Hello. On behalf of Dr. Paul Kuo and me, I'd like to welcome you to the 2020 Liver Meeting Clinical Research Workshop on Research Methodologies Using Alcohol-Associated Liver Disease as a Model. Our objective is to provide an overview of laboratory-based and clinical research methodologies using alcohol-associated liver disease as a model. We have a very exciting program developed for you with a star-studded cast, starting with Dr. Liang Punsakal, who will be giving a talk on translating findings from bench to bedside to identify biomarkers for alcohol-associated liver injury, followed by a talk by Dr. Szabo speaking on bench-based research methods and strategies to assess mechanisms of injury of hepatocyte inflammation and hepatic fibrosis. Following these two talks, we'll switch to clinical research methodologies with a talk by Dr. Theroux on developing prediction models for clinically meaningful endpoints in patients with alcoholic hepatitis, followed by Dr. Ellinger giving a talk on qualitative research methodologies to understand barriers to sobriety in alcohol use disorder. And last but not least, we will close out our session with a lecture by Dr. Kim on leveraging large administrative databases and big data methodologies to develop cohorts. We hope you enjoy the program. I would like to thank ASLD and course organizers, Dr. Pao Kuo, Jennifer, for the invitation at this year's clinical research workshop. The topic of my talk today is translating findings from bench to bedside to identify biomarkers for alcohol-associated liver injury. This is my disclosures. Alcohol-associated liver disease, or ALD, is one of the leading causes of chronic liver disease. It is a broad spectrum of disorders ranging from simple fatty liver to more severe forms of liver injury, including alcoholic steatohepatitis, advanced fibrosis, and cirrhosis. Fatty liver is an early response to alcohol consumption, develops in most, if not all, heavy drinkers. Interestingly, only about 30% of heavy drinkers develop more severe forms of ALD and cirrhosis. To determine what level of alcohol consumption is considered too much, we should first understand the concept of standard drink. Standard drink is varied by countries. Different types of alcoholic beverage can have very different amounts of alcohol content. That's why it is important to know how much alcohol each drink contains. In the United States, one standard drink contains roughly 14 grams of pure alcohol, which is found in 12 ounces of regular beer, 5 ounces of wine, and 1.5 ounces of distilled spirits. The study has shown the gender difference in the risk of alcohol-associated liver disease in men drinking above 40 grams and in women above 10 grams of alcohol, increasing risk of alcohol-induced liver injury. So if we go back to the definition of one standard drink of 14 grams, moderate drinking is defined as men drinking up to two drinks and in women up to one drink per day. Excessive drinkers, per NIAAA definition, are defined as men with more than or equal to four drinks and in women more than or equal to three drinks per day. First, let us visit if there are any non-invasive biomarkers to screen for excessive alcohol use. Non-invasive biomarkers to screen for excessive alcohol use can be stratified into two categories, acute and chronic alcohol use biomarkers. Acute alcohol biomarkers are those with a very short window of assessment, such as blood and urinary alcohol concentration, or the measurement of alcohol metabolites, such as ethyl glucuronide and ethyl sulfate. Because its presence is relatively to the timing of acute alcohol ingestion, these are not useful as a marker for chronic alcohol consumption. For this latter group, there are commonly used biomarkers, which are quite nonspecific, such as AST, ALT, MCV, and GGT. It is not surprising that the sensitivity and specificity of these tests are low. There are two non-invasive markers which are mechanism-based with better diagnostic performance, such as carbohydrate deficient transferrin. Transferrin molecules in the blood usually contain several carbohydrate components, the process which is inhibited by alcohol, resulting in an increasing in the percentage of CDT among excessive alcohol users. CDT has a window of assessment about two to three weeks. Phosphatidyl ethanol, or PEF, is formed directly after alcohol intake via the enzyme phospholipase D from phosphatidylcholine in the presence of alcohol, and this has a window of assessment about one to two weeks. Why is the identifying non-invasive biomarkers for alcoholic liver disease important? This is because ALD is rarely detected at the early stage. This is a paper by Ramon Batalla and his colleagues looking at almost 3,500 patients worldwide with chronic liver disease. Only 3.8% of patients with ALD were seen at early stages, whereas 29% of these patients were seen at advanced stages, suggesting that ALD is normally diagnosed at the later stage when liver-related complications developed. First, let's look at the diagnosis and evaluation of steatosis in ALD. ALD is accepted as an initial screening modality for fatty liver with good sensitivity and specificity, depending on the severity or degree of steatosis. Magnetic resonance spectroscopy, or MRS, quantifies the proton density fat fraction, a standardized measure of liver tissue. It is highly accurate and reproducible in measuring hepatic fat, with a sensitivity and specificity around 80 to 90%. Controlled attenuation parameter, or CAP, is a recent tool to non-invasively assess liver steatosis. The CAP software is incorporated into the fibroscan equipment during the measurement of liver stiffness, which enables CAP to be combined with non-invasive liver fibrosis assessment. The area under the curve of CAP to diagnose steatosis for ALD is around 80%. What about non-invasive biomarkers for fibrosis measurement in patients with ALD? There are several non-invasive blood tests which have been studied to determine the degree of underlying fibrosis as shown in this slide. Overall, depending on the cut-off being used, the sensitivity of these tests to screen or detect for fibrosis or cirrhosis are not ideal, with a higher sensitivity around 80%. Liver stiffness measurement has become one of the non-invasive tools to screen for liver fibrosis or cirrhosis. However, despite its wide use, there are multiple factors affecting the measurement, such as underlying hepatic inflammation, cholestasis, food intake, and more importantly, the timing and quantity of alcohol consumption itself. In this meta-analysis, the study shows a better performance of liver stiffness measurement to detect F4 fibrosis in patients with ALD. However, the optimal cut-off varied drastically, ranging from 11.5 to 25.8 kPa. This is because of underlying hepatic inflammation indicated by the level of AST. In patients with low or normal AST, the cut-off values were comparable to those observed in patients with viral hepatitis. The cut-off values increased exponentially as a function of median AST level. What about non-invasive markers for alcoholic hepatitis? There are two components for alcoholic hepatitis diagnosis, history of alcohol consumption, and clinical presentation or lab tests. We expect the excessive alcohol use history should have occurred for more than six months and within 60 days of abstinence before the onset of jaundice. Serum bilirubin cut-off is more than 3 milligrams per deciliter, with the AST more than 50 international units per milliliter and AST to ALT ratio of more than 1.5. Liver biopsy is not always required unless the clinical diagnosis is unclear. Recently, the NIAAA alcoholic hepatitis consortium proposed a classification of alcoholic hepatitis diagnosis into three degrees of confidence. Definite alcoholic hepatitis means that they have clinical and lab comparable with the definition, and these patients had biopsy-proven alcoholic hepatitis. Probable alcoholic hepatitis are those with clinical and lab fit the definition but have not had liver biopsy and without confounding factors such as the presence of autoantibodies, sepsis, shock, or drug-induced liver injury. Possible alcoholic hepatitis are those with compatible clinical diagnosis but with other confounders. In this group, liver biopsy is recommended to confirm the diagnosis. What about the lab-based assay for the diagnosis of alcoholic hepatitis? Mollulidenc bodies are cytoplasmic inclusions representing rearrangement of the cell cytoskeleton. Their presence is considered highly suggestive of alcoholic hepatitis. A main constituent of Mollulidenc bodies is keratin 18, which can be detected after hepatocyte cell death from apoptosis or necrosis. Two forms of K18 fragments can be detected. The M65 represents both full-length and fragmented forms, and M30 detects the cleaved fragment. This is the study from clinical sites in France and Spain, including 83 patients in the test cohort and 68 patients in the validation cohort, to test the utility of several non-invasive markers to diagnose alcoholic hepatitis. All patients underwent liver biopsy and were reviewed by three expert liver pathologists to stratify patients in each cohort into no AH and with AH. Among several non-invasive markers, K18 has emerged to be a useful test for the diagnosis of alcoholic hepatitis. Patients with histological