TWed Talk: Neha Keshan on "Stress and Machine Learning"

Key Takeaways

is it alright yeah you can do it you can sit down so good evening everybody welcome to our last normal twit of the fall 2018 season in a couple weeks we'll have our end of the term lightning talks which everybody is excited to do you'll be getting notices about signing up for that shortly and next week is an hour pirates may hang up in Amos Eaton it's a little bit logisitics like usual we are streaming this on the interweb for posterity and you feel free to ask questions yes we're happy to help Nihad do our presentation tonight she's going to be talking about work that she's done and I called the fascia show talked through that and it's it's when when I was looking for volunteers for twitch she said can I talk about anything I want to talk about this so we're really excited Twitter's for us to learn and to understand these things and and this is very cool topic so take it away here thank you yeah please feel free to ask questions and if at any point of time so topic for the day is chess and machine learning and the work was done at advanced wireless systems Research Center at SUNY Oswego and system security and machine learning lab at Syracuse it was so before we understand what is the connection between stress and machine learning let's no-fee major components of our whole talk the first thing is stress now many people just listen this word I am they're like oh it's chest has to do something with negative thing all it has to do with something like ill-effects or something but what stresses stress is nothing but the way our body responses to every life event that we go through at every moment in our life and it also helps us to it says whether we should fight or flight say for example you are in a forest I know and I'm a lion comes in front of you then how do you know what you to do that is where stress comes in action it tells you need to fly from and it is a place or in a situation where you to stand and you need to fight for yourself that is also where stress comes into action and it gives you more strength than you think you have so that you can fight for yourself now as I said stress love two majorly two different kinds first is you stress which happens to us when we are happy for events like promotion and making new friends getting a graduate degree need marriage or going on a vacation or similar kinds of events they are termed as useless because they give us positive energy and it give us motivation to do something new and exciting and helps us to perform better but on the other end which mostly people connect to a stress is the negative feelings or the click if someone close to you dies or work issues or a feeling of getting neglected if those kind of feelings are there that creates negative stress which is D stress and that really takes a lot out of you it gives you anxiety and if it's a long term it can create much more trouble and make you a chronic patient and can give you a mental and physical problems and these kind of stresses are outside of coping capabilities so how so we know this is nothing but respond power what he does for flight-or-fight how it is happening whenever we are in a situation as I said every moment in our life we are undergo some kind of stress be it gold be it bad some low amount of stress or high amount of stress what happens then these two hormones adrenaline and cortisol they get into action they help us to understand what our body should behave like and what should it be its response now it helps us by increasing our heartbeat and simulating force force variations and it also helps us to access a tissue repair substance and alters the response of immune system but if we undergo stress for a long time it means that our heartbeat is working more than it should be for a longer time which gnorm effect which is implied bases can diminish or you get hypertension which can end up like which can also give you make you at a risk of cardiac arrest it can also lead to type 2 diabetes and so on and so forth so paper talking about stress it depends from person to person for some person an event can be a stressful event whereas for the other it might be like joyous stressful event event C so like an example I love swatting max so for me any problem of my I can't give me any stress in her like a T stress man but a person who hates Matz even as normal addition algebra problem can make them person cool like and if these stress motion like anxiety and other stuff so it depends a person to person and the level of stress for me which might be like low might be high for someone else and my see Bursa so we need to know that how the level of stress differs from person to person and that is why it is important to make a personalized system which can detect stress based on the level of that person which is good for that person and not make a change in one morning what happens and why this system is so necessary because when we are growing up with the similar kind of stress our body gets used to it and even when it is having of facing some kind of stress our body tries to ignore it deny that it's in a state of stress so that is why if we have a system it can help us to understand that yes we are under stress and this can also help us to take care of it and not get into a problem in the future so the next important part is electrocardiogram so this one is a very simple Bible forget the wave on the top what is important here is this heart and how these ECG is made so being this portion over here the small or white portion which is which are pointing with my mouse is the sinoatrial node which has a pacemaker cells within it which actually starts P my heartbeat so when it says R we should start that is when B series comes on these are the electric charge which is fast and it comes here so when this is happening on your left atrial the basis electrical charge which is made by which this form of var wave comes in in the similar fashion when this stays like a blood is pumped to the ventricles and you can see the ventricle size the mass is so huge so anyway whatever signal is passing there will be very evident and we can see that okay this is what is happening and so all these waves are made by the contract contraction and expansion of these ventricles