Computer generated transcript

Warning!
The following transcript was generated automatically from the content and has not been checked or corrected manually.

All right. So there we go. Is everyone here, come to your eye can't see. So you just let me know and also for the questions, I'm not used to them at all. So I'll leave that to you if there are any burning questions. Yeah, we've got some people here. Um People can just join whenever so you don't have to let them in or anything. So just fine, I'll, I'll get started. Great. Let's do it. Um Well, for those who are here, um Thank you for joining on a Tuesday evening. I won't keep you for too long. Um Just as an introduction. My name is Simon. Be, I'm one of the academic foundation doctors. I'm currently working in Royal Preston Hospital. Um and I was a Manchester grad. So if you're a man, just medical student, I have been in your place. Um And today, what I'd just like to do is talk to you guys about, if you are interested in the field of academics and research, it'd be nice for you to have an idea of basic statistics and how to apply them clinically and practically in what we call critical appraisal. So not every research uh project is done equally, not every study is done equally. Um And it's important to be able to identify um or what we call validate um a study. So if you remember, we've got different tiers of research. Um your level one evidence being your meta analysis and systematic reviews and your um level four, level five evidence being your case report and editorials. Um And so we'll touch upon those in a sec. Um There is a lot of detail, I'm happy for the slides to be sent out to everyone after because I think there's some important information on there that can hopefully help you out. Um But as as always, if you've got any questions, let just let me know, type it in the chart. Um and we'll get through it. All right. In essence, we'll talk mostly about things you need to know um from a basic statistics and clinical interpretation point of view. And then the last slide will talk to you about some not so basic analyses that you can perform. All right. Um In essence, you can break up every clinical problem as a set of variables. So you have an outcome variable ie the variable that you're trying to answer and variables that are independent or what we call your predictive variables ie variables that you are analyzing in order to impact in order to evaluate the impact that they have on your outcome. And that goes for any um problem you're trying to solve. And so as researchers, one of the first thing that I would recommend is you have a look at one of the principal components that build up your question and what do you need in order to answer that? So, always work from hypothesis to um testing rather than the other way. So instead of just collecting variables and see what you can do though, sometimes that is important in what we call unsupervised learning or exploratory data analysis is often better at the start of your career to look at hypotheses and then try and test them out. Um So that's your rigorous hy hypothesis um testing. Um Anyway, going back to variables, this is very important and very basic. You've got two variable types, you got your categorical and your continuous variables, you can subdivide them your continuous variables, uh your numerical variables. So any number. So BMI height age, those are all continuous variables. There is a second type of continuous variable called a discrete variable. So GCS, for example, ranges from three, all the way to 15. It's a discrete variable because there's nothing called G CS 14.5 or 4.7. Um but it is a continuous variable. However, some might also say that that's wrong because it might fall under a categorical variable. And why is that? So you've got three types of categorical variable ordinal, nominal or binary or dichotomous ordinal is in the name, they have an order to it. So TNM staging, so as you move on from T three T four N one M one or M zero, you've got your staging, your stage 1234 and the higher your stage, the higher your order. So you're ordered. Same for your uh visual analog score for pain zero is no pain. 10 is a lot of pain. So there's an order to it, I ea person with a score of 10 is higher ranked than a person with a score of zero. And so you could say the same for G CS, you could say, oh a, a patient with G CS 15 is better in a patient with GC S3 and you're right. Um And so there's sometimes it can get murky in terms of what's a discrete continuous or what's a categorical ordinal, but that's outside the scope of this lecture. And most of the time it's gonna be a categorical ordinal variable, a nominal variable. You hardly ever see it in clinical practice. Um because most of the things you're trying to do will have a rank to it. Um But an example that's just names of employees and then the last one is a binary or dichotomous variable um where it's either a yes or a no or a one or a zero. So those are just questions you're asking. So, you know, what is the sex of the patient is? The patient smoking does the patient consume alcohol. Does the patient have a headache? Any symptoms? Those are ones and zeros you either have it or you don't? There's no middle ground. Um, again, that obviously depends on the type of question you're asking. Um IV smoking status. So if your question is, does the patient smoke, um ie if it's one packet a day or 20 packets a day, they all fall under. Yes. But a problem with doing