Voices of Community Medicine- Episode 2 (AI in Community Medicine) #psm #medico #ai #trending

本站添加: 2026-05-28

[00:00:00]Hi and hello everyone welcome to GKS community. So today we have a special guest uh we have Dr. Sep Das senior resident name as buneswar sir first a warm welcome to this channel sir so I'm glad to I'm glad to do a podcast with you so today uh in voices of community medicine we have a second episode that is a in community medicine uh you are the better person I think so I thought you are the better person in uh a that I have worked with so I thought do a podcast with seps A in community medicine. So no better person than you sir. So so just uh give a overview sir.

[00:00:45]What is the what is exactly the a in community medicine in simple terms to the audience please?

[00:00:52]>> Okay first of all thank you for inviting me to this podcast and um hi I'm Sep Das. I work as a senior resident in uh aswa in community in the department of community medicine and family medicine and uh to answer your question what exactly does AI in community medicine mean?

[00:01:16]>> Yes sir.

[00:01:17]>> We first have to understand what community medicine actually means.

[00:01:21]>> Yes. Yes.

[00:01:22]>> And to do that I'll first give you a summary. Okay. Yes.

[00:01:27]>> By telling you a quote.

[00:01:30]A jack of all trades is a master of none. But in some cases, he's better than a master of one. Okay? So if you look into community medicine, I would say that community medicine is a super speciality.

[00:01:51]Now many might disagree, but why am I saying that community medicine is a super specialty? Because it first of all it does not look at uh just one body part. It looks into the entire body.

[00:02:06]>> Exactly.

[00:02:06]>> How it works together and not only that it does not does not look at only one human person. It focuses on population levels. Okay. It takes the knowledge of each uh subsp specialtities such as your cardiology, neurology, the nephrology and many others. They combine all that knowledge and they try to see how can we be applied not only on an individual level but also in a family level and then in a community or a population level.

[00:02:41]>> Yeah. Yes. So and we also work here with uh not only ourselves but we also work with people like Ashas and ALMs.

[00:02:52]>> Exactly sir.

[00:02:53]>> Okay. Now coming to AI in community medicine.

[00:02:58]AI currently has been has been used uh is being used by almost all departments of >> AI is widely used every >> Yes. In radiology it is being used. In dermatology it is being used and now other applications of AI is also being explored.

[00:03:18]>> Yes. Yes.

[00:03:19]>> But in our case it goes beyond that. For us AI is not just a robot to which we'll give instructions and we'll and we'll get answers.

[00:03:30]AI for us is a tool that helps us make sense of large messy population data.

[00:03:39]flag where a problem is emerging, translate the health information ac across the languages and reduce the paperwork burden on ourselves as well as other people who are working with us the allied health professionals such as the nurses and your&ms your ashas etc. Now I should add one thing out here right away because this will change how you will hear the rest of the conversation actually. AI is not just one thing.

[00:04:09]>> Let me be very clear.

[00:04:11]>> Yeah.

[00:04:11]>> AI is not just chat GPD. AI is not just clawed.

[00:04:14]>> Yeah. Yeah.

[00:04:15]>> It is actually a full stack.

[00:04:17]>> Exactly.

[00:04:17]>> A ladder of methods.

[00:04:19]>> And this is a message to community medicine residents. Those who are currently not using AI actually don't realize that you are already using the basics of AI >> such as in uh in statistical software such as STA or in in R or in Jamov and all >> the first basic analysis is the regression analysis. You're doing the linear regression, the logistic regression that is one of the basic method of machine learning. The m machine learning forms the foundation of artificial intelligence.

[00:04:57]>> Exactly.

[00:04:58]>> And these are the are the base methods of machine learning.

[00:05:02]>> So you are already into AI.

[00:05:04]>> Yeah. Yeah.

[00:05:05]>> You should not think that you are outside AI. You are already into AI.

[00:05:08]>> Now today I will talk to you about how you can put it one step further. So that is what I'm trying to convey in AI in community medicine.

[00:05:19]>> Yes. Ex actually you're saying it's so widely used. You should never avoid using uh AI in community medicine.

[00:05:26]Actually you are already using community medicine in basis of any models in bas of any software like that. So uh my second question is so why a becoming important in public health and community medicine now? So why it is important now?