proven AH had higher levels of serum M65 and M30. Using the appropriate cutoff for either M65 and M30, the presence of alcoholic hepatitis could be ruled in or ruled out in over two-thirds of the patients, with their diagnostic accuracy in the range of 80 to 90 percent in either test or validation cohort. Taken together, this study suggested that plasma level of K18 fragments are reliable non-invasive markers for alcoholic hepatitis. What about the scoring system for prognosis? There are multiple scoring systems as a non-invasive marker for prognosis in patients with alcoholic hepatitis, either using alone or in combination. However, these do not reflect the magnitude of cell death or the form of cell death, which may be important in distinguishing various forms of liver injury and may impact the drug therapy. So K18, as mentioned earlier, is not only useful as a non-invasive biomarker to diagnose patients with alcoholic hepatitis. In this large multi-center prospective study from the NIAAA-sponsored DASH consortium of 84 patients with moderate or severe alcoholic hepatitis, both K18 biomarkers perform better than the traditional biomarkers with regard to the area under the curve in predicting three months mortality. What about non-invasive markers for prognosis in alcoholic cirrhosis? MEL score is widely used. However, using a high-throughput transcriptomic analysis, we found two long non-coding RNAs, AK12A652 and AK054921, with the same pattern of expression in both peripheral blood and in the liver, and they are independently associated with survival. Using the estimate regression coefficients, we constructed the RISC score and found that this linked RNA-based model had achieved statistics of 0.9 compared to 0.82 for MEL in predicting survival in patients with alcoholic cirrhosis. Summary, I would like to break down the non-invasive biomarker discoveries in this area into three categories. First is a diagnostic biomarkers. Because ALD is a spectrum of histopathological changes, the phenotypes matter. For alcoholic cirrhosis, imaging-based diagnostic modalities seem to be a reasonable tool. For alcoholic hepatitis, K18 seems to be a good diagnostic biomarker. Liver stiffness is a very good tool to detect liver fibrosis. However, we are still in need to explore other diagnostic markers to detect early fibrosis among heavy drinkers. Second is the prognostic biomarker. Several scoring system has been used to predict prognosis in alcoholic hepatitis and alcoholic cirrhosis patients. However, we should also explore other mechanism-based biomarkers similar to K18 to providing more accurate diagnosis and prognosis in these patients. Lastly, we are in need of the biomarkers to predict treatment responses in ALD, notably in alcoholic hepatitis patients. We have new score as a predictor for those who responded to steroid therapy. Several new therapies are being explored in this area. Future areas of research should also focus on biomarkers in predicting treatment responses for specific therapies so that the right treatment can be administered for the disease with high mortality. I would like to thank NIAAA and the VA to support part of the work being presented today. And thank you again to the ASLD and course organizer for the invitation to be part of this year clinical research workshop. Thank you very much. Welcome to the clinical research workshop. My name is Junji Sobo, and I have the pleasure to talk about the bench-based research methods and strategies to assess the mechanism of injury of hepatocyte inflammation and hepatic fibrosis. I'm a professor of medicine and faculty dean of academic affairs at the Harvard Medical School and serve as chief academic officer for the Beth Israel Deaconess Medical Center and VIH Health. My disclosures are listed here and pretty much have nothing to do with the current presentation. So the topics of discussion for today will include assessment strategies and the bench-based research methods that will target mechanisms of injury, inflammation, and fibrosis in liver diseases. And on bench-based research methods, I will talk about standard assessments, discovery approaches, and other novel methods to assess the injury, inflammation, and fibrosis. So I'm going to start by talking about strategies to assess mechanisms of liver injury. Obviously, you're very familiar with that several biomarkers and mass models are available to assess liver injury. So biomarkers in the circulation, some of them are very obvious, such as ELT, EST, and lactate dehydrogenase levels that can indicate, not necessarily specifically, but typically used as biomarkers of liver injury. More recently, there has been attention on certain microRNAs, and I would like to particularly point out here microRNA 122. That is the microRNA that is very abundant in hepatocytes. And in fact, pretty much any kind of hepatocyte injury results in a rapid release of 122 that could be found in the circulation. When one assesses cell death, then various cell death-associated molecules that could be evaluated. And depending on the cell death type, this could become very complicated. And as you know, the cell death comes in many different flavors. It could include apoptosis, necrosis, necroptosis, pyroptosis, and so forth. In inflammation, we are looking for immune cell populations. And the result of the inflammatory cell activation typically is a release of cytokine and chemokines, and measurements of these is relatively easily available. And finally, if one is interested in various types of models for acute versus chronic liver injury, then typically we can use the alcohol binge, for example, in the alcoholic liver disease area as an acute injury, or a combination of that with chronic alcohol. Acetaminophen-induced liver injury is a classic acute liver injury model. And in case of chronic injuries, they use chronic alcohol plus-minus binge. There are various models for non-alcoholic liver disease that are diet-induced or genetically-based. And for fibrosis models, typically the chronic carbon nitrochloride administration or bind-off ligation are used as prototypes of these type of liver injuries. So if you look at the markers of liver injury, then the biomarkers in circulation could be the ELT, EST, and LDH. And typically, these are enzymatic activities that could be measured in the laboratory and not necessarily only in the clinical laboratory setting, but in the basic labs too. And the microRNA-122 is another good measure for this. Liver injury, particularly in alcoholic hepatitis, it has been shown that cytokeratin-18M65 increase actually is a very good marker of liver and hepatocyte injury. And more recently, if you look at the microRNA-122, it appears that in animal models, it very much mirrors changes in ELT. And this increase in serum microRNA-122 can be found both in the plasma and serum, but most of this is actually in exosomes, which are small membrane-bound vesicles that are released from injured hepatocytes. And these exosomes and extracellular vesicles are typically enriched in microRNAs. So this could be another way of assessing liver injury. Now, when it comes to evaluation of liver injuries, you may need to keep in mind what type of cell death could occur in the particular model that you are looking at. And here, the regulated cell death pathways are listed. And as you appreciated, it's very complex from an apoptotic morphology going to the necrotic morphology that is mirrored by the intracellular signaling pathways that are involved in these various processes. So in a 15-minute talk, I don't have time to get into details, but to give an idea, in apoptosis, typically caspase-8 activation is one of the markers, and also caspase-3. And the TNF-induced cell death is another type of cell injury that can trigger either survival or death pathways. And another way of inducing apoptosis is through defective complex 1 formation that results in, again, caspase-8 activation and or in necroptosis that is a cascade of activation of RIP-1 and RIP-3 kinases and induction of NRK phosphorylation that could be another way of measuring these type of specific cell death pathways. Cell death and inflammation could be linked. So for example, pyroptosis is a cell death mechanism that is triggered by inflammasome activation, for example, through NRK-3 or to the P27X receptor. And here, the caspase-1 that is induced by inflammasome triggers gastroimmune D activation and cleavage. And this N-terminal gastroimmune D actually results in a poor formation to lead to the classical canonical pathway of pyroptosis. But we know now that caspase-11 activation can also induce gastroimmune D to trigger a non-canonical pathway activation for pyroptosis. All of these have been indicated as potential mechanisms, for example, in alcohol and liver disease. So I just wanted to show this recent paper where the microRNA208 deficiency was assessed and in a very acute carbon tetrachloride-induced liver injury model and just show that how liver damage, necrosis, apoptosis and inflammation markers can be sort of assessed to come to a picture. And here, it was found that liver damage is induced by indicated by increased ALT. Extracellular vesicle number was increased, indicating, again, liver damage. Necrosis and apoptosis markers in the form of NF-kB and TNF activation. Upregulation