that, obviously, as you can see easier than making a very heterogenous population of people who smoke 20 packets a day and who smokes one pack a day under the same umbrella. But obviously the the repercussions of that on their physical and clinical health is different. And so that is another thing that comes with experience is knowing when to use a variable as binary and when to use a variable as or no. All right. Um I've touched on this. So basically any predicted variable ie the variables you're trying to analyze are your independent variables ie they are independent of each other and the outcome variable of the variable that you wanna see if they have an impact on it is your dependent variable. And just if it, if it helps you from a visual point of view, that's how it looks on a graph. The Y axis is your dependent variable axis uh and your x axis is an independent variable. So an easy example is you want to predict length of stay between male and female patients. So length of stay is your outcome variable. It goes on your y axis. Um The X axis, you have two categories, male and female. And then the height of the bar chart represents the length of stay and you can then compare it. So that's an idea um of how it works in, in practice. All right. Um I think this is one of the most important slides in the presentation, which is P values, you to understand what AP value is because that is one thing that unfortunately, this day and age is what is important um for you to get published. And what I mean by that is if your entire research project has no significant P values, you are statistically likely to not get it published. However, there's a lot of papers in the literature that what we call are negative papers ie there's something controversial in the literature and you are trying to recreate it, which is one of the core foundations of science. Um And you can't, in that case, that's OK. But if it's a foundational paper and you're trying to be something or demonstrate something and you don't have a significant P value, um then it's harder to get published. But in essence, what does it mean? We keep saying P value P value, what does it mean? It all has to go back to your hypothesis testing. So you have something called a null hypothesis and something called your alternate hypothesis. So my hypothesis say is on a pain scale, female patients have a higher tolerance to pain than male patients. That is my hypothesis. And now I wanna test it. How am I gonna test it? I'm gonna look at all of the visual analog pain scores between our male patients and our female patients. And my hypothesis is female patients will rank lower because they have a higher tolerance than male patients. My null hypothesis is that there will be no difference between the two patient groups. So, so, so basically, there is a no relationship or the relationship is null. So that's where the name comes from null hypothesis. And the alternate hypothesis is the hypothesis that you've generated before you've started the paper or the project. So that all the P values is the probability of obtaining the test results under the assumption, the null hypothesis is correct. Basically, all it means is never in real life. Can you have a 100% certainty? Never. And so you need to generate a risk you're, you're willing to take. And the risk in all of medicine research is 5% ie in every paper that you read the threshold for their conclusion to be correct can be wrong. 5% of the time, that's what it means. It means that the results that you see in a paper um can be generated by pure chance. 5% of the time and if it is more than that 5% threshold, then sorry, you are luck. But if it's less than the 5% threshold, then good. Now, this is where the problem arises is why 5%. And if 5% what's the difference between 4.9% and 5.1%? And that's called the transposed conditional fallacy. And that's unfortunately where medicine is, is rather than providing a range of accepted P values, they've provided us with a cut off. And you could argue that between 49% and 51% sorry, 4.9% and 5.1% probability. There isn't a whole lot of difference. However, any paper with a 4.9% probability or AP value of 0.049 is likely to get published than those with the results of 0.051. Because when a researcher sees 0.049 their conclusion is that there is a statistically significant relationship and they're saying that their hypothesis is correct. But if it's 0.051 you'll often see papers do something called trending towards significance or nonsignificant. So when you hear that word statistically significant, this is what it means. So it's important to know what it means. It simply means the probability that the null hypothesis is correct and that your results are wrong basically. And that for medicine is 5% or AP value of 0.05. All right. Um Moving more into our evidence based medicine um analytics, um what we call measures of central tendency. Um your mean median mode. We've done this in middle school, high school and then the start of university as well. Um But I just wanna point your attention to the um F on the top, right. A perfect normal distribution is in the middle where your mean median and mode are all the same, your mean is your average or your arithmetic mean, you have different types of mean like your harmonic mean and stuff. But I'm only talking about the arithmetic mean. Your median is what value represents 50% of all values ie the one in the middle and your mode represents what's the most common occurring value. But in a