[00:05:41]>> Okay. So to that I will need to tell you >> yes sir >> the current picture three three things have converged at this moment three number one we are now generating enormous amounts of health data enormous >> yes yes exactly sir >> we have surveys like the national family health survey we have the longitudinal aging study of India we have now the Aayish manh digital mission >> we have uh we have the registries.

[00:06:13]>> Yeah.

[00:06:13]>> Like you have the stroke registry, the cancer registry. Now, previously it was probably paper based. Now it is being shifted to being completely digital.

[00:06:22]>> Nowadays, each and every hospital has a has a hospital management uh information software.

[00:06:29]>> Exactly. Yeah. Yeah.

[00:06:30]>> This generates a lot of information, a lot of data, far more than any human can handle by himself. for that uh that person will need a very large team of analysts and even then they are uh it takes lot of time to for them to to analyze >> the if you if I'll tell you one example the national family health survey >> the sixth round of national family health survey the data collection has already been done in around 2024 and 25 the data collection has already been done but the results are not yet out in the public because uh because uh the statisticians >> Yeah.

[00:07:14]>> are still now analyzing it.

[00:07:16]>> Yeah.

[00:07:16]>> They're still now analyzing that huge amount of raw data that they have in their hands.

[00:07:22]>> Yeah. Yeah.

[00:07:23]>> Okay. And and data analysis is not just you know application of methods. It is also regarding data cleaning. The huge part which sometimes we skip. Okay. So this is one this is one point. This is one thing that has has come in front.

[00:07:40]Second, the tools which we were previously using before the advent of artificial intelligence has crossed a usability threshold.

[00:07:49]>> Yeah. Yeah. Now, now the capability of AI models have increased so much that now you no longer need a computer scientist to uh to make you a capable model and that matters enormously in in uh resource limited settings.

[00:08:07]>> Exactly. Exactly.

[00:08:08]>> Such as a lone researcher, public health researcher like me sitting here.

[00:08:12]>> Okay. If in the if the AI would not have been possible, I would have to call up a computer scientist or a coder to make a model for me that analyzes this data and gives me the inference.

[00:08:24]>> Yeah. Yeah. Reduce more burden more burden like >> Yeah. But but that burden has reduced.

[00:08:29]>> Yeah. Yeah. Exactly. Exactly.

[00:08:31]>> Now the third Now the third thing is our health work workforce is still not finite.

[00:08:37]Okay. Our demand is rising and rising and rising. So but our health workers workforce is very finite. So, anything that can absorb the routine clerical load or the cognitive load such as filling out of forms, uh transportation and and many kinds of scut work, anything that can that can lift uh lift us off this load and they can and take take these instead while we focus more on focus focus more on inference and eventual interventions.

[00:09:12]>> Exactly.

[00:09:12]>> That is very valuable. So three things one is the huge amount of data. Yeah.

[00:09:17]>> Second is uh second is that now uh now making tools has become much more easier than before and third the demand is high and the uh and so the clerical workload has increased and we need something to reduce that workload.

[00:09:35]>> Exactly. Exactly. So there's a still lack of very important.

[00:09:40]>> Yeah. There's still lack of knowledge and awareness on using AI. So what's your point on this s because uh nobody is uh ready to adopt this AI tools you know just as you discussed it will reduce the burden of the healthare workers but there is a people like like experts also they are lack of knowledge and awareness they have failed to adopt that so what's the what's your opinion on that what's your perception sir >> see in this regard I would say that the PE uh that people who have just started into AI who have started into things like just chat GBD and plot I will not denigrate them because chat GBD and plot is a very good way to act as a starting point. Yeah, that's the starting point.

[00:10:29]That's that's where we all all started.

[00:10:31]Right.

[00:10:32]>> Yes.

[00:10:33]>> Yeah.

[00:10:33]>> But the thing is that artificial intelligence goes beyond that.

[00:10:38]>> Mhm.

[00:10:39]>> Okay.

[00:10:40]>> Yeah.

[00:10:41]>> Artificial intelligence goes beyond that. And to go beyond that also you need to you need to understand in greater detail in which stream you are in which area are you working on because let me tell you one thing okay you are not going to make an AI model you are not going to make a machine learning model >> exactly >> what you are going to do is you are going to find out which is the uh uh which uh you are going to first find out the problem.

[00:11:14]>> Yeah, >> you are going to first find uh then the second is you are going to find out that what kind of data is emerging from that problem. And the third thing is what AI tool that you have across the world will be most suitable to solve this kind of problem by using this kind of data.