of BEX, caspase-3 and caspase-8 could be another marker of liver injury. And looking at inflammation markers in the form of cytokines. And all of this kind of correlates with some of the changes that you can see, for example, in this necrotic liver injury in the acute CCL4-induced model. Now, in addition to injury, inflammation is another parameter that we often assess in liver damage. And in fact, inflammation and cell death can be very much linked as I showed you earlier. So for assessment of inflammation, one can look at the composition of the various immune cell populations. In normal homoestasis, Kupfer cells are in the liver, but in response to any kind of inflammatory signal, recruitment of neutrophils, peripheral monocytes and monocyte-derived macrophages in terms of polarized M1 or M2 macrophages are recruited to the liver. So one can use looking at the cell types and using the various markers that are expressed on the surface of these cells to assess the frequency of these cells in the various liver disease conditions. Some of the assays that we use for this include immunohistochemistry, where you can just visualize the liver pathology and quantify population of immune cells by using immune cell-specific antibody staining. Then one can take the next step to actually isolate at these immune cell populations from perfused liver and use those in a fluorescent activated cell sorter that can quantify the distinct cell populations of cells. And here, multiple different color-linked antibodies can be used to do complex assessments of these various macrophages or neutrophil or Kupfer cell populations. And using FACTS analysis, actually, we can also do what's called the image stream analysis system through MNIST that allows individual cell analysis by flow cytometry. And here, we can look at fluorochrome-labeled proteins, for example, NF-kB, and look at the nuclear translocation within the cells through this system that can give yet another way of evaluation of this cell activation. In addition to that, intracellularly expressed cytokines can also be combined with the surface marker analysis in FACTS to get a sense for cytokine production. So inflammation can also be assessed in terms of the produced cytokines and chemokines. And here, we're just showing a not exhausting list of pro-inflammatory, anti-inflammatory cytokines and chemokines. And in terms of measurement for these, one can look at the protein levels and do quantitative assessment of that through either ELISA methods or multiplex that allows up to 40, 50 different cytokines being measured in the same sample. And this multiplex becomes very, very useful when you want to assess a small, limited amount of sample. Less quantitative is the Western blot analysis or proteomic analysis. That particular proteomic analysis can give you a large number of information using various existing scans, like the somatologic scans or using mass spectrometry. Cytokines can be easily measured at the messenger RNA level. Obviously, the messenger RNA measurement shows a transcription regulation and not necessarily a biologically active product. And here, if you want to look at a very quantitative measurement, the ddPCR is the method for that. But typically, what we use is the quantitative qPCR that provides the semi-quantitative analysis of messenger RNA expression relative to a known housekeeping gene. All of these messenger RNA measurements can be done kind of in a larger scale by using various platforms in nanostring analysis. And if somebody is interested in unique levels of messenger RNAs or RNAs in general, then can do an RNA-seq, more of a discovery type of analysis. Recently, on the single-cell transcriptomics is a method that's being used that essentially allows single-cell populations and information on these various cell populations with regards to their transcriptome profile that provides yet another deeper layer insight into the inflammation assessment. I also wanted to talk a little bit about non-coding RNAs that could give information about not only inflammation, but in general about changes in the liver. And these would include microRNAs, new RNAs, and so forth that we may not have time to talk about. Just wanted to show you how biomarker discovery, for example, in a context of a disease can help. And here, this was a study where multiplex analysis was used for looking at circulating biomarkers of inflammation, and in a large scale in about 100 patients. And using some of these multiplex platforms, we were able to assess key molecules that indicate inflammation. So, for example, soluble CD163, which is a protein cleaved off from activated macrophages or soluble CD206. HMGB1, which is a danger molecule or acetylpentene that's a cytokine, were actually related to various clinical parameters in alcoholic hepatitis. And we found that 90-day mortality, for example, was correlated with increased soluble CD14 levels in the circulation. Infection could be predicted by increased soluble CD163 levels in the circulation of patients with severe alcoholic hepatitis. And those who had increased acetylpentene level were more likely to develop organ failure. So the large-scale evaluation of these potential inflammation-related biomarkers could help us in the future to prognosticate patients and potentially to even assess responses to therapy. Another area where the inflammation, injury, and fibrosis markers could be kind of looked at in a concerted way is looking at microRNAs. And many of these microRNAs are packaged in exocellular vesicles that are emerging biomarkers in alcoholic liver disease, but other types of liver disease as well. And particularly, the various circulating microRNA signatures in alcoholic liver disease, at least in mice that have been identified, include increases in miR155, increased miR122, miR192, and miR38 levels. In humans, some of these were correlating with the finding from mice that increased miR192 and miR38 levels actually are increased in the human exocellular vesicles as well in alcoholic hepatitis. So another way of getting deeper in the biomarker and injury analysis is to look at the protein profiling of plasma or exocellular vesicles. And here, this is a study that is under development that we're comparing plasma protein and exocellular vesicle protein expression in alcoholic hepatitis patients compared to normal controls. And as you see, the protein profiling allows us to identify a cluster of proteins that are unique to alcoholic hepatitis compared to normal controls, and there is some overlap here. Furthermore, then we can take it a step further and ask the question whether looking at the protein profiling of plasma or potentially looking at the isolated exocellular vesicle protein profiling could provide us more information. So as a next step, we are comparing the exocellular vesicles compared to the plasma in these patients. And then, of course, the picture can be further complicated by comparing the various severity of the disease between mild or severe alcoholic hepatitis. But these type of measurements identify numerous unique or known proteins and pathways that can be further analyzed and allow us to get some either biomarkers or further insight into the pathology of disease by identifying pathways and proteins that are either upward or downward related. So finally, I'm going to talk about some strategies to assess liver fibrosis. And here, the typical methods that we use include serious red staining, trichrome staining, measuring hydroxycholine activity in the liver as a matrix remodeling marker, looking at fibrosis markers such as alpha-synthetic actin and collagen expression at the messenger or other protein level, looking at tissue remodeling matrix methyl or protein expressions, or isolating cellic cells and myofibroblasts and looking at their characteristics. So again, here I'm showing just an example in a study where the inhibition of the CCR2N5 signaling with semi-criviroc was applied in a mass model of alcoholic liver disease. And as you see that on the left, we're looking at ELT levels as markers of the liver injury and find that when the inhibition of CCR2, CCR5 was introduced, then all of a sudden we find normalization of the ELT levels in alcohol-fibronized. This nicely correlated with a reduction in infiltrating macrophages and those were using a flow cytometry analysis, the isolated cells and measuring the frequency of F480 low and CD11B high cell populations. So these are surface markers that identify infiltrating macrophages and that nicely correlated with changes in liver TNF protein levels. Again, showing that inhibition of CCR25 attenuates the recruitment of macrophages to the liver. Furthermore, in a chronic alcohol feeding, we find that this semi-criviroc administration attenuated the liver steatosis on an ORADO staining. And on the top right, you see that the serious red staining indicating liver fibrosis nicely demonstrated that this semi-criviroc can attenuate and essentially prevent alcohol-induced fibrosis in this mass model. So these are kind of ways of showing how the various indices of liver injury, inflammation, and fibrosis link together. So finally, I just wanted to tell you about new approaches to assess liver injury and some of the discovery approaches including mass spectrometry where we can