normal distribution, it's all the same over here what we call a negatively skewed um normal distribution. But I mean, you can't call it normal distribution by, by a negatively skewed distribution. Your mode is gonna be higher, followed by a median, followed by a mean and the same in terms of a positively skewed, your mean is gonna be higher than your median, which is gonna be higher than your mode. And these are just things to think about when you're reading a paper, you will never see mode ever. So you can basically forget about it. So now you're just dealing with mean and medium. And if you see that they're quite close together. Then the assumption you can make is what they're reporting follows to a certain extrapolation, a normal distribution. But if you, if you, if you have a huge variety between or what we call variants between your mean and your median, then it's a skew deviation. And that's important because that will determine this next analytical steps that you take or the next analytical steps that you can analyze whether or not the researchers of the paper you're analyzing have taken. So if it's a normal distribution, you will do what we call parametric testing. And if it's a non normal distribution, then you will do nonparametric testing. And this is where a lot of papers fail in the sense that they just apply whatever they know despite it not being the right thing. All right. Um In point number four, there's something called SD and SE MSD stands for standard deviation. SE M is the standard error of the mean. Um You will most often only see standard deviation um which is the square root of variants. Um You then have, you don't have to worry about square root of mean. Um All right. Fifth is I QR or your interquartile range. It's the range of values between 25% or the 25th percentile and the 75th percentile and then smack bang in the middle is your medium. And so if you look at the graph at the bottom and this is your 25th percentile and this is your 75th percentile hypothetically though it isn't. But if you were to take it that way, the red would be your 50th percentile. Um So that's how you can think about it. Is it smack dab in the middle and it represents the range of values between the 25th and the 75th percentile. Um The last part is your confidence interval. You will see this again and again and again, this is extremely important to understand all the confidence interval tells you is that if you repeated the paper again and again and again and again, using different cohorts and different population groups, 95% of time, the observed value will fall within this range or this confidence interval. Why? 95? Because your P value is five percent and so 95 is 100 minus five. You can, you can move it straight up to 99% confidence in total if you'd like to be that thorough. Um In fact, it, you can apply it for the conference interval for your me, the conference interval, for your media and the conference interval for basically anything. Um All right. All right. So the next thing in your critical analysis or your appraisal or your emergence into evidence based medicine is what do you do next? You've identified your variables, categorical continuous. You've then performed measures of central tendency in order to see whether or not you have a normal or a non normal distribution. Why is that important? Because it's gonna help you guide whether you need parametric or nonparametric testing. This is where you move on to step three, which is correlation analysis. So in Falon, it's what we call a Pearson correlation or A R squared, you will see this everywhere. And it's probably it may be something that you've done previously simply it is to determine association between two variables or it's what we call a bivariate or a bi variable analysis. So it's only comparing two variables. So if you have 14 variables, you'll have to do it multiple times. You can only compare two variables at one specific time. So in figure one, all I am doing is I'm comparing two scoring systems. Um you can't appreciate in this figure and it's something I have to change. But basically, you're comparing the X and the Y axis, they are continuous variables. And that is key. You should not use categorical variables for A PS and correlation analysis. It is only comparing continuous variables. And that is the first assumption. This is very important that people forget is each test that you do has certain assumptions that need to be met before that test can be the right test for you to do. So for example, for Pearson, both variables have to be continuous and both variables have to be linearly correlated. What do I mean by that? If you were to make a scatter plot between the two continuous variables. So say age and BMI if it follows a linear trend all well and good Pierson correlation is the test of choice. However, if you plot the scatter plot between the two continuous variables and you see that it's parabolic or U shaped and you try to do a PSN correlation, you won't get a good result at all. But if you didn't know that and your RWA value is zero, your conclusion then is, oh there is no correlation between the two variables. So that's not right because there is, there's simply a parabolic correlation between the two variables, not a linear correlation. And so you need to do a different type of test, you can do line of best fit analysis and then look at your mean squared error. But the PSN correlation does not, is not able to um account for that. However, you have another