19,000 participants where after a cumulative, their model didn't exactly put how many years this was followed up, but it was long enough that they were able to create

[15:00] a model that looked at a 10 year cumulative probability of gastrocancy occurrence and developed a risk scoring system. So the risk scoring system again is on the right hand side. You can see that for the variables that are in the model that a score was assigned. And with that using a cutoff score of 16 or more.

[15:20] or the sensitivity and specificity was 70, sorry, 69% and 70%. The paper below that shows the external validation of this using a previous cohort and the.

[15:40] Basically, the model was very stable with similar sensitivity and specificities. Next slide. Similarly, a study out of China that also followed people, this was data from a population-based screening program where the median

[16:00] and follow-up was for 11 years and with 12,000 patients they found 0.7 had developed gastric cancer and they developed the model with this. On the right side is the scoring system that they used. They used five different biomarkers, pepcinogen 1, 2, the ratio, helicobacter pylori sero T.

[16:20] to zero positivity and gastrin, 17, and had a C statistic of 0.803. And if you had a score greater than 14%, your gastric cancer rate was 122 per 100,000, with a sensitivity of 30% and specificity of 97%. And obviously, depending on the...

[16:40] cutoffs this altered. Next slide. Some efforts have also been made to not just use pepsynogen and helicobacter pylori, but also using singling nuclei polymorphisms. This study looked at this using a case control study methodology and had a use

[17:00] of 0.65, so still into works. Next slide. Another very interesting study. This study looked at antibody response to helicobacter pylori and they had a helicobacter pylori multiplex serology.

[17:20] that looked at human antibody to 13 different helicobacter pylori proteins. And with this, they were able to look at a cohort, very extensive cohort called the helicobacter pylori biomarker cohort consortium and use that data to apply this.

[17:40] and found and was able to develop a six variable model with a AUC of 73.76, sensitivity of 73% and specificity of 60%. Next slide. So with all that said, those were all studies outside

[18:00] of the United States. So within the United States, there have been several, a few case control studies that were able to examine helicobacter pylori and pepsynogen as risk predictors, mostly in the Japanese American population in Hawaii, as well as in the US.

[18:20] as a study out of Alaska, showing that helicobacter pylori and pepsynogen were possible biomarkers. Next slide. So we were able to conduct what seems to be the first study using this, using, looking at pepsynogen in a prospective study.

[18:40] body. And so we recently published this where we looked at the data that was collected for the PLCO, cancer screening trial. So that was a randomized clinical trial of cancer screening, conducted over 10 centers in the United States, and enrolled $155,000.

[19:00] men and women that were ages 55 to 74 and they got randomized to a screening study. And the blood collected at the time of enrollment was used to test helicobacter pylori and pipsinogen. So we conducted a nested case control study.

[19:20] 205 gastrocancer cases that had developed and 209 match controls were identified. A medium followup was 13 years. Next slide. And with this, we found that pepcinogen positivity conferred an odds ratio risk of 10.6. So very pro-

[19:40] predictive of gastric cancer. Obviously small sample, but one of the first to have prospective data in the United States. Interestingly, helicobacter pylori antigen was not predictive, but that might be due to different cutoffs that we need. At this point, at this analysis, we had used standard cutoffs.

[20:00] that are being used in Asia. Next slide. I have a smaller study that is, that we're still in the works, but this is some preliminary data that I could share. So in my former institution, we, I had collected patients

[20:20] who developed gastric cancer, who got gastric cancer and collected blood before they got any treatment. And with this, and I also had, and we also designed this as a case control study, so collected controls as well. So with 101 participants of which 33 were gastric cancer cases,

[20:40] We tested helicobacter pylori and pipsinogen as well as a couple of other different biomarkers. This population in particular was when I was working in the Bronx and so we had a lot of minorities, 15% black, nearly 30% Hispanic, not a lot of Asians unfortunately.

[21:00] And with this, try to create a model that has very high sensitivity as well as negative predicted value. And pepsynogen positivity and helicobacter pylori, or having either pepsynogen positivity or helicobacter pylori positivity

[21:20] confer the best ability to get a really high sensitivity and negative predictive value upwards of 85% for both. Next slide. So once again, with the prediction tools, you know, we, you can see that we don't have a lot of data

[21:40] to be able to do that. So we really need some better data to develop these. From a screening tool perspective, next slide. So we currently are talking about endoscopy currently as the screening tool. However, in the United States, if you think about it,

[22:00] Endoscopy is very expensive and if there was a way to get it, you know, to screen in a cheaper way, that would be ideal. There is a lot of excitement right now around these multi-cancer detection tools. So I decided to, you know, so I took a look at the two major ones.