[00:11:34]These are the three things.

[00:11:35]>> Okay. Yeah. as a community medicine resident or even as au as a as a researcher in community medicine, you are not going to make an AI model. That is the job of an of a machine learning engineer or an AI scientist.

[00:11:48]>> Yeah. Yeah. Your job is that to uh to understand the problem acutely to understand what data is coming from from that and then which AI tool is best to solve this problem in a most effective and feasible way or in the more economical.

[00:12:08]>> Yes sir. How to how to find that this tool is very helpful in community medicine? just uh some people don't understand which tool to use and uh what kind of EA method or what kind of models use in community medicine. Uh any point on that?

[00:12:24]>> See first of all the thing is that uh in uh in this question I would say that how to use or how to know which model to use.

[00:12:33]>> Yeah. Yeah.

[00:12:33]>> The starting because uh even the internet and this one is a little bit sketchy can be sometimes.

[00:12:40]>> Yeah. Yeah. the very starting point you can ask to the very starting point itself the claude or a chibi or gro or anything. Yeah. Yeah. Yeah.

[00:12:49]>> You can just you can just tell them your problem.

[00:12:53]>> But while telling telling them your problem, tell it in in very minute details and tell them tell them your problem in very minute details along with them. Tell them that what you actually want. Okay. So let me tell you one example. So there is something called an exploratory data analysis.

[00:13:16]>> Yeah.

[00:13:17]>> Okay. Most of us we do not do exploratory data analysis.

[00:13:21]>> Exactly sir.

[00:13:23]>> We do the uh only the level of analysis which our protocol tells us.

[00:13:29]>> Yeah. Yeah.

[00:13:30]>> We when we make the protocol whatever we have written as statistical analysis we do only that we do first present a descriptive analysis then we do a a linear or logistic regression on only a few of the variables and that's just it.

[00:13:45]Yeah, we are not going right or right or left.

[00:13:47]>> We are not we are not trying to find associations between two seemingly unrelated variables or what can be their interaction between them.

[00:13:56]>> Exactly. Okay. Exactly.

[00:13:57]>> How are they interacting? How it is is it an additive effect or a multiplicative effect? We are not doing that.

[00:14:04]>> Maybe the first thing is exploitative analysis. And for that and for that many uh let me tell you uh to your audience that the first thing you can do is take the take the data set give it to the your AI uh service such as charge or claude and tell them that I want to do an exploratory data analysis of this data set >> and and u and I want to go beyond the protocol whatever the protocol I had made.

[00:14:35]Now what I uh so help me in this but but but do not tell the uh do not tell the uh tell the AI model to analyze it and give you the results.

[00:14:46]>> Exactly. What this >> this is the problem in like they're using AI for analysis. The problem is like most of the people's like they want the analysis from AI. So that is the issue.

[00:14:58]>> No that is the mistake that is the mistake that that that most people do >> in this case. what you uh what you should do is you already and here I'll tell you that at this moment it is advisable to shift yourself from softwares like jamovi and SPSS to uh to R or Python >> m >> okay then what then the next step would be that you don't do the you will tell the AI that you don't do the analysis instead you see the variables You see the d uh you see the variables, you see the data and you see my protocol and based on that write me a full code script. Write me a full code script that does the analysis on these variables based on whatever methods of analysis that we have in statistics so far.

[00:15:52]>> Exactly.

[00:15:52]>> And it will gen and for example in in R they'll generate you an R file >> or in STA they will give you a STA 2 file or in uh or in Python they'll give you a Python file. You then download it and then you run it. It will give you the results.

[00:16:09]>> Yeah. Yeah.

[00:16:10]>> It will give you the results. It this results will be uh will have no hallucinations >> which the AI which the AI model is prone to make hallucinations but this won't be prone to making hallucinations.

[00:16:22]>> Exactly. Now whatever the results you have you then transfer it back uh back to the AI just tell them that what what relationships do you do you see in uh in this in this data based on these results.

[00:16:34]>> Exactly.

[00:16:35]>> Can you get it?

[00:16:36]>> Yes. Yes. I think more misconception has been broken down by this separ. So next question my question is like uh sir do you think a will transfer our uh community medicine landscape or healthcare system here? So what do you think about the A in future sir in especially in communication?

[00:16:56]>> Okay. So the thing is this will AI transform healthcare or not?