test called Spearman's rank test or the Spearman's test where you can compare an ordinal continuous variable and uh ordinal categorical and a continuous variable. What's an example? You can look, look at length of stay, for example, between um different age groups. So you have age group, 0 to 18, under age group, 20 to 40 under age group 40 plus, you've got three age groups that are ordinal in nature because a 40 year old person is older than an 18 year old. Um And but then you on the other side, you've got length of stay, which is a continuous variable. And in order to trend that data or perform correlation analysis, you do a Spearman's rank. Hopefully that makes sense. If you have two continuous Du Pearson, if you have a ordinal categorical and a continuous doumen, all right, if there's one thing you can take away is this table. So take a screenshot, take a photo, whatever you have to do, take this table. Um I'm very proud of this table, but in essence, it is all you have to know when you are performing primary data analysis. This is the table that's gonna tell you everything you need to know. And what do I mean by that? If you look at the points 1 to 3, we've spoken about it already. You determine the normality ie whether or not it's a normal distribution. And then you see whether or not you're gonna use parametric or nonparametric tests. How do you determine normality is the next question? You could just calculate the mean median mode, see if it's the same and if it is you have a normal distribution, however, it can be a bit time consuming. And so you could do quantitative tests, we'll give you one number and if that number is closer to one or closer to zero, you have a normal distribution or you don't. And those are your Shapiro Wilk and the KS test. And in the brackets, I've said N less than 50 n more than 50 simply represents that the Shapiro Wilk test is better for small groups. Um IE is your sample size. So if you've got less than 50 participants or less than 50 patients, you can use the WK in order to determine normality. But if you got a really, really, really big data set, um or anything more than 50 you can use the KS test. Um But anyway, let's let's look at this chart or this uh table, if you wanna compare the means of two independent groups, all right. So length of stay between males and females, your dependent variable is continuous, which is your length of stay. Your independent variables are categorical. In this, in this case, males and females, if it is normally distributed, ie parametric, you would do what we call an independent samples. T test. That's it. And if it is not normally distributed, you do a man Whitney U test, that's it. Move on to the second row. The mean of two paired samples, example, weight before and after diet for one group of subjects. Perfect. It's given us an example. So you're looking at the same group of patients. So say you take every medical student in year four, that's your cohort and you take them through a diet program and you measure all their weight before that program and then you measure all their weight off. It's still the same group of people, but you've you, you've done an intervention and so you simply wanna compare before and after, in that case, we do what we call a pad samples t test if it's normal or parametric or you can do a nonparametric Wilcoxon sign, right test. Now if you have three groups. So let's go back to our age group, 0 to 1820 to 4040 plus. And you wanna compare length of stay if they are normal, you do a one way Anova, Anova simply stands for analysis of variants or if it's nonparametric ie not a normal distribution, you can do your skull Wallace test. And the next one is three plus measurements of the same subject between two continuous which we've spoken about, which is your spearman's and your Pearson. The last one is for categorical variables. We spoke about PS and just comparing two continuous, we spoke about spearman, which is, which is comparing a continuous and a categorical. But what about two categorical? In that case, you do something called the chi squared test. So if you wanna compare, for example, do they have headaches? So that's your question. Headache. Yes and no. And then sex, male or female, those are two binary categorical variables. You wanna see if they have an association, you do a chi square test and you're done, it's gonna give you a value and if the P value is less than 0.05 then you have a statistically significant correlation between these two categorical variables. So that is the table of choice that's gonna tell you and basically guide you to what test you need to do. All you have to do from your perspective is determine the normality of the data. All right, we focused a lot on our independent variables and our correlation or associative analysis. But what about outcome analysis? So let's take the example of does the patient go home day two, POSTOP? That's the question. So your outcome is yes and no. The patient goes home 48 hours, the operation or not. And your predicted variables are, did they have a blood transfusion? Did they have a fall? Were they male or female? What age group did they belong to? What was their BMI, et cetera, et cetera? So, so that's the question you wanna ask. There's two ways of achieving an outcome and doing an outcome analysis. If your outcome is categorical, which is in this case, it is, you can do something called a logistic regression. All it's gonna tell you is what is the probability of this patient