and it is of the heart when it comes fly so why have put these different kinds of heartbeat so because this is very important as we already saw in a previous slide that whenever person is is stressed sometimes the heartbeat starts increasing so if say that is a normal heartbeat so say you have 60 heartbeats in a minute when you're in stress it might happen that it goes faster and now you are 80 so that is how it changes in the same minute and when it's slow you can just have 40 so it's really very evident okay this is for different parts and you've seen the irregular heartbeat is no pattern but which can be matched or which you can see that the same pattern is happening again and again but there might be some irregular heartbeats where you might have patterns also so this really tells us something that if we want to depict which kind of stress level a person is undergoing and you want to analyze that you can actually use this as our tool and as a signal because first thing these are always unique so if you take a heartbeat of a person now and if you take it after 10 minutes or after 6 months or after your or 10 years it changes so in the machine learning field and on it's not very easy especially in the authentication by matrix P it's not easy to replicate it or to forge it so whenever we do a research on ECG there are two basic things that we look into one is non fiducial features which is nothing but this whole heartbeat as a whole and we use some statistical tools on it say like need me there are not any that we want but on the old days and the other one is the fiducial features which are these wave points the pqrst which is on the wave those are the production features so when we are working with reduce your features you can still use the statistics method but we use it only on those points to see if I want to have the p2p they show the p2p distance then I take the same so I take two consecutive waves and I seen a Pacific Indian me and that is how I got latest it's not a good question so is can you use only like it's the EKGs review to determine whether a given stress is you know not just single one so I come to that point at least five seconds of data is required because in five seconds of data you at least get five four heartbeats four to five heartbeats which is efficient or which is a itself okay to detect which level of stress are you doing because you need to know the patterns and if you Quincy so with just one hardware it's not evident that what is the heartbeat or how it is changing would it be possible to see people who have the same EKG reading but the undergoing different levels of stress lately so generally no two people can have the same ECG it's always different but again if if it is possible which are not heard but still if it is possible then it it is also possible that both of you are and the will different level of stresses that depends on the base ECG that we have collected from you so if your PC CG and you're currently sit is almost the same it means that you are not undergoing stress but for the other person why is the same ECG their phd CG is different it means that person is undergoing different kind of stress so that is why personalised analysis is really wavering okay so how do we do that this is an overall concept of how we can use machine learning so you get this ECG really extract each of us using each a mining then we use different classifiers and we are walk on them and see we compare them and see which one works best we can use this one because this walk was a part of a big walk so we just wanted to know the stepping stone and how we can go ahead so we tried all the possible ways and how big into it and using these classifiers we can make a personalized wireless system giving you a value of its detecting which level of stress on are you in low stress medium stress or high stress high it depends on the individual so that is why we say personalized or video because the low threshold for you might be a high treasure for someone else but again just as a general thing so it's like if I have to get to know your baseline that is very important so you just make you sit in a very relaxed environment so that you feel relaxed and then is calculate your normal ECG which we call a pace ECG and then based on that you see how it is to play at every point which helps us to determine me at which level of stress you are so part of research purpose we use this data and this data is a part of we take our dataset collected at MIT and in the media lab by Jennifer Healey for her PhD project so they had a lot of trials and this was done based on the if they started the experiment at the MIT garage from bathe the driver had to take a like a bath for the first right and then get on the road and then go to highway and City and come back so they had a lot of drivers but we just selected these ten drivers based on the data that they had many other drivers their data was either missing or they were corrupted and put of noise so we decided to just choose this ten triangles and here when you are seeing what we are doing is so in the actual data set they had other sensors also but we are just focusing on ECG for two purpose to reduce the computational time and because publication so for that if you can't just use one signal to detect it's the best thing because it's faster and more feasible so over your so what do you say that initial dress and we find the rest is considered as a low stress because you are in the setting where you have meditation song going on and you just sit and relax and this city one city to city 3 they are termed as high stress because at MIT when they go out there to hide our two roads which merge and they make a bottleneck which creates a lot of stress on the drivers to drive over there because at every second we have to stop and it's like very busy route and you might just also take like more than one our teachers clause that one ma not not even cut my leg some path fraction of that school and this highway 1 and highway 2 is considered as medium stress because all this cars tears and all these driving portions were done in a controlled environment so the drivers were told that on highway you have to go at this speed and you have to keep on right or left and walk accordingly so that is how they are