[22:20] cancer-seq and GRAIL. And what you can see is that, so this is particularly data for stomach cancer. And you can see that for cancer-seq, they actually did not even report stomach cancer. For GRAIL, you can see that for stage zero, they didn't report the data. For stage zero, they did not report the data.

[22:40] stage one, their sensitivity was extremely low and it was only when you got to the higher stages of cancer that these detection tools could actually detect the cancer, which isn't a very good thing for a screening study because with screening studies, the real point that you want to get to is to detect them early. Next slide. So one of

[23:00] One of the more exciting papers that I kind of ran across, and I'm trying to see if we can use some of our data to validate this work, is a serum microRNA biomarker that was used for gastric cancer detection. This was a multicancer study out of Korea.

[23:20] and Singapore. They did it as a three phase study where they first had a discovery phase and a validation study, validation phase followed by a prospective study. They started off with a lot of mRNA candidates, boiled it down to 12 micro RNA biomarker panel.

[23:40] And in their prospective study, the sensitivity was 87% specificity of 68% with an AUC of 0.848, which is actually quite good. Next slide. And if you put it together and compare it with the cancer-seq and GRAIL, you'll see that they report

[24:00] reported that they report a sensitivity that was quite excellent even in the very earlier stages of cancer with high grade dysplasia at 60%, stage one and stage two nearly at 90%. Next slide.

[24:20] talk briefly just as ending some clinical trial ideas that some of you know that I've been kind of working with and trying to get up and going. We have to really generate good data. We're going to need clinical trials, but it's been extremely hard to

[24:40] push this forward. Next slide. So in the United States, to develop the clinical trials that we really need is going to be to develop risk prediction models. So we're going to, but we're going to need data.

[25:00] that's going to be able to capture the variables that are important, such as race, ethnicity, immigration, acculturation. We're also going to need some blood-based biomarkers probably. And so, you know, the other big questions for clinical trials is going to be, does EGD work? You know, what kind of biomarkers work? Next slide.

[25:20] So, after a lot of, you know, trial and error, this is one of the clinical trial schemas that we came up with. You know, when we looked at gastric cancer alone, it was very difficult to justify the number we would need to be able to

[25:40] effectively prove that EGD actually worked. So one of the clinical trial areas that we're missing is, does screening actually work for this population? So we came up with a study design where we combined the risk of esophagia disease,

[26:00] cancer as well as stomach cancer to see if we could to screen for them. And unfortunately, in order to get to a cancer-specific mortality reduction of 30%, the sample size needed to be 88,000 patients. And so this was a little difficult to justify.

[26:20] identify without more data. Next slide. So, more recently, I'm currently working on this study that I hope that people might show some interest in because we're going to be expanding this to other sites as well. So, right now,

[26:40] The pilot study is the idea is for an opportunistic EGD. So persons getting screened for colonoscopy will be screened if they're, whether or not they're high risk and offer them addition of EGD. Since this is a pilot study, the numbers are still small.

[27:00] But so with that, the aim is really going to be to determine the feasibility and safety of EGD at the time of colonoscopy. You know, is there added, how much added time is needed? Are there adverse events? Aim 2 is the documentation of pre-cancer and cancerous lesions. Aim 3 is to examine the predictors of uptake for the EGD.

[27:20] So beliefs and attitudes, perception of risk, moderators, motivators and barriers. And also patient preference and satisfaction with getting the EGD at the time of colonoscopy. With the exploratory aims, you know, this will be an opportunity for us to get blood as well as other samples.

[27:40] And so, examining if pepsynogen is a biomarker of gastrocancer and pre-cancerous lesions in the United States, as well as potentially working with Gastroclear in their validation study. Next slide. So just in summary and closing,

[28:00] identifying persons at high risk for gastric cancer screening is necessary to improve gastric cancer mortality in the United States and pips in a helical bacter pylori have potential but really need to be examined for US populations and optimal risk cutoff for risk prediction tools will depend

[28:20] depend on the cost of the screening test. So currently, EGD is very expensive, and cheaper screening tests will allow for lower cutoffs, allowing more people to get screened. So thank you. APPLAUSE