having these characteristics to have this outcome? So what do I mean, what is the probability that this patient is gonna be discharged within two days of a surgery? If they had this BMI at this age, they were male or female had this blood transfusion or not? That's all it is. And it produces a table like this that you can see. So instead of platelets, if you think about that said blood and instead of constant, if you think that said um male or female, then what it's gonna give you is this XB or your odds ratio? It's gonna tell you what are the odds of a patient having a blood transfusion and going home before two days POSTOP? And if the odds ratio is less than one that it's having a negative impact on your outcome and if it's more than one that it's having a positive impact on your outcome, what does it mean? So say that they said blood transfusion, if you had a blood transfusion, you are less likely to go home more than two days POSTOP. So let us in because the odds ratio is less than one, it has a negative impact on the outcome. And our outcome of choices can we predict which patients are gonna go home less than two days POSTOP? But because it's less than one, then we know it's negatively impacting that outcome, which means these patients are more likely to go home more than two days POSTOP. Um That's your logistic regression and you can, you can have different types, you have your b logistic regression, you multinomial, your ordinal, you step wise which you're not gonna get into. But we'll cover this when we come to the paper. But the paper that you know, we sent out earlier was creating a scoring system based on logistic regression. And the key to analyzing a logistic regression is looking at your X to B or your odds ratio and then looking at your significance because each variable has a significance. So in this table, for example, you can see that platelets was a statistically significant uh variable and it had a negative impact on the outcome. All right, a classification table. This is more important when you do a specific type of research that has to do with testing or diagnostics or prognostication where you're trying to calculate sensitivity, specificity accuracy. All of those types of numbers, not every research project you do has this in it, but it's very important for you to know this. Why? Because in clinical practical medicine, you see this, for example, something called ad dimer test. If you are worried that a patient has a pulmonary embolism or a deep vein thrombosis or any type of venous thromboembolism, um You wanna do a ddimer test that's wrong. What you first have to do is do a history and then do a physical exam, calculate the wells score for DVT or for PE and then do the D Dilema test. Why the D dimer test is highly specific but not sensitive? What it means is that if your D dimer is normal, then you do not have AP but if it is abnormal, there is no faith that you most definitely have a key. I ea lot of people with a high or a positive DDIMER will not have a pe and that is sensitivity and specificity. So if you look at this classification matrix and you look at sensitivity, which is this purple box, it is your true positives divided by your true positives plus false positives. Ie how many people who had a positive test result actually had the disease? That is what sensitivity is dealing with specificity is dealing with the opposite ie your true negatives divided by your true negatives plus false negatives. Ie if you had a negative test, how many people actually did not have the disease? And this is very important if you're interested in cancer work because if Amri scan misses a cancer ie has poor specificity and that's really bad because you'd much rather not miss a cancer. Um So that's something very important to know is the difference between sensitivity and specificity. Because in clinical medicine alone, you will see a lot of this. For example, when the RT PCR COVID-19 test came out, they were more sensitive and more specific um than the previous gold standard testing, which was your rapid uh panel where you just dropped it onto a, a slab. Um I forget what they were called. Um But anyway, the R TECR was compared to that and showed higher sensitivity and higher specificity in actually identifying patients with COVID-19, but it's never gonna be 100% in cli in clinical work because simply you cannot account for every single heterogenous patient group ever. It's simply not gonna happen. And so you must always account for error. Um Going back to the RT PCR for COVID-19, it was 2%. That was the error. So the sensitivity and specificity was 98%. All right. Um Coming to the towards the end of the talk now, which is more focused on the paper, which is your critical appraisal. How do you critically appraise the paper? The introduction does not matter. So you can forget about it. What matters from a evidence or quantitative point of view is the methodology of the paper. So you look at the methods, you then look at the results that the methods have generated. And though I have not said it here, the last part of this is the conclusion is that people have drawn from that. All right. So how do you go about doing this? It's easier than done. My first thing is what is the question the paper is trying to answer? All right, why I find this important is because then I can generate my own methodology. So the paper that we look at was trying to predict mortality after hip surgery. If I know that's my question, I then am now generating