identify protein structures and modifications associated with disease. The various omics platforms, proteomics, transcriptomics, and lipomics could be listed here that will assess these unique characteristics. New methods that are more frequently used include CyTOF, which essentially is a mass cytometry method that can detect up to 40 plus parameters and quantify cell lineages, markers of differential function and activation. Single-cell transcriptomics is another way how messenger RNA expression profile can be identified in isolated cell populations. And spatial transcriptomic, which allows messenger RNA expression profiling in a spatial location within a tissue sample. So that, again, is a very exciting and new method that is gaining a lot of attention. So the key takeaway is that if you want to do a quick and easy assessment, then probably looking at EST, ELT, and HNA staining will give a good assessment on the type of injury. Looking at TNF, IL-1, MCP-1, either at the qPCR or ELISA level will briefly assess the presence of inflammation. And alpha-smooth ethylethylamine and collagen measurements at the qPCR or ELISA level will give a very good rough estimate of fibrosis. When the types of hepatocyte injury is the goal to be assessed, then think about the fact that there are many different and interconnected ways of cell death mechanisms, and you need to look for signature of the various death pathway molecules. Typically, confidence in any kind of finding is increased if complementary assessment methods are used, or different methods are used to confirm the same finding. So I would suggest that before selecting strategies to assess liver injury, inflammation, or fibrosis, you should ask a few questions. Number one, that are you looking for quantitative changes of known molecules? Because that will inform you what type of methods of assessment should you use. How much sample specimen do you have? Because, again, that may determine that if you have just a small sample size, then you may want to use, for example, a platform technology as opposed to individual measurements. And finally, are you looking for discovering novel molecules? Because in that case, you need to get a deeper dive in technologies rather than using just standard methods. So with this, I would like to thank you and enjoy the rest of the seminar. Thank you. It's a pleasure to be here, and I'm grateful to the organizers for inviting me. I'll be addressing developing prediction models for clinically meaningful endpoints in patients with alcohol hepatitis. I'm Nora Turo from the University of Southern California. These are my disclosures. Alcoholic hepatitis, or alcohol-associated hepatitis, has had a rich tradition in terms of prediction models dating back to the Madre discriminant function back several decades ago. Shown on this slide are others that have evolved over the years. Many of them focused on prediction of survival as being the primary outcome, or the need in particular for corticosteroids usage and or discontinuation. So this really speaks to what are clinically relevant endpoints for this disease entity. The natural history of AH is well known to you. Particularly in its severe form, it is associated with a high mortality. Shown here on the left is data on alcohol-associated hepatitis in the TREAT consortium. On the right is a study of which alcohol-associated hepatitis was treated with transplantation. You can see that the matched untreated controls have a pretty dismal survival. So the endpoint of survival is actually very useful and is very typical for that of studies in AH with early and late survival, both reasonable entities depending upon the question being asked. Other important endpoints can be recurrence of the entity of alcohol-associated hepatitis. It can be looking at progression of cirrhosis and liver-related complications. And it can be alcohol use if the intervention is more focused on alcohol use. I would say that these particular endpoints have become very relevant in the realm of looking at individuals with alcohol-associated hepatitis who have undergone liver transplantation. I just want to point out that alcohol use measures are evolving. It's still an imprecise science, I would say. Typically, validated self-reporting is used as an endpoint to measure alcohol use, but biomarkers are increasingly being considered. And in the future, perhaps biosensors will have a greater role. So in developing predictive models, I'm going to focus on two models, studying different endpoints and using different methodologies. And it's really by means of comparing and contrasting how you do model development. So the first is a little score. You're all very familiar with this. It was a score used in patients with severe alcohol-associated hepatitis who were treated with steroids. And the other I'm going to look at is the SALT score, which is used in patients with severe alcohol-associated hepatitis undergoing liver transplantation. And it's looking at predicting sustained alcohol use after transplant. So step one in developing predictive models is to, indeed, define a clinically relevant endpoint. And just to highlight that for these two predictive scores, the little score looks at six months survival among patients treated with corticosteroids. The SALT score, in contrast, is looking at likelihood of sustained alcohol use after liver transplantation. The target population for both these studies are patients with severe AH, defined by a discriminant function of 32 or higher. They did define their diagnosis of alcohol-associated hepatitis in that target population somewhat differently. But just to make the point that the endpoints here are quite different. Step two is to, you want to explain that clinically relevant endpoint. So you are going to look at the risk factors or the explanatory variables. Ideally, they should be based on an a priori hypothesis or hypotheses, literature that exists, or expert opinion. They should be readily available in clinical practice. Objective is preferred over subjective. Both baseline and or dynamic, meaning at some interval, can be incorporated. And the little score is a good example of that. And just to note that poor measures and the ones you should avoid are those that are time consuming or costly to collect, invasive, or imprecise. So let's look at the little score. The variables shown in the table are those that they evaluated in univariate analysis as potential variables for their predictive score. What do you see when you look down the list? Well, first is that there's a lot of numeric, so precise data in terms of the variables that are chosen. You'll notice that both the discriminant function and the components of the discriminant function are included. So they're evaluating whether a combo versus a single test is more predictive. Same is true for Child's Pew score. They did look at presence and absence of ascites and encephalopathy, but they note in their methodology that they eliminated these from consideration even though they're statistically associated with the end point of interest because they felt they were too subjective. So coming back to the importance of precision around the estimate around the variables. The SALT score which is looking at the endpoint of sustained alcohol use has variables of course that are more focused on psychosocial but just to point out that there can be precise measures here as well. Of course this is dependent on having them measured in a way that's consistent so it was dependent on high quality addiction medicine and social work assessments. You'll see it here too though that there are examples of variables that they viewed to be too subjective. Examples of those would be insight into diagnosis and complete acceptance of diagnosis. They were statistically significant in univariate analysis but excluded. So again you take these variables you evaluate them in univariate methods. So what what method do you use? Well the selection of the model type is somewhat dependent on the outcome you are looking at. So logistic regression is often the the test model of type because the outcome is binary. So survival, not survival, recurrent AH, not recurrent AH for example. If there's an element of time and censoring that's needed for incomplete follow-up then Cox regression is preferred. And then lasso regression is perhaps newer to some of you but this is where you're you're using either a linear or binary outcome. So it can be in this case sustained alcohol use, yes or no, but where there's a desire for shrinkage in variable selection meaning you want a parsimonious model. And this is ideal for multivariate analysis where the outcome is rare and it's hypothesized to be predicted by a weak number of predictors or it's also very good for models where there's a high level of multicollinearity that you're expecting. So coming back to again the LIL and the SALT score development. The LIL score used a logistic regression. The SALT score used lasso regression. In part you can see the reasons why when you look at the development cohorts and the frequency of the events. So you'll see in the LIL score it's a very nice size cohort 295 patients with a 35% mortality so a high number of events. The SALT score in contrast is a very decent size development cohort 129 patients but in terms of