[00:00:05]
>> Cited you