variables from my own knowledge and creating my own methodology pipeline that is grounded on the foundations of evidence based medicine. So I know that my outcome is binary. I know that I want to look at these, these, these predictive variables. And my next question is OK, how do I go about doing it? And so look at the predictive variables and the outcome variables, then you look at the tests that were conducted now that you have that beautiful table that I had created. You can now see for yourself how the paper should have been done, how the paper should have been conducted. And if they aren't following what you have predicted, why have the authors given a valid statistical evidence based reason for it or not? Uh The reason that's important is because you can then say, you know what, that's the wrong test. So many a times I'll read a paper and they'll just throw a test in there because it has a fancy eponymous name to fancy people. Their names are attached to it and they think that, oh this is a test that we should do. For example, there's something called the Point by serial analysis sounds fancy. Really isn't. And you should never see it in a research paper because it doesn't add a lot. Not because it's a, not, not a good test to do. It's simply because the logistic regression, all the correlation analysis you're gonna do is gonna be uh good enough. So that's the next thing. So the usual method that I follow is cohort demographics which are your independent and dependent variables. So your predicted variables and your outcome variables and making a broader outline of what needs to be done, then you have your correlation tests. So your Pearson or your spearman or your chi square, then you have determination of normality. If you have continuous variables, because those will then play into whether or not you go for parametric or nonparametric tests. And lastly, if you have a continuous outcome, you can do a linear regression. But if you have a binary outcome or a categorical outcome, you can do logistic regression. And that's it that in essence will cover 99.99% of papers is this methodology. All right. So if you have the paper open, that's good. I appreciate that people are busy. So if you haven't done, that's all right. But basically, we will go through this paper, all it is is trying to create a scoring system to predict 30 day mortality in those undergoing hip surgery. A very easy, simple question that they're trying to answer. They've looked at a bunch of variables and their outcome variable is 30 day mortality. So yes or no. Was the patient alive or not alive? That's it. And if you look at the table, the variables that they looked at were age sex, hemoglobin on admission, your mini mental score, whether or not they were living in an institution, the number of comorbidities that they had and whether or not they had some sort of underlying cancer which is good. Um They could have had 50 variables and may have only chosen this. That's all right. The next thing you need to see is what have they used these variables. As for example, age usually is a continuous variable. But if you look here deeper, they have bin it ie they only have two age groups, 66 to 85 and 86 or more. And so they have what we call reduced a continuous variable into a categorical variable. And I'll come into why that's important. Same for hemoglobin, hemoglobin is a continuous scale. You know, you can have 10.2 10.19 0.019 0.0 to whatever. But here they've binner it, it's either more than 10 or it's not the same for your mini mental state exam score. Is it either more than six or less than six? And the same for the number of comorbidities? So if you have heart failure, co PD diabetes, chronic kidney disease, dementia, all of that, you would still fall under the same category as someone with just eczema and Acne cos they still have two co co morbidities. And but if you only had um say stage four heart failure, then you'd only have one comorbidity. What I'm trying to get to uh get to is that breaking down a continuous variable into a more simplified categorical variable may not always be a good thing to do. The simple reason is that you lose granularity and if you are um understanding of clinical practice, then you know that there is a huge difference with a patient having one very serious comorbidity compared to someone having multiple smaller comorbidities. So they've got very good diet controlled diabetes type two versus insulin dependent diabetes. Though they are one comorbidity each, they're very different. And w that's point number one. So when you're critically appraising this, you could say, oh, there's oversimplification of continuous variables is why is H ba 10? OK. Why not 11? Why not? Nine? Um patients with an age group of nine aren't actively bleeding most of the time or? OK. Why not 9.5? Is there a difference? What if it's 10.01? It's still more than 10. And so questions like these is why oftentimes it's good to have a continuous variable, not always the case. Sometimes you have to break into groups simply because it makes clinical utility easier. Um All right. So in this case, their methodology was they had a bunch of predictive variables and they put it in a multivariable, stepwise logistic regression analysis, multivariable in the sense that they had multiple predictive variables. But it's still a binary logistic regression analysis because your outcome is binary. So what are some of the disadvantages of this paper? The first one is there is no comparative