sustained alcohol use only 10% of the patients or 13 had that endpoint. So you can see the number of events is relatively modest so lasso regression quite well suited for that analysis. Okay step four is actually model selection. So you've identified the model approach you're going to use. You've identified your explanatory variables and your outcome. So now you do you use traditional covariate selection methods forward backward or both. Often statistical packages do this for you and they're using a criterion of optimizing p-value in order to do that. More recently there has been a move towards using alternative criteria other than p-value and and the reason for this is that the selection criteria based on p-values can be more unstable they can be sensitive to small changes in the data as well as changes in the number of observations which is less an issue with the alternative approaches. And what are the alternative approaches? They use a criterion called a Kiki or Bayes AIC or BIC and what the criterion here is attempting to do is to continue to find the overall best fitting model with an optimal number of explanatory variables and so it's a approach that discourages overfitting of the model and the optimal model is based on choosing a subset that achieves the lowest AIC. So that's the criterion ultimately that's used. Okay so how does this play out for the LIL and the SALT scores? So here's the model building for LIL. It was a logistic regression. They used forward selection, they detailed, didn't give much beyond that but they ended up with a multivariate model that includes both a baseline and a dynamic variable, the evolution of Billy Rubin at day 7. So they did their forward selection ultimately ended up with five variables as shown here. The SALT score using LASSO regression also ultimately identified four variables. Again remember this is a an approach that yields fewer variables generally through a mechanism of regularization or shrinkage. And so in this with LASSO regression four variables you can see strong odds ratios in the multivariate analysis all of them with the significant p-values. So you have the variables they're all significant in your multivariate model and you've gone through your forward, backward or both selection criteria to get to the final or you've used your AIC or BIC criteria. Now you have your final model. How do you display that into a risk score? Well you can either use a complex or a simplified approach. The complex approach basically derives the score directly from the model equations. So the multi model coefficients weight is used in the actual risk score. And then alternative is a simplified method in which you assign a point to an integer point to each risk factor where you scale the points based on their relative contribution. And again this is based on their coefficients. So just to show you this by way of example with a little score they use the complex approach. So in the paper they provide you with the risk equation and ultimately it links you with a an online link that will calculate this automatically for you. So it obviously looks rather complex but because it's with a link these are all actually very easy to use now because most people would provide sort of an app approach to generating their score. At the salt score used the simplified or point system approach. So you can see here the coefficients in the final model and here are the assigned points. And you can see that the coefficient which was the weakest predictor was assigned the lowest point and the variable that had the strongest association was given the highest points. And then you know based on relative contributions of those in between they were assigned in integers and then ultimately that point system was evaluated to ensure that it was robust. Additional things that you need to do once you've come up with your final model and your final score is to evaluate its performance. And there's different ways to do this. I'm going to highlight the two most common ones are discrimination and calibration. Discrimination being the ability of the model to discriminate who will develop the event and that's with an AUROC, see statistic. Calibration is the measurement of agreement between observed and predicted events and there's a statistic to evaluate that formally. For both LIL and SALT scores they use the AUROC method to speak to the performance of the models and you can see that the LIL score the AUROC was 0.89 and for SALT score 0.76. Another way to look at how this score performs is to actually demonstrate diagnostic features. Sensitivity, specificity, positive and negative predictive value and that's shown here for the SALT score. They show the actual score here and then the performance characteristics and you can see that where this score performs very well is it has a very high negative predictive value but has very poor positive predictive value. So that's very impactful in terms of how this score is used because what it tells us is that we're using this as a way to identify those that are very unlikely to drink after transplant in a sustained way but this is not something that should be used to exclude patients from consideration of transplants as it has a very low positive predictive value for alcohol use after transplant. The LIL score similarly moves one step further on how the score performs, remembering that their goal is to try to identify those with a high likelihood of not surviving despite corticosteroid therapy. So then they went one step further to identify a threshold or a cutoff that provided them with a maximum performance characteristics. The specificity of 81% and sensitivity, sorry, sensitivity of 81%, specificity of 76% and that cutoff was 0.45 here and you can see that when they look at survival that it in fact is quite discriminatory. And then the final step is you've got your score, you've looked at its performance characteristics and now you want to see if it performs well when you validate it outside of your development cohort and that was done with the LIL score and indeed external validation showed also that the AUROC was very robust and in fact it was also better than other predictive scores. So this is a very nice way to show what this score adds to the existing literature. What about the SALT score? Well they did not perform external validation because they couldn't find a validation cohort with a large number of patients with alcohol-associated hepatitis have undergone transplant. So they did something else, they did internal validation and indeed this is something to consider when there isn't a readily available external outside cohort to validate the initial original score is to do some and there are various methods of doing internal validation to show that the score is valid at least to that level. So to conclude the takeaways here, alcohol-associated hepatitis, I highlighted this is an area where there's lots of useful prognostic models and scores, still room for improvement I would say as well. The key clinically relevant endpoints are survival and alcohol relapse. Those are the two models I showed you. Other ones that are emerging include recurrent disease, recurrent development of cirrhosis and other liver related complications. And what I showed in sort of a stepwise fashion is the importance of choosing explanatory variables carefully. Choose your regression model to maximize the success of your model building. Scores ultimately can be of the complex or simple simple varieties. Model performance should be assessed in terms of how well it discriminates, how well it calibrates and ultimately how it's going to be used and that external validation is ideal. If not possible include internal validation. Thank you very much. Hello and thanks for joining us today. I'm going to be talking about qualitative and a little bit of mixed method research methodologies to understanding barriers and sobriety in ALD. My name is Jessica and I have nothing to disclose. I am an assistant professor in transplant hepatology at the University of Michigan and my clinical and research interests are in ALD where I run a multidisciplinary ALD clinic and I also do behavioral and outcomes research in alcohol cessation for ALD patients with funding from. So to begin we'll start with looking at what is qualitative research and then look at some of the key issues to consider as you apply this technique to studying sobriety barriers in ALD. We'll talk a little bit about how to choose which approach to use and then I'll also introduce a little bit about mixed methods research where we mix qualitative and quantitative because that really is a powerful way to get at more about and more understanding about the barriers to sobriety in these patients. So qualitative, so excuse me, talking about barriers to sobriety in ALD, certainly this is a major challenge for us here in the U.S. and elsewhere. There barriers tend to fall kind of in three big areas and this comes from broader studies of just substance use barrier treatment barriers in general but also from some direct qualitative work that we did in our group here with ALD patients looking at barriers to access to care to substance use care. The classic lack of insurance coverage is certainly in play and definitely for those who have Medicaid that can be a real problem. There