it was instructed to them and it was done so City one city to city three instances are taken as high stress instances and they one have into its medium and initial and final rest is the low stress yes yeah this 151 right 0 5 sells like 1600 and so I got short time for mrs. Thatcher a 60 minutes yeah thank you okay so now we have the data and we need features right and because it was a basic walk so we decided to do a brute force matter we know we are using me arm we are using the fiducial points and I said QRS is the best one so if you remember the Ripper from yawn this P and this email they almost look similar right and if they is are if a person is in hydrostatic when C increases like a lot then this P and P wave tends to coincide and it's difficult to say which is the MST by this QRS is very separate and you can know if this is QRS so we just focus on the QRS complex of the ECG later so based on that we extracted these fourteen features and over the other this part happens it will be it is very important to our design because that is not any different speed it is the difference of your low stress and your medium stress your low stress and your high stress so that helps us to understand the stress level of an individual and that came out to be a very important feature in our whole experiment because the decision tree also first pick this one in the pre Iowa hierarchy so this was really important because this tells us a difference and how an individual's heartbeat is changing from time to time and this are extracted and QQ is like the distance between the are of the heartbeat one and hard of heart be two in that way it goes we used to wake up for people who are not familiar with wave heights like a tool there are which helps us in machine learning and they can play with the parameters and because it has all the machinery on mostly I won't say all mostly the only machine learning algorithms build in different sections and you can give any parameters and fix in the way you want it Omega you can do it jaw bones so it has two motion but yeah we use that and these were the classifiers that I use for the comparative study and the details of each classifier is mentioned in awake and as I said initial rest and final rest is low stress so yeah because we have to give classes names and also we had three classes a zero one and two later we move on to just two classes because in machine learning it has seen that it's easier for them to say yes or no which is binary classification rather than doing a multi-class classification even if it's just three classes so this is how we did the assessment we wanted in the world Lord Li Li were not cross validation that is what we what in general and reality happens so say a person is there and the system knows about any seconds and if they know the details so what happens after n plus 1 second where does that well whether the system is able to say whether that person is in stress or not so I am sick in second you can big ass for that instance that we are taking either five seconds or 10 seconds or 1 minute whatever that is so over there what happens then full if you have 10 instances or say 10 or data points it takes nine data points as training and the last one is users testing and this keeps on happening for each and every date and then it's averages and gives you the final answer so that was really important for us and we had to for 10 folds 90 split to save me like 90% of the old data is taking us training and the rest NSS testing I basically need to be done on a per individual basis right it cannot be done for the entire know patients because it's basically that I know this one main ability of the whole dataset as a rule so ignore data so that is what I'm saying the average different speed that we are having that feature that helps us to do so we are not setting a threshold as in over here to say that okay below this disaster this dose was not above this is high stress oh we know we have these data points right data says 1 X since 2 months is 3 4 5 6 right so it should be 5 not 6 so every driver say has 5 instance ok and you want to know if this instance is in highway 2 so this instance should go in a load find a medium stress so because your do your class right so we provide this whole data set to our system and then it classifies which instance is a part of which stress level training data yeah so this is my so when I'm doing a leave one out cross validation it takes all these data leaves this one out uses this as a testing data it trains on the rest see they're they're like and cow instances trains on the N minus one instance s with this and instance then it includes this one into a training one and use this one so in that way it leaves each and every one out and then keeps on texting okay talk about this week as cortisol levels and that's it again yeah so you mentioned at the very beginning that using individual yeah so that will be doing when we are doing authentication so so see they even than we're doing the whole way when it's personalizing you know which incense is going to be cut predicts and tells you that this instance was originally from this person and at this level now it has gone to this person and this level so wake up days will their output but if you do the prediction output so in that way you can match and see or else you can also there's a crawl on correlation matrix in or which you can see and see okay how it is happening but here the first thing that we wanted to do is where our system can actually classify stress level as separate but less stress levels so if it can do that then we can go to the next level yeah so when we did only three classes are the first eight features then we did not get a very good reason but it was not even bad because it's for some for like according to our knowledge this was like one of the very few studies where we just use ECG to detect stress because till did researchers said that they had to use some other sensors also with theory of it is easy to get on nice above 80% or 90% accuracy so with that also when we were getting like 85.7 and Percy were like okay we can do some kind but we need to work on it because the we are doing a very beautiful smell very nice method I'm using everything that we have and they have not done any feature like