[00:17:02]>> Yes. Yeah.

[00:17:03]>> Yes. AI can transform healthare but I would be very cautious with the word transform.

[00:17:09]>> Yeah.

[00:17:09]>> Because in public health this word has disappointed us before. Okay. Plus I'll say AI will meaningfully change how care is delivered. Okay. And the change is already underway.

[00:17:23]>> But the transformation of the outcomes which we desire depends on the more unglamorous thing. AI is a very glamorous thing at this moment and it is a hot p like selling like hot cakes.

[00:17:36]>> But the transformation of but the but what transform do we desire b is based on more unglamorous things. What >> first is data data quality >> then you have then you need electricity and connectivity in places like a a subcenter or a PHC and then the third is whether this tool fits the actual workflow of our of our workers for example Asha and M and the fourth is whether people trust it or not. M >> okay the real risk in in artificial intelligence is that it widens the gap between well resourced and poorly resourced systems rather than closing it.

[00:18:18]>> Exactly.

[00:18:18]>> Hence my honest answer is that AI artificial intelligence is a powerful enabler but whether it transforms healthcare equitably it is a policy and implementation question not a technology question.

[00:18:32]>> Okay.

[00:18:33]>> Yeah. Yeah. So at the end I would say all models are wrong in most of the cases but some are very useful. So that's why no model no model artificial intelligence or otherwise perfectly captures reality. It just approximates the parts that matter well enough to act on. So that is not a flaw in AI.

[00:18:58]>> It is the whole nature of the data you are collecting.

[00:19:01]>> Exactly. How you how are you uh putting relationships across them? It is uh that matters that uh it is not the flaw of AI.

[00:19:11]>> Yeah. Yeah.

[00:19:12]>> Yeah. Yeah.

[00:19:13]>> This is the healthiest way to approach.

[00:19:14]>> Yeah. Basically, you're saying uh we have to use the A with caution and uh you should know you should have a basic Yeah.

[00:19:22]>> No, I say no.

[00:19:24]>> See the thing is whenever we hear the word caution >> then at the time we become very pessimistic. No. No, we should be selfaware.

[00:19:33]>> Exactly. Exactly. Yeah, >> we should be self-aware. We should not be cautious. We should be selfaware. We should be optimistic >> but at the same time we should be self-aware.

[00:19:42]>> Okay. We should not blindly uh we should take the help >> but do not but do not blindly accept whatever results come from that.

[00:19:52]>> Exactly. Exactly. So uh my second question is the fourth question is like what are the common aid tools that are currently used in healthcare sir you any idea what are the a tools now evolving or using in medicine >> now now this is a question which then goes beyond the common tools that we use >> okay there are specialized AI based software >> exactly exactly >> you have diagnostic imaging tools >> artificial intelligence reading the chest x-rays of of fiber gloss process, >> retinal scans for diabetic retnopathy, they're already being deployed.

[00:20:28]>> Yes. Yes.

[00:20:29]>> Then you have large language models, okay, which are helping in documentation, literature review, uh, drafting and decision and decision support.

[00:20:38]>> Yes, sir.

[00:20:39]>> Okay. There are conversational translation tools that let patients interact in their own language.

[00:20:44]>> Exactly.

[00:20:44]>> But again, here is the part most of us miss >> almost all of us. Again, I say it is experiencing a very single layer of of artificial religion. M >> one person is typing one prompt into the model and that is the shallowest possible use. Okay. This above it sits two whole dimensions of depth. Number one there is depth in the models themselves.

[00:21:05]>> Yeah. Yeah.

[00:21:05]>> At the base there is there is a logistic regression, linear regression. Then you have then you have more richer methods of analysis like decision trees >> uh support vector machines etc. Then you have neural networks and deep learning.

[00:21:22]Okay.

[00:21:24]And then the and the next depth is composition. How you take two or three different models and assemble them into something useful. Okay.

[00:21:34]>> The lowest the lowest level is one human, one prompt, one model.

[00:21:39]>> Exactly.

[00:21:39]>> But now we have to go towards multimodel pipeline.

[00:21:43]>> I'll tell you one uh because I am not only working, I'm trying to develop some products also, AI based products also.

[00:21:51]>> Exactly. So in the in those AI products um I hope to to roll it out in some in a few months. Okay. In that I uh in in that there there will be uh there will be two two models. One will be translating >> from from from the local language to English and then the second model will interpret it will interpret it and and give us the results.