analysis. All they did was they did the core demographics and they went straight to logistic regression. They did not do any comparative analysis, like your independent samples, T test your paired samples, T test your um anova your chi squared, et cetera, your correlation analysis. Um And so you don't, yes. The logistical aggression is what you will eventually do. But if you do not do these parametric comparison tests, you cannot truly say that there was a difference between those who lived and those who died. Um There was also no test of normality. So the hemoglobin, for example, you don't know if it's normally distributed, but to be honest, it doesn't really matter because they're using it as a binary variable. And so you can see that they jumped through and cut corners in some aspects of methodology for a perfect paper, I would say. Um And the last one we we've talked about which is your dichotomisation of continuous variables. So basically, when you convert them into a categorical variable, um for example, 10 is an arbitrary cut off. 9.5 is good enough. Um So these days, a lot of mathematicians and statisticians are coming up with different ways of identifying what we call an optimal cutoff. So is there a mathematical way to come to this conclusion rather than just selecting 10? You know, why age 66 to 85 why not 65 to 85? Um And so some of the fancy said they're coming out with something called L OE ss which is smoothing and lowest efficient coefficient, et cetera. But, uh, out of the scope of this paper. Um, so there you go. Uh, that's your, um, critical appraisal done. Uh, you've analyzed the paper, um, you know, what should have been done. And so as a result of that, if it's not there, you just tick box exercise that, oh, they don't have this, they should do this. They don't have that. They should do this. Um, and then you can come to the conclusions. Um And you know, you can think of the conclusions as you read the results and if they don't match up, then you know that there's a problem. Uh you can then go on to the limitation sections of papers to see what have they mentioned. All these further things that you've thought about. All right, last slide, I won't keep you here for too long. Not so basic analysis. Once you've moved on from doing all this analysis, you can move on to more um complex um statistical uh methods. Um something called discrimination analysis or your rock curves, your receiver operating characteristic curves and area under the curve or AUC or your or whichever one you wanna say basically. Um That's how the curve looks like on your x axis. You have, have one minus specimen and on the y axis, you have your sensitivity and you're simply plotting each value. Um ie what was the diagnostic sensitivity and specificity of the RT PCR test? You can just look at it at all, you're just plotting, it's just a scatter plot um for every single uh point on your um for the, you have every single patient in your cohort. Um It's just a scat alot. The last one is survival analysis. So if your question is, will the patient be discharged in 30 days? That's a time or a temporal analysis. So you're trying to identify something in time and the things that happened before can affect that. And so you can do something called Kaplan Mayer or your cox proportional hazards or your log rank test, um which are similar to logistic regression and everything, but they account for your temporal variable. And as a result of that sort of an odds ratio, you would have something called your hazard ratio. Um this is something again to be aware of, you might see in a lot of clinical trials that they'll report hazard ratio rather than on ratio. Yeah. All right. Um The last type of advanced test, you can do something called a post hoc test. So what, what does a post hoc test mean the most eas the easiest example to understand is um when you have to do a anova. So if you remember from your table, if you had an independent variable with three categories, and you were trying to determine the outcome, which was a continuous variables across these three categories, you would do an anova. So for example, if you wanna compare the length of stay of three different age groups, 0 to 1820 to 4040 plus, you would do an Anova, however, the Anova will generate one singular P value. It's all it's gonna tell you is, is there a statistically significant difference in length of stay across these three groups? So it's gonna give you ap value more than, or less than 0.05. However, you don't know between each group. So you don't know whether that difference is between 0 to 18 and 20 to 40 or is it between 0 to 18 and 40 plus or is it between 40 plus and 20 to 40? That is where you do a POSTOP test. A POSTOP, a POSTOP test simply breaks down an anova into its principal components. It's gonna give you ap value for every single pair wise, pair wise, simply means a pair. So two variables for each pairwise association. Um Those are your POSTOP tests. Um You have different types, you're gonna have two keys bon for it doesn't really matter. Um Which one you choose? Um I would stick to two key. Um And then the last one is penalized logistic regression. Um I'm not gonna get too much into it, but basically, you can have an error term where if the model or your analysis is diverging, you can give a error term to converge um your analysis. Um But that's not as good as in simple legislative regression, simply because if your legislative regression model is diverging and that's a