selection or any other process over you but the important thing was we had like zero false positive rate and zero false negative rate so I see poor decision tree which was really nice because if there is no false positive and false negative means there's something good that is happening and you can work on that okay so because you are not getting that good results in three classes and as you said machine learnings are much better for two classes because it loves to say yes or no whether then asking you to say maybe so that is why we did just two classes and in two class will be used of very nice thing we just took the low stress and the high stress so that they have two very distinct stresses and there's no medium stress which can't just create trouble for the machine to understand so we just took the low stress and the high stress and we feel and we will experiment the same set in the same experiment and now we saw that and itself just increase alone so we thought maybe there is some biasness or something that we need to look into because what is happening but then we saw that remember the important feature which I said the average different speed so that really helped us to understand what is happening because when there is no stress and mini on high stress your stress and we are looking at the frequency of the QRS in in a minute or something so in that way if in your normal heartbeat rate is 60 it goes to like 100 110 so it's very easy for the system to understand better this one is in the low stress or the high stress so because we are analysis said that average difference because really very important so this time we just said leave around all the other 13 features that we have just a one feature that is opposites and we didn't see what happens because every of these the most or theme I Gordon's they were first became this one this particular feature to start the classifier classifications and then we did this we actually saw that j14 is the same because they always used this feature to classify our to classify these stress levels but rest of the of classifiers also increase if you just compare this one with the previous one we see they're more like nine please just keep hundred-person it all kind of so that was really interesting for us and we asked different people who are much more experienced in this or experts in this domain to help us out why is this happening and they said because of this feature which surely helps them to understand and how numerically it can go so in that way it is really good MIT MIT data said all of these yes all of us hungry how many data points so for our walk we just had this solution we had 50 and then we divided them in 2 minutes 1 minute 10 minutes and they're different thing but I'm not explaining all those because practice believe in the people but here I'm just trying to but I can explain as in the fire yeah so yeah I see as I already discuss these points so for us it was very interesting point to a point because till did whenever they had even by getting low percentage of accuracy with 3 classes it was much better than what previous researchers said because most of the paper said that it can't be done only using only ECG and something of the atom so but then we we saw that it is possible to different levels of stresses and if we just speak to stress levels then it's really very good and the other reason to why exclude the highly data points and just take the lower and high mine was because in the data said to me a highly data points were not very nice and I personally contacted or Jennifer Healy to understand what they are doing and how it is done and I rendered PCs and I had a talk with them so they had some trouble in capturing the data points can be high B so that made sense why has such a lot of noise and by many data points are missing from that particular sections so yeah and it was also clear that yes you see ECG alone we can detect stress in a person but over here as we saw the data points like the whole data instance if I take just one instance it waited from seven minutes to 15 minutes as it was put it in the table we wanted to know how less time do we need to detect the stress in a person so for that the first thing so we did a temporal analysis for him so in this what we did we were looking again into a data so by this time I hadn't talked with or Jennifer Ehle and I knew what is happening and they've told that all these drivers which were there all of them are not unique drivers but in the first study but we did not require them to be unique we just wanted to know what is the stress level whether it's the same person or different person that doesn't matter to us and you know what we wanted to do is what is the minimum time required so we took five minutes of each section for five drivers whose data points for the best out of every other drivers that we had and just based based on the data points that we had the noise level and the outliers and all we decided the five drivers and I'll show you all the data said that we used and we divided them so we had five minutes for each so we had a total of twenty five data points a big data points we divided each five minutes dataset into 60 30 20 15 10 and five seconds and we did not move below 5 seconds because then they might just have one point or two point which will not make sense to under for an analysis so yeah and again over here because in the previous one we had got to know that machine is working better for to low and high so we thought just to know how less it happens for the same models and the same explaining method we stick to low and high and because anywhere highway or papers or not that actually that we can walk on them so we took like these tail points so those five through 10 is the minutes from five minutes to ten minute was the data that we took for that driver so all those four and we just need we kept the same name as it was in the dataset so that is easier for us to map and also check what is happening so we do we did that and again IRS ignition rest frac final rest C 1 C 2 C 3 Rd CP 162 and city 3 so the force and the last column represents the low stress and the medium one semi high stress so this is our new data set now but this is a subset of the additional data set so we are not children editor said at