[00:22:18]>> Okay. I won't go into further details about my product when it will be released you all you all guys will know.

[00:22:24]>> Exactly. Exactly.

[00:22:25]>> And and I will hope sincerely that when you guys see my products >> I hope you love it.

[00:22:31]>> Exactly. Just wait and watch and we will give feedback on that sir. So uh exactly what we are speaking about the AI tools as we as you mentioned like uh every everyone seeing the first layer of AI we are not able to go beyond that point. So that's that's the uh point where we have to uh gain a knowledge about the ASR that's what I that's what I mean like so my next question will be like um since how this as we speak in community medicine so here epidemic pandemic outbreaks so how this AI will useful in this kind of uh situations like epidem ep epidemic or outbreaks >> okay yeah that's a very good question >> yeah So uh you have read about the headons matrix right?

[00:23:21]>> Yeah. Yeah.

[00:23:22]>> Headons matrix. Headons matrix it integrates uh agent host environment and then you have pre-event event postevent.

[00:23:31]>> Yeah. Exactly.

[00:23:32]>> Okay.

[00:23:33]>> So here the agent will be your AI.

[00:23:41]>> Okay. Okay. The various models of AI post will be us and the environment here would be the the particular epidemic or pandemic. Okay.

[00:23:50]>> Yes. Yes.

[00:23:51]>> And on the other side we have pre-event event and postevent. Pre-event is just before the pandemic or just before the epidemic.

[00:23:58]>> Yeah.

[00:23:58]>> Then it uh then event would be the epidemic and the pandemic.

[00:24:02]>> And the third would be post event was post epidemic or post pandemic. Okay. So let's look at into a special headon matrix. Okay. So in the pre-event like before the epidemic starts before the epidemic starts also just like we know in a in any infectious disease we have something called a incubation period.

[00:24:24]Okay.

[00:24:24]>> Yeah. Yeah. So uh I believe that every epidemic has an latent period also and and in the during this latent period data comes signals do come but they are so they are so small and filled with a lot of noise uh that ordinary humans are unable to analyze it. In this case AI can scan those diverse signals. They can they can look into search trends. They can look into news. They can look into syndroic data. They can look into even animal health and they can give us an early warning. That is the basis of event based. We we already have event based surveillance. Yes.

[00:25:08]>> But AI can empower it even further. Now that is the pre-event.

[00:25:14]>> Yeah.

[00:25:14]>> Then is then comes the epidemic.

[00:25:17]>> That is the event area.

[00:25:19]>> Yeah.

[00:25:19]>> In out here it can help in rapid diagnosis. it can sift through routine surveillance data for anomalies that that a human would miss.

[00:25:30]>> Okay.

[00:25:32]>> Then at post event this is not post does not mean that the epidemic is over.

[00:25:37]>> Mhm.

[00:25:38]>> This means that when the epidemic has has reached its peak and is starting and and is now in it way to wreck havoc in the health systems. In this case, artificial intelligence can help to anticipate what amount of case loads can we get and then they will allow they will help us in judicious allocation of beds, oxygen and staff.

[00:26:02]Something that that we had learned the hard way during during COVID 19.

[00:26:06]>> Yeah.

[00:26:07]>> Okay. And then AI can also help us counter the information pandemic through rapid multilingual communication and misinformation detection.

[00:26:17]>> Okay. Again I would say this caveat. A model is only as good as the surveillance data feeding it.

[00:26:26]If you are not feeding the uh the right data then any model any model will always be wrong. That's why all models are wrong but some of them are very useful. Okay.

[00:26:38]>> Yeah.

[00:26:38]>> So the limiting factor is the limiting factor out here is the strength of the underlying public health system and the people inside it not the algorithm because algorithm is not a human being.

[00:26:49]You are the human beings you are here to decide.

[00:26:52]>> Yeah. Yeah. Exactly sir. So um so as we speaking about community medicine main thing is research research sir. So in research what are the ethical limits of using AI sir? This is the most important question I think because most of the people's use writing a manuscript using AI. So what are the ethical limitation in AI? When to use where to where not to use? So just uh clarify in this point sir.

[00:27:21]>> Okay. Okay. So so there so there are some limits.

[00:27:24]>> Yeah.

[00:27:24]>> Especially in research. Okay. And it's not that AI cannot be used in research.