are definitely logistical issues with transportation, lack of time off, not enough providers or maybe people living in rural areas who lack access to lack a provider in their area but far and away the biggest barriers and the ones that that really mandate using a qualitative or mixed methods approach to evaluate are these attitudinal barriers. You know most studies that look at barriers for alcohol use disordered patients find that the vast majority, 75-80 percent of the barriers that they find really reside in attitudes of the patients themselves. Largely feeling like they don't need treatment, big concerns about stigma, feeling stigmatized going to treatment and also concerns about privacy as well as social anxiety. So substance use treatment rates are low in general for all substance use disordered patients but they're also low for our patients with advanced ALD which is somewhat surprising given that an accumulation of negative physical consequences as a result of alcohol often propels people into treatment but for the broader population with ALD that pays. So when we're thinking about these different barriers to sobriety and looking at particularly evaluating those powerful attitudinal barriers, qualitative research really comes across as a major tool in our toolbox for this and what it is is a type of inquiry that looks at applying a theoretical or interpretive framework to certain research problems sort of a priori that actively address the meaning that different individuals or groups are ascribing to a social or human problem. So it's not really taking you know a 64,000 foot view, it's really going right down right next to the patient, talking to patients, groups of patients and using an emerging approach to that inquiry. So we'll see that when you when you do qualitative research your protocols sometimes change over time. You know who you're talking to, how many people you're talking to, the nature of the questions that you're asking may actually change and shift as the study goes on as you delve in and find more and that's termed an emerging approach. We tend to collect data in a more naturalistic setting and we use both inductive and deductive pattern reasoning to analyze and themes. When we go to mixed methods research then we essentially combine that qualitative data analysis and collection with quantitative, what many of us are much more familiar with, and we mix that data and we'll talk a little bit very briefly about how we do that, though we're going to focus mostly on qualitative. Mixed methods also has a theoretical if you're purely doing quantitative, but it really all starts with that and really where qualitative and mixed methods are the most powerful is in this realm of clinical and health services research, where we're starting to look at not just what works to cure disease or to treat disease, but why some people do or don't get that benefit, why are some people able to stop drinking and others are not, and that's really in this realm of on the research chain kind of here more on the right in the clinical. These are powerful techniques. So just a few examples of qualitative research questions applicable to sobriety, so you could examine why ALD patients do or do not maintain alcohol abstinence, maybe using focus groups or interviews. You could look at factors that influence a provider's decision to prescribe or not prescribe alcohol relapse medications in ALD care, looking at it from the provider end. What are we giving to people to help them with sobriety? Determining if stigma is playing a role in how we evaluate patients and transplant committees can certainly have an effect on outcomes for patients, and then looking at ALD patients' views on alcohol abstinence and treatment using qualitative interviews would be very powerful. These are just a few examples, certainly in your own practice, and as you're going, you will find more that will really guide what kind of a qualitative approach is going to best fit your research needs. So the work of John Creswell, who's really a giant in the field of qualitative and mixed methods, in this book, Qualitative Inquiry and Research Design, puts out roughly five different approaches to qualitative research. There's certainly more than this, but these are often the most commonly used and the most commonly seen in the literature. And just by way of disclosure, my own personal experience is really largely with the grounded theory approach, and this is an approach that really matches well with a lot of the health sciences questions that I use in my research and many of the health sciences questions that qualitative is useful for, in that we're looking to use our research to develop a theory about a specific behavior or problem that is actually grounded in the views of the participants. So we're not necessarily bringing with us lots of other behavioral theories. Those may inform our interview guide, for example, as we explore what factors are important maybe to sobriety or abstinence, but we ground our theory then and we come up with our theory in what our subjects are saying and what we find with that. Narrative research is really about telling the stories and telling about the life and the stories of lived experiences of usually one or two people, so we're not talking about large groups of people. Narrative research is very intensive, very time intensive, and you really wind up getting quite close to the individual and delving quite deeply into their lived experience. Phenomenological research looks at specific phenomenon, a lived phenomenon, so this could be anything. You know, this could be the experience of homelessness in a patient with ALD and how that affects their sobriety, the experience of stigma. You want to understand the essence of that experience through your subject's eyes, through the eyes of those who are living it, and typically this involves maybe a few more people, you know, 5, 10, maybe 15, but again we're not talking about big groups here. Ethnographic research looks more at culture sharing groups, so looking at shared patterns of a culture or a group, so if one were to look at, for example, alcohol use in a specific race or ethnic group. Ethnographic research might be a good way to answer some of the questions around issues of sobriety barriers for those individuals. And then finally, case study research, as it sounds, is really about developing an in-depth description of a case or multiple cases and really coming to an understanding of that case around as a form of your research. So you can use any of those, but you're really going to want to start first with that research question. Why do you want to study it? And then matching up your research approach or qualitative approach to that. So then once you've decided on your approach, thinking about what exactly do we analyze, what are the data sets in qualitative? Unlike quantitative, which is as it sounds, quantitative, it is often numeric. It's instruments, demographics, experimental data, surveys, administrative data. Qualitative data focuses on interviews, focus groups, focus group transcripts, written records, and this includes, if you're doing the interview, your own notes and your own field notes of what you're seeing during the interview. What are the facial expressions? What is the posture? What, you know, what's the feel in the room? What are you seeing? That also becomes part of qualitative analysis. Videos and images can also be a part of it, and increasingly we're seeing more and more work coming out using social media. So Twitter feeds, Snapchat, TikTok, a lot of these social media, Facebook posts, etc., that really contain a lot of rich data, but often are very ephemeral. They come and they go and they take a long time. Qualitative research sampling is different. So if you've read much qualitative, you know that the sample sizes are often very small compared to what we're used to seeing in large randomized trials where we have hundreds if not thousands of people. So sample sizes are often smaller and they're not fixed. So you may have a study that has 10, you may have 15, you may have 20 or 30, but as you are doing your analysis, if you find that thematic saturation is happening, you're really not encountering new information as you read more interviews. Oftentimes interviewing will be stopped, and so you may not get to, you know, maybe you said you were going to do 30 interviews, but you may only do, you know, 22 if you hit thematic saturation. You can do many different types of sampling as this slide talks about. It's much more than just the random sampling that we see in a randomized control trials, but you often want to be thinking about who do you want to, who are you, what population are you really looking for so that you can get representativeness. So if you are doing a study looking at sobriety barriers from a gendered standpoint, you're gonna really want to focus on making sure that you sample to achieve equal representation of both gender, of male and female gender. You might want to look at special cases or outliers, you know, finding outliers and selecting for those or doing what's called snowball sampling, where one sample, one subject gives ideas for other people they can sample on, for perhaps family members, etc. When we start with data exploration, you really