any point it's the same data set but we're just refining the data that we want based on the data and the quality of data present and how hard we saw was so in the presentation I am just putting the result of 2004 like 5 minutes and 5 seconds the rest we have put in this book Chatham so over here P so that our OC is the curve area term which was important so this work was done when I was in the static University and in the system log security lab roc plays a vital tool along with accuracy because accuracy can be biased a lot of time so it's area under the curve so when so whenever machine learning algorithm make some decision there's always you can find an area within what is happening like the false positive and false negative what is the area and how it is happening so based on that we get that ioc level over them and but I will double check a relative oh it's even what is an actual standpoint it's a really it's like a receiver I think at least yeah it's it's goes back to like yeah it's yeah it's sown in the audience it it goes to the original how illusionator the whole notion is user area under - yes but but they read it as I don't see because of the only store in town that is that but nowadays it is taken as area at the top there and that's really important in any system security system so yeah so this time we we taught it down the IROC and the accuracy percentage together because this is always useful for further analysis and see how hidden your data with their results are by or not and how accurate your day our results can be so yeah when we did this so we can see that the level of stress is detected and it's like a tip 0.8 percent accuracy which is the highest way that we are getting using j14 and our re but are you serious is important we are we are getting 10.7 4 to 1 so when we check on that after 0.75 it's a very nice ROC so it means you're in Tulsa metal but over here one mid is not important for us because when a person goes to a physician to get a check-up or get to know whether the person is in stress or not position leads one minute to check because they have their own way to see high order physical like you know action center so if a physician can do it in one minute then there is no point for the system to have there if he is doing in the one minute only so I'm just skipping the other results this is for 5 seconds and they see the 5 second is increased to 88 so yeah so it says that if we have one winner data point or we have 5 second data when the system is really able to classify where does the stress for instance if this coming is belongs to the no stress or high stress so and this then we also check manually for each instance how many data points they have and how it is working and before discarding the data points below 5 seconds we also try it on one or two of them and then we saw that it is just one or two data points and we went to or we went back to the literature to see why is it happening and based on all the based on the Veda a Hardwicke's generator and it power happens so we we came to a conclusion that yes we should just stick to 5 seconds and the explanation we have given in this chapter by we stick to 5 seconds and all so this is a comparison bow clock that we made so all these different powers are actually for the different time instances from one minute to 5 seconds so like 60 seconds 45 30 15 10 and 5 and P training set 1 for 2 holy temple we see that it every place the love this starts from 60 seconds and this is the five-second so at every test experiment we see the five seconds is either giving the same amount like the same accuracy level or a better accuracy so we can see that yes this is an increase in accuracy so initially remember we said that I'm taking five minutes of data for each individual neuron expensing so that five minutes is divided into different instances so one minute is divided upon minute now has to have instance so this five minutes if I am dividing in five seconds I'll have like almost 1000 something so that would so my mom is my excess yes five seconds it is enough because we were doing computer analysis so we divided the whole data then we had a 50 minute eight we divided them we just took five minute data of that 15 minute data and we divided that five minute data into five seconds each so I Han gives you like beginning of the end although this ones which I showed you so five minutes or so each of the classes yes from beach so five minutes from the Alicia that is prime minister up City one five minutes from city to five no subsidy three and five minutes from the hundred rest if they are these were chosen based on the theses and the information that we had so these are the outcomes that we thought that but if you see in both these experiments all the caches were not balanced blessings right and specially in the last one when we subdivided them into five second instances there was a very highly imbalanced classes at Lehigh because for each individual we had ten minutes of low stress and 15 minutes of high stress so when you divide the into five seconds the ratio of imbalance increases so we are assuming that if we do a feature selection of all the features that we have and if we can is balance these classes and then redo our experiments then maybe the results will have some effect it might increase that's our assumption but now now we know that yes stress can be analyzed and 5-second is enough to protect the level of stress in an individual also because we are just using the QRS complex of God ect does the computational time and the cost of computation decreases now the next thing is can we authenticate so authentication is a totally two thread research area where a lot of researchers have tried to authenticate an individual using ECG but in the starting of my love of my talk when I say that is it is unique and that is where it can be used as a biometric signal and that is why it's very important in the environment authentication research area because it's unique you can't duplicate it you can't replicate it moreover if were ECG is in your body if you are alive only then you can produce ECG signals so even when you have fingerprints or IDs or something it can still be used after a person is there right but ECG can't be done it can be deeply so it helps if a system can