[00:27:30]Let be very clear. Uhhuh. Um there are some there are some people in our fraternity uh who are very very cautious about AI.

[00:27:41]>> Yeah.

[00:27:42]>> There are some who are who are still who have started to adopt it but they do not know how much further it can go. Some of them take it blindly and there are many kinds of people who are using AI at various levels. Okay.

[00:27:56]>> So what are the genuine limiting factors that I can tell you? Okay. First one is the honesty of authorship. First one is the honesty of authorship. AI can assist with your language and structure of your manuscript but it cannot be an author.

[00:28:11]>> Exactly.

[00:28:11]>> And it and it cannot take responsibility.

[00:28:15]>> And any substantiative use beyond language and structure should be disclosed. Okay.

[00:28:22]>> The second limiting factor is data consent and data provenence. Okay.

[00:28:28]And why is it because much of these AI based models which had been made whenever these models are made they are uh if you go into very basic details of AI or after making a model a model is first trained to refine its inference and then it is given to us for use.

[00:28:55]Okay.

[00:28:57]Much of these models were trained on had been trained on data that has been collected without anyone of the people imagining that it would feed a machine.

[00:29:08]For example, Wikipedia. Wikipedia was an encyclopedia which in which people had had done edits and all of that.

[00:29:20]>> Then you have Facebook.

[00:29:21]>> Yes. Then you have you have X or Twitter, many other websites with which we constantly interact >> and we share our personal feelings, our personal information and all of that into it. AI models had collected this data to train themselves. Okay. in health research. This is an unresolved ethical question that we are only beginning to confront because those data which had been which had been to which they had been trained had some of them have been taken without consent. Now if we are going to give health data uh to that model so will it be without with the patient's consent or without the patient's consent and if we had collected we had collected data from the patient with consent have have we taken an additional consent from the patient that this data will be provided to an AI model.

[00:30:18]>> We have we have not we have not introduced this aspect of it in our consent forms.

[00:30:23]>> Exactly. Exactly. Third thing, now the third limit is fabrication risk. What happens is that many times that uh whenever we feed into the uh we feed data into the uh into the AI, we tell them to analyze it.

[00:30:42]>> The AI the AI artificial analyzes it. It gives us the results. M >> they have a particular uh you know uh they have a particular bias.

[00:30:56]>> They a psychopy bias. They tend to tell us they tend to be yesmen.

[00:31:01]>> Yes. Yeah. Yeah. It will give positive positive kind of output.

[00:31:06]>> They'll they'll tend to give us a more pos Exactly. They'll tend to give us a more positive kind of output.

[00:31:11]>> Yeah. Yeah.

[00:31:11]>> Okay.

[00:31:12]>> They will be very fluent. They'll be very confident that sir your the results which we have got are substantiative >> even though you can see that there is not much substance in it but they will give you as much as one >> they will completely invent the references and and findings hallucinate findings. Yeah, exactly.

[00:31:30]>> Okay.

[00:31:30]>> If you go into reference, there will be no sentence like this.

[00:31:34]>> So that's why suppose suppose you have you have asked the EI to generate this a particular manuscript and they have given references also at that time you should read the manuscript and not only that you should you should copy paste you should copy each of these references paste it in the Google search and try to see if that either actual research article exists or not. it won't be there >> and whatever and whatever they are citing and and and which and whichever reference they have has been used to site a particular sentence in this one does that actual reference say that or not >> okay >> you should verify >> yeah it reduces the burden of you going and searching for the references >> but then it introduces another another burden that you have to verify >> verify except class verification should be done >> because uh >> yeah so my bottom line is that AI is a legitimate tool for thinking to some extent and drafting to a larger extent but the scientific responsibility remains fully non-transferably human.

[00:32:41]>> Exactly sir. Uh so it mainly this ethics ethical issue in AI is like I'm seeing like um currently in most of the journals they asking for if you used AI please give a concern to that they also they also like ready to accept if you are using AI but it should be like as you it should be human author should be human but in case of rephrasing grammar this kind of paraphrasing can be done using AI >> so this will be uh this will be a take So, so uh my last question will be sir like uh so what a skills should future doctor will learn. So this a skills should be learned for the f uh doctors in the future to cope up with the a as a evolving force.

[00:33:32]>> Okay. What what courses should they learn? Okay. That's >> not about the courses. It's a skills.