just start by reading it. So it's very much like if you remember from university doing literature courses, you're really reading the text, you're getting a feel for it, you're remembering that, as in many cases, how you're doing your theory. If you're doing grounded theory, remember you're going to be approaching that analysis where you're going to build your theory from the data, as opposed to necessarily bringing it in with you from another a priori theory or conceptual model. After you get the feel for it, you start memoing, you start making notes, kind of reading, reflecting on it, moving, you know, between interviews, comparing things, comparing notes, and this is that constant comparison where we constantly are going back to other team members looking at previous, you know, what we wrote earlier in the interview, and this kind of emerging data begins to take shape into categories, codes, and themes, and that is what we build our code book out of. So typically in my work, my team and I will take maybe three or four transcripts. We will all code them, and then we will come back together and talk about it iteratively and decide on which codes really were meaningful, which codes can be collapsed into other codes. Finish by making a code book. This is an example of a code book from one of my qualitative studies, part of which was looking at ALD patients and barriers to change, and we looked at, as you can see on the left, motivation to change was the parent code, and underneath it subcodes included medical symptoms, external factors versus internal factors, and then representative. Remember that when you are coding, it is an iterative process and often happens in a group or a team, so you code a few, you go come back as a group, you talk about it, maybe we change things, maybe we find codes aren't working, or we need to collapse codes down, and then you do that over and over, triangulating your data potentially with your own observations and field notes with the medical record to ensure validity, until you hit thematic saturation, where you're just not seeing new codes or new meaning. You're getting code saturation, but also you're saturating the meaning. You're not, and that's when we typically know that we are done. There are a lot of really interesting visual representations of qualitative data when you're thinking about presenting it. You can certainly do standard tables, histograms, etc., but many of the qualitative software companies now have software that can do some pretty interesting data representation. So this was out using MaxQDA, looking at barriers to treatment amongst four groups, males and females, with compensated or decompensated ALD, and you can see here that the amount of time, the number of codes spent on these different topics is visually represented by the size of the square, and so you could get a quick visual representation of how important a topic might have been to different members of different groups. Mixed methods, just very quickly, is about, again, mixing it. So it's about mixing qualitative with quantitative, which you may want to do if you want to use them to strengthen each other's weaknesses, if you feel it would be more comprehensive than either one alone. Maybe you have some survey findings that you want to use qualitative to flesh out and explain, or you have some qualitative data and you want to see if it expands. There are different decisions that you may want to make when deciding how to do a mixed method study, and it will depend on how interactive you want qualitative and quantitative arms to be, the priority you want to give them, do you want them to occur at the same time, be sequential, or be multi-phase. Remember to match a design to your research purpose. So convergent designs happening at the same time may be appropriate if you don't have a lot of time or money and need to do both at once. Explanatory versus exploratory, you want to evaluate those to see which one fits with your particular research question. Maybe you have a survey and then you want to expand more and dig deeper into it, you might go into qualitative after that. And then embedding RCTs can be interesting. So overall, qualitative and mixed methods is very challenging but very fun. I would suggest gaining experience in quantitative and qualitative separately and then trying to mix them and really understanding why you're using qualitative or why you're mixing methods. Know that there are some barriers to it, and that there are several resources for qualitative and mixed methods which I'm labeling here, particularly this website on mixed methods resources that also will show you where you can get this research published, like which publications and journals accept mixed methods research. I thank you and I'll be open for questions. Thank you for letting me participate in this workshop. I have nothing to disclose. I would like to begin with this example, which is a paper by Drs. Tapper and Parikh who used a CDC mortality registry data to evaluate the burden of ALD in the U.S. And this figure just shows that the burden of ALD has been increasing in younger age groups between ages 25 and 34. The details of this figure is not as important as the point that I'm trying to make here, and that is that one, you can get publicly available data sets like this one from CDC and many other different sources at a very low cost or no cost at all. And two, if you have the right questions to ask and have the ability to answer those questions in a scientifically valid fashion, then you can generate high-impact scientific knowledge that would be acknowledged by high-impact journals such as BMJ. As brilliant as that example may be for a secondary data analysis, I would like to emphasize what epidemiology does, and that is that it is a science of discovering causal relationships. So, useful descriptive analysis is great, but we can take a step further and see if we can prove all of these things. We can see if we can prove or get close to demonstrating causal relationships. This is what I call causality matrix. As you think about the study designs for your database, there may be several different kinds that vary in terms of its ability to show internal validity and ability to control exposure. Today, we're talking about cohort study design, and in the grand scheme of things, it is not as strong as randomized control trial, but it has a decent chance of showing strong associations, but also getting fairly close to demonstrating causality. As we pursue truth and internal validity with our data, there are some pitfalls to avoid, and they are biases and errors. Errors are simply measurement variability for the things that we want to measure, whereas biases takes us away from the true value in terms of what we can measure. An example may be portal pressure versus HPPG. We have errors in measuring HPPG, but portal pressure is not necessarily the same thing as HPPG, so we need to recognize that bias. But in terms of the administrative large database analysis, to the degree that you're not making those measurements yourself, the data is what you have, so it is important for you to think about what biases and what errors may be entailed in your data sets. If you focus on biases, you may have heard of many different kinds of biases, but ultimately, there are only two types of biases. One is selection bias, and the other is misclassification bias. Selection bias gets at whether you have the right population to study, whereas misclassification talks about making the right measurement for the exposure variable and the outcome variable that you're pursuing in your study. So with those concepts, let's think about designing a cohort study to look at the impact of ALD. You want to construct a cohort, which is shown here as a pink arrow, and you want to identify predictor or exposure variable, and over time, the cohort members will develop intermediate outcomes, and ultimately, the outcome variable that you're looking for. So at the end of the study, you want to demonstrate the causes and effects as you have those data. For an ALD question, let's say my question is, does obesity increase the risk of death from ALD? So in order for you to do the study, as an example, you would need an ALD cohort, and in order to do that, you will need to typically use a diagnostic code to define who your cohort is. At the beginning, who your cohort is. At the beginning of the study, you want to define obesity as a potential cause, and ultimately, you want to measure death from ALD, and the intermediate outcome in between might be cirrhosis or HCC or other comorbid factors. So, first step, how do you define an ALD cohort? These are ICD-10 codes for ALD. Under the heading of K70, there are several categories, including fatty liver, alcoholic hepatitis, fibrosis, cirrhosis, and hepatic failure. I'd like to make two points. One is that, depending on the study question that you're asking, you need to carefully select which categories of patients that you need to include in your cohort, and more importantly, perhaps, whether the ICD-9 codes are appropriate to select the patients that you need for your question. The second point may be ascertainment bias. For example, if you're studying alcoholic fatty liver, then you need to think about whether the database you have adequately