be built just using ECG and because he's really changes from stress - stress so it's more difficult to keep a track and do it but if a system can we just made using ECG proper education it will be almost full I still see almost foolproof because there are ways you can train to system there's always these net lending to systems so yeah so we wanted to see that now now this was the model so here you see I create formula we know that side of it is node and everything so this is the data that is reading from this heart rate so this is the actual data of five seconds from the data set that we have that no data its annotated teachers are extracted and fix your selector the selector features are the annotated feature sometimes either one or both goes to a machine learning model this modeling decides this model then decides whether this person whose is it is coming is the same person or not it's basically one thing which Aquaman and is the conversion of the waves then they'd act like basically the pictures I was like this outfit condensing the fight with it for like weeks I'm a compulsive eater how do you figure out those shows so this data was available on physio net it's a biggest of physiological signals collection and it has and you can use sequin and all what do you say syntax in segment two annual convert these data into annotation forms so see at one minute five seconds sorry at our one minute five and second ten the user heartbeat was showing was at cue okay cue of the QRS complex so it shows that the Open bracket MV file when it ended this is a notice in that and then if it is our it a note it has N and when it is s dramatizes close parentheses okay and then there are the columns where you can write and you can know it's say that okay this is our this is Q this is what happened this is how it was taken you can do that so they have this whole system built which you can use and I know tape these signals in the way you want there was just you care of that one so I personally you had a slide right that's why they have a whole bunch of different sort of things yeah but those features were extracted from the annotated or signal right five seconds or 10 seconds of the whole earth processor as in whole so we first annotated that signal so if I were to say earlier so that raw signal which comes to us because we are just extracting curious which I showed in the figure I can raus photographers yeah so see those city signals are the raw signals right we then annotate them so do you have this whole customized program written for it which uses these door signals and conversely our negative bias the way you want and then outputs they have into the file the way you wanted an attorney so like you can take the word signal in cubes and to be you like the way I wrote it was it was automatic it took like 10 10 minutes data and it kept on separating so the first type we like the first one minute we used to go in one fight the next one with if we create new files where it keeps not so what that happens automatically but you need to understand the whole sequin at the good is unit and you see which command you want to use exactly and how you want it so like these people have actually worked a lot and within this Visio net and they were only two week eight available whatever enough is enough and from those annotated signals that is when we extract all these features you have a question Wow so we get unrest in the Pakistan what a shape signature work the situate there's a another so what this is doing it's pulling out features that are characteristic the waveform but they're intervals different intervals yes um but any given waveform hmm you can represent that as a polynomial yes with that this menu using a non-traditional feature so now I said you have two ways fiducial features and non-traditional features okay so then you're doing a polynomial then you're taking this whole way yes and expressing it into a polynomial P or you are you're taking some statistical matter right but I'm using fiducial meth future might notice when I'm just using these points right right shortcut so then I have this okay so this is Q RS q RS right so this typically this and this so if you compare these on graphs over on you see this distance is increasing or decreasing right right Shawn Shawn I understand if I guess what where I was kind of in my mind try to which would be are there any advantages or disadvantages to particulars represented yes oh yeah so resources have done in both the ways the best thing that they have sought sauce or till date is come combination of good oh but whenever they have done this it's almost like they have other signals also but because he was just starting and it was a knife but I'd like to do it the way we will understanding so we just taught to go in one way first and see how it is booked so that we can take further decisions in the same any other question I'm almost in the last part I knows a lot of information where I'm trying to give as much high level and make it clear but please ask me questions so before that just giving an information about authentication means so there is like we all use our phones every day right and why this was important is we know that our phones get hacked every time we use our laptops that every technological institute gets hacked now and then but if those don't use password or don't use any thumb print or something but uses our ECG as an authentication process it means our life easier because we know no one else can get into a system and it's protected 99.9 percent it's protected and we can still work on it being safe and you can just keep it anywhere like there many people who don't want to keep the phone away from them because they think someone else will take it and they'll hack into the system right but this is because it's coming from your own body you know it will just take that so that is what is happening we are given the Roy City data and then if the machinery Caesar yes you're the same person now I'm going to a detail of authentication because it's in a whole area itself where you can do identification of verification so I just say whether that person is the genuine person or not see Amy have an eye on my laptop so if the system is played in my laptop it can see a better the person working on the