[00:33:37]Yeah. What kind of a as we as we speaking like as we mentioned like there only one layer we are seeing beyond that point nobody's going there. So how to go through the beyond how to go beyond that point. So what is the way?

[00:33:52]>> Yeah.

[00:33:52]>> Okay. Okay. Okay. Yeah.

[00:33:55]out here. I would say that sometimes some y doctors might feel that uh because we're not using because in this aspect most of the most of the people in the market when they when they are they are into artificial intelligence into machine learning and all they tend to say that uh you need to learn how to code how to go into the very core of things, how to manipulate models, tinker the models and all of that. Okay?

[00:34:30]But I but I would raise caution to the wind and I would say to all my peers, my juniors and all, you do not learn you do not need to learn how to code. Okay? The most important skills for using AI in community medicine are very much conceptual.

[00:34:51]>> Very much conceptual. The first one is you need to learn how to prompt well how to give data properly to the to the AI. How to make the AI understand the nature of the problem that you are currently facing >> and how to use it judiciously. This is the first thing. Okay. How and also how to distrust and verify their answers.

[00:35:16]>> That is what what I saying judiciously.

[00:35:18]How to distrust and verify their answers.

[00:35:20]>> Exactly. Secondly, data literacy. Data literacy is understanding what makes data good or biased.

[00:35:29]>> Because that judgment separates us from useful AI use between that that separates us that separates useful AIUS from dangerous AI use.

[00:35:41]And thirdly, an understanding of ethics and governance because because you guys will be the ones that will be uh that will be asking the patients for consent and safeguarding the data. Okay. And on the third point point it uh I would like to tell that it is not it is not a law lawless lawless anarchal ground. Okay.

[00:36:03]Because India standalone medical AI is now regulated as a software as a medical device. SMD.

[00:36:09]>> Yeah.

[00:36:10]Under the central drug services uh committee which is a CDSO >> the ICMR has also issued ethical guidelines specifically for AI in healthcare. Patient data actually sits under the digital personal data protection act and a and a man bharat digital mission also has a particular sandbox a testing area where these AI tools can be tested against the national health registries before any formal audit. A doctor does not need the fine print. Like I said, a doctor does not need to know the nitty-g gritties, but knowing this framework exists is part of being a responsible user rather than being a nave one. And fourth, lastly, for all community medicine residents to not stick into the one model, one human, one use thing.

[00:37:02]>> I would say go beyond that into one very specific skill set called workflow imagination and workflow integration.

[00:37:10]Imagine the workflow.

[00:37:13]>> You imagine you imagine problems and you imagine solutions as workflows. That step one, step two, step three, step four, component one, component two, component three, component four. When you break it down, then you'll be able to realize that okay, one model is not being able to solve all of all of these component. One model is not being able to to handle all these components. Maybe the the model which I have currently is useful for one particular component or two particular components but but for the rest of the models but for the rest of the components maybe uh another model is much better equipped to handle that.

[00:37:50]M yeah >> for example let me tell you okay >> many people many general people today nowadays upload their chest X-rays ultrasound ultrasound images into into uh Gemini or Chad GPD >> hoping to get well furnished results and all of that but because Chad GPD claude and all of that has been trained not only on X-rays and all they have been trained on a large amount of of non-medical data. So sometimes while trying to answer your input of an X-ray of an X-ray image they will introduce some non-medical things also they will try to in infer it in a way which is far away from the truth. So in this matter Chad GBT, Claude, Gemini even though they have been tested and and they have been and uh model and Gemini has been has been ranked very high in interpretation of medical information still I would say you should be very cautious in in using that.

[00:38:55]>> Now what to use there is one model called Medjma. Okay. There is one. So Gemini if if you go into the history of of how Gemini Google Gemini has come into form. Before that there were there were models called JMA models. Okay.

[00:39:13]Gemma models. Then you came into Gemini and now you use the Gemini right? Gemma models many of them would not know.

[00:39:20]>> Yeah. Yeah.

[00:39:21]>> Some people had some people had taken the some of these Gemma models and had specialized them in medical information only. So they are the medjma models and I would say that medge gemma models would be much more useful >> in suppose interpretation of an X-ray than Gemini which is much much of a general purpose AI.

[00:39:42]>> Exactly. Exactly.

[00:39:42]>> Okay. Yes.