captures all alcoholic fatty liver patients, or there's some sort of ascertainment bias included, why some patients with fatty liver disease may be captured in the same cohort. Some patients with fatty liver disease may be captured in the database, while others are not. There are other challenges in studying ALD. The first may be how to quantify or characterize alcohol use patterns, and using administrative data, it will be difficult to discern binge drinking or identify the type or amount of alcohol in an individual, unless the data is specifically designed to answer those questions. Second, ALD is associated with other socioeconomic factors, income, education, deprivation, and other comorbid mental illnesses, and these factors may be confounders in the analysis, sometimes not apparent that you can adjust for. Third, we just discussed early stages of fibrosis or asymptomatic stage of ALD. It would be difficult to identify, and the concern for ascertainment bias would be there. What about the other data variables? For obesity, you ideally would like to have BMI data, but very few administrative data will have BMI, so you may have to resort to diagnostic codes for obesity, and you can imagine how inaccurate or accurate that may be in a patient who has simple obesity without comorbid disorders associated with it. For cirrhosis and HCC, you would be looking for a diagnosis of cirrhosis and HCC, and then the ultimate outcome of death from ALD, you'll be looking for a cause of death, and for these situations, again, you will be using diagnostic codes, and you will need to think about how representative and how sensitive the diagnostic codes will be for the purpose of the study. I hope I demonstrated some of the pros and cons of database research. The pros, data are already available, obviously, and many of them are affordable. Second, many databases are large enough that it gives you statistical power to demonstrate significant findings. Many databases tend to be generalizable to the population that you're studying. Databases are useful to generate preliminary data, and ultimately, it is a hypothesis-generating exercise. The cons, database research is limited to available variables. There may be many missing data that will make it difficult for you to have adequate power. Accuracy or validity of the measurements contained in the database is something to be careful about. In order for you to execute good database research, you will need to have statistical expertise. And finally, it is only hypothesis-generating most of the time, not hypothesis-proving. These are some of the examples of publicly available databases. This is a paper by Philip Okafor, published a few years ago. Many of you will be familiar with OPTN and NHINS data, but there are BRFSS data, which may contain behavioral risk factors associated with ALD. There are other survey data, as you can see here. More data sets here, MEPS is an expenditure panel survey, SEER program for cancer data, NSQIP is surgery quality program, and HCUP, Healthcare Cost and Utilization Project, data is often used for hospital discharge analysis. There are many other data sets that one can consider using, and some of the often-used databases are Claims Database, as you can see here, OPTN, Truven, Cerner, or IQVIA database have been used. Health System Database are great to use, Kaiser, VA, particularly VA has Audit C data included there annually. Mortality data from CDC, as you saw in the example shown earlier. I'd like to end this talk with three tips, and the first is about statistics. There are two aspects to this. The first is just the knowledge of what tests to use, what's appropriate, and so forth, and the second is the programming aspect. With regard to that, I think it's important for people in the early phase of their career to have a good command of a statistical package, such as SAS or R, and I prefer not to use a spreadsheet software for data management or analysis because you can often make a mistake and cannot reverse. It is important to have a good nurturing relationship with a biostatistician partner or mentor, and not make a mistake of thinking of them as a service provider. You need to have respect, commitment, and time to make the relationship correctly that will be conducive to productive career with regard to data analysis. The second tip is about p-values. In determining statistical significance, there are three factors. The first is the true difference that we want to demonstrate, defect size, but in addition to that, there's sample size and data variability. Larger sample size and smaller data variability will drive smaller p-values. However, statistical significance is not the same as clinical or practical significance, and this is where your content expertise come in to interpret the data correctly. If you do multiple comparisons, you need to adjust your p-value cutoff, and importantly, just because you have small p-values, it doesn't mean that you have the truth. You cannot overcome a poor study design just because you have small p-values. Lastly, it is important to know your data. This is an example four-line code that will drive MELD score, but only if you have the right data set to drive it. So, it is simple to write the code. It will take a few seconds to run the code by the computer, but it is super important for you to know the data inside and out so you can have confidence about the result of the analysis. In summary, large observational data sets can provide valuable insights at a population level. Acquiring data is much easier than collecting them on your own. These analyses are great for hypothesis generation, but you need to be aware of its pros and cons. ALD research can be particularly challenging to use these data because alcohol use patterns are difficult to characterize retrospectively. We talked about the potential ascertainment bias of asymptomatic ALD cases, as well as selection bias related to psychosocial comorbidity. Secondary analysis definitely has a role in academic career. It is an easy way to enhance productivity. It may potentially be highly cited, which is attractive to journals. It is a great way to provide background data in grant applications. It can be used to justify hypotheses in the significance section. So, my final advice is proceed, but proceed with caution. Thank you very much. So, this concludes our 2020 Clinical Research Workshop, Research Methodologies Using Alcohol-Associated Liver Disease as a Model. With the increase in prevalence worldwide of alcohol-associated liver disease, particularly in the setting of the COVID-19 pandemic, it will be essential to identify novel biomarkers and validate these biomarkers to predict those who are at highest risk for progressive liver disease, as well as increased morbidity and mortality. Translating these findings from bench to bedside should help us identify biomarkers with the highest chance of predicting liver injury. We also have a greater understanding now of the mechanisms of injury of alcohol-associated liver disease. As we continue to identify these pathways, we can hopefully also identify existing and additional targets to engage to allow the greatest chance to ameliorate liver injury and also reduce subsequent hepatic fibrosis in those with alcohol-associated liver disease. With the identification of biomarkers, the identification of biomarkers, the identification of mechanisms of injury, and potential therapeutic targets, it's going to be important that we create meaningful endpoints to assess the efficacy of all of these interventions in those with alcoholic hepatitis and alcohol-associated liver disease. This is going to be extremely important as we want to ensure that our endpoints are designed in such a manner that they do not miss potential efficacy signals for an entire cohort or in subgroups who are treated for alcohol-associated liver disease. In addition to identifying those who are most at risk for progressive alcohol-associated liver disease and treating these individuals, we need to provide resources to help patients achieve and maintain sobriety. Understanding these barriers faced by patients, their families, and the physicians who care for them will be important in our efforts to reduce the morbidity and the risk of liver in our efforts to reduce the morbidity and mortality associated with this disorder. Finally, we need to develop cohorts to better assess the challenges we face with the changing epidemiology of alcohol-associated liver disease, as well as to track how well our interventions are working. In liver disease, we have used large administrative databases, and they have been essential in helping us to direct our resources to those most likely to benefit from our interventions. And we will continue to move forward with these large databases and gauge our efforts in this disorder. Again, thanks very much to the faculty who presented here, including Drs. Liang-Poonsakol, Zabo, Turow, Mellinger, and Kim. And please enjoy the rest of your meeting. This concludes this clinical research workshop.

Liver Meeting Clinical Research Workshop

Alcohol-Associated Liver Disease

ALD

Biomarkers

Inflammation

Fibrosis

Alcoholic Hepatitis

Non-invasive Markers

Research Methodologies

Translational Research

Sobriety Barriers

Prognosis

Cohort Study Designs