is Neha or not okay and because we know stress level changes and ECT changes with P stress level so can the system still authenticate me as Neha when it's trained on one stress level and testing on something else so oh yeah I just say what was the result that we had in this so we took the same same fight people over there and we did it so one minute instances and five minute this is the confusion madness which I was talking about so just all the odd confusion matrix is conditioned matrix takes the individuals that we have on left and on the top and then if so okay so over here if I have a one it means and this is a and if said this was also a then it means yes it has accurately found the person to be the same person but if it is not there it means it has misclassified that person has someone else and as food we know by this so see over here this was a it but it was classified as a it means that three instances from user a was identified as user H so here is where we come as false negative and false positive comes into play so this really helps us to know and the so in via command uses two like predictions and all it gives you exactly maps it like which instance was map wrong and then you can go back and analyze and see why did that happen so you can work on it so over your brain you're looking at identification and all what we see is this is an individual how many instances did it had and how many was Ekta car was correctly classified as that individuals instance okay so in battery you know that okay it is identified or not so that's one part about it it's very identifying that person as the same person based on the instance or the data point supplied to the system so the probability of identifying the individual for us was oh and we use ten subjects why not five was ninety nine so it was really and really important and we use the same analysis method as it was done my very first researchers can be your fear and the method which was accepted by the whole community so we use the same method to come up with this 99 probability method and we also saw that because then we see one minute instance and final instance like this time duration doesn't play a major role in identifying a person so these were the two outcomes the results that we had this is just so lots of replication was a part of the poster that I presented in the next conference at Syracuse University just because it's not available online like this so I just wanted to show that and yeah thank you while you resolve our offices swaddle cortisol levels and staff back oh the adrenaline cortisol yeah I just said that to show that how stress is affecting in this and which my brother was in one of her classes and also over the side did some work with his measuring and partly measuring the stress levels and the physiological effects of different kinds of air treatment plant wall versus conventional meet you at E versus night and part of the data in addition to the usual physiological symptoms being collected was that the cortisol couples stop they purposely that was that was actually the indicator of stress there was they were observing yeah on the heart rate effects of co2 and all that I understand or do see my dad said differently such that had that takes place and they have been research done on that just based on the cortisol level in a human person because the more you are stressed or thing you have cortisol level and adrenaline hormones in her body but I took that into just a mere conditional because when we are stressed these hormones are secreted which increases heartbeat so and that is a main reason why we are using ECG because that is directly related to the heartbeat right I guess what I was saying I I would have thought it could have been used as part of your confirmation that you were indeed getting light indeed because because it's where you think you think that the heart increased heart rate is an increase in stress this is it would be more possible it sounds like just the categorization when things should be a measure that was kind of a positive was we know that this is this city city driver city driving these are our representations of stress right so which is the interesting way to do it well I also understand that the purpose was to be able to figure this out just it was always great job and it tried to give like I could have gone into quantities because I can throw on this like four eyes and I have no problems with that but I just tried to incise myself and try to put some important points but I'm happy to answer and go any detail any area yeah like that day I did just because I wanted to make a connection between stress and why I'm using ECG because those are really very important thing and why those heartbeat changes so to make that clear that yes that is really very important and can be a big factor to detect which level of stress they are because that increases only when these hormones are secreted so question so on I'm trying to forgive ever I'm trying to envision the practical application of this okay so let's use that example my laptop is going to be easy T is that so so there's got to be a face when it's going to learn to learn me and then there's and then it's got to learn me with sufficient generality that yes or I can come back so that John you gotta come back when you're not stressed yeah so that was one thing you know what people did what we started this we were thinking like a person isn't driving because we are drivers database so we'll again see a person is driving and we know that if that person is in high stress if he listens to this music then the person stress level comes down so if a person is ready instead of saying to the person of your stress the system just gives an under to the music system that played this song and the strong stands playing so in that way the stress so it was a really very automated thing and so I have some unpublished work which is under review now which is working and when we did much more on the authentication part and the whole degree of easy and everything so yeah these are looking well know that if we're in meetings in here and all of a sudden comes I stress it means have you ever had a music it would bring down your stress to go on the ring tone ring studies the pizza selfie dream do you know other questions