[00:39:43]>> And when and whenever you are putting it in in a workflow. So one component in for example inter uh uh for example translating translating uh odia to English translating what a person is speaking I would say better than gemini would be something a model called the server mayi model >> you have sus v3 model of server mayi which is much more capable in honestly translating so when you are partic you're building suppose a system which acts as a clinical ical decision which acts as a clinical decision model you will uh so first you'll make a workflow then for patient input whatever the patient is saying for that you'll put SARS AI SARS V3 model of of server AI then for uh for things like X-ray images ultrasound images you'll put MEMA >> then for uh then for interpretation of all these information that is coming from the from these models then you put a final Gemini model >> so you can see a workflow. Okay. So, this is the way you have to start visualizing and and using the AI in the next level. Okay.

[00:40:55]>> Exactly.

[00:40:55]>> Uh you and let me be very clear. Let me be very clear.

[00:40:59]>> You do not need and now I'm speaking more and more about AI. Use AI. Use AI this way. Use AI that way. Then you also have to recognize when not to use AI.

[00:41:10]>> Okay? When not to use AI. Now you do not need AI to you know uh sort the numbers and calculate the mean and the median and all. You do not need artificial intelligence for that. You can also do it by yourself.

[00:41:26]>> And we do in Excel also. No issues.

[00:41:28]>> We can do it in Excel also. Okay. And then and then even linear and logistic regression. Yeah they are they form the basics of the machine machine learning.

[00:41:36]>> But you do not need a very highly advanced model to do that. You can do it on your own.

[00:41:41]>> Okay. So these are places where you do not need an AI. Now then you are in the actually in the middle. This one model, one human, one uh one use, you're actually in the middle.

[00:41:54]>> So you need to know that sometimes you have to shift to places where you actually don't need AI and using AI actually overuse it.

[00:42:01]>> Yeah. Yeah. and and and you need to sometimes shift into that place where under use of AI is happening and you need to know which model to use how to put it in an integrated workflow and all >> exactly >> the doctor and the final message would be a doctor who can who has real domain knowledge about what he or she is is doing in this particular field like for example in community medicine if he knows what kind of data is he working on what kind of people are are is he working with in every minute detail and then he will try to see which uh which tools can be used for very specific purposes that doctor will be far more valuable than either a pure technologist or a computer scientist or a doctor who refuses to engage at all.

[00:42:54]>> Exactly. Exactly. So what my understanding is for the future doctors there's a way distance to go you have to go beyond that point and you have to see and you have to explore more models like as we you told like many a tools are there try to explore more for translation there will be a separate AI tools there and kind of analysis there will another tool is there so try to explore more and uh I hope this will help somewhere of the future Dr. sir. So thank you. Thank you so much sir uh for this opportunity to host you sir. So I hope some of the some of the residents or some of the UG students also they will learn uh what is AI in community medicine what will the future and because A is the future as everybody is saying like uh India is becoming a center for AI. So everybody will use uh this AI in future and uh um as we already mentioned like uh so many people's they are sitting in that first layer. So uh yeah yeah yeah so this podcast will uh uh just uh motivate them to go and search behind the point and uh I hope this video will be helpful for all the students medical students and residents and all general peoples also. Uh so thank you so much sir uh for >> thank you thank you >> for come and uh speak of these words. So we'll do more uh podcaster uh like this in the future. Yeah. Y >> and wish you all the best and to all the audience also uh that that you guys are are not away from AI. Those who are not using chat GB and all of that you're not away from it. You are unconsciously already started using it.

[00:44:46]>> Exactly.

[00:44:47]>> Now you need to take the next steps. You need to know when to use AI, not not when to use AI, how to use AI, how you can put everything together and and make a product, >> do some difference in the world.

[00:44:59]>> Exactly. Exactly. So one sentence I would say like use AI. You can use AI but in cautious in cautious thing like you have to be cautious while using AI.

[00:45:09]Don't try >> again. I would say don't use the word caution because that's a pessimistic word. I would say use you uh use AI judiciously with self awareness.

[00:45:19]>> Yes. Yes.

[00:45:20]>> You need to have self awareness.

[00:45:22]>> Exactly. Don't don't expert AI will do analysis and all. Just just uh it will give the code and all and uh so use a more specifically. Uh this is the take-h home message I think.

[00:45:35]>> Okay sir. Thank you. We'll uh do more podcast and we'll meet in the future also. Thank you sir.

[00:45:41]>> Thank you.

[00:45:41]>> Thank you. Thank you.