problem with your data. Not a problem with the model. It's either the question you're trying to answer can't be answered with the data that you've provided it with. And you have to do these further analyses in order to uh gain a model that gives you some good answers. Um But that's it. Um That's me, everyone. Um Hopefully you've learned something. Um But the two things are obviously that table and the last one is what I've said here, it's not daunting um statistics, evidence based medicine and critical appraisal. They can seem daunting, but you just have to get into the gro uh the groove of things uh and continue doing them. Um And the more you do them and the more you see other people do them, the easier it gets. Um But yeah, I'm more than happy to answer any questions, but I appreciate the fact that I've been speaking now for almost an hour um at you guys. Um But yeah, thanks for listening. Have you answer any questions if you have? I'm, I'm not sure if there are any questions I can't say much. There aren't any questions. No. Um I will once they uh oh there is one actually. Um So someone is asking de de Zymer has good negative predictive value as I know how, how does that relate to low sensitivity? So that's a really good question. Um What you need to understand is how you do a ddimer is very important and under what clinical context do you do it? So what does sensitivity mean? It means if your D dimer is positive, do you have that certain thing you could ask? You could ask the same question for proponents if you have an increased troponin. T what is the probability? You have a um, myocardial infarct? It's pretty high. It's the same reason why even in you, even in Ami, you're gonna have a high DDIMER. Always remember ddimer has good specificity, not sensitivity because you can have a high DDIMER. If you hit your toe on the side of a chair, what is DDIMER? It is when your fibrinogen breaks down into your fibrin forms the fibrin mesh that fibrin then gets degraded by your fibroblasts into DDIMER. So the degradation product of fibrin is DDIMER. So in your body at all times, you have this balance between coagulation and anticoagulation. And for example, you could have a high D dimer in D IC where you, where you're clotting and you're not clotting. At the same time, you can have a high D dimer in an M I, you can have a high D dimer in sepsis. You can have a high D dimer in chronic kidney disease where you can't excrete that protein. So in those cases, if your D dimer is high, it does not mean you've got a venous thromboembolism So, which is why it's got below sensitivity. But if your D dimer is low, then, you know, for a fact that you do not have degradation of fibrin at a level above normal. Um And so I'm not sure where you've heard that it's got low specificity, but it's actually, it's got a very high specificity because if it is low, then you know that you do not have excess coagulation. And as a result of that excess breakdown of your fibrin clot, um That's how I like to think about it. Um Hopefully, that makes a bit of sense, which is always why I never ordered ddimer unless I am, unless I've done the history, the physical exam calculated the wel score. Um And then you order a ddimer because um if you look at a lot of the nice guidance, it tells you lose specifically that um you or the D dima, if your dima is low, you can basically forget about it. Ie your entire um differential diagnosis, clinical reasoning was wrong um because it's not high. Um So that, that's my answer. So um a lot of things can increase it. Um But it's not gonna be raised. Um If you don't have it is the way you can think about it. The same for most, any biomarker, creatinine troponin and B um K. Um any biomarker, you always have to look at sensitivity and specificity. All right, any other questions, there aren't any currently I'm not sure if you can see them. But, um, let me hold on if it's on the right. I see a chat function. Yeah. Yeah, it's in the chat. There's, there was only the D 10, I see one. Hold on. Uh, Kaplan Meyer may have been possible with longer follow up times. Yeah. Absolutely. Um, you can do a Kaplan Meyer for any time to event analysis. Um, and that event could be anything so that mortality paper that we saw, you could do ak my for that. Absolutely because it's a time to some event. Um And they didn't do that. Uh They did a logistic regression which is not bad. But um what statistics, um statistical research has shown that when you're doing temporal analysis or time to event analysis, a cox aggression is able to provide you with more significant hazards or more accurate hazards um than the odds ratio from your logistic aggression. Um So any temporal analysis, you do cox aggression or Kappa. Um And anything that is nontemporal, you do a logistic aggression. But if you just did logistical aggression, you will not be penalized is what I have to say. Great, amazing. I don't see any other questions. Um But the my email is at the last slide and I'm more than happy to send the slides out. Um Hopefully, that was helpful. Um But yeah, yeah, so everyone who fills in the feedback form which I've sent out will get uh will be able to access the slides. And I've also, once we get all the feedback in, I'll create like a PDF export it to you so that you have it if you haven't wanted to. I think it's helpful, but otherwise I think, I think that's it. Thank you again. No worries at all. Uh Thanks so much. Um Yeah, see you around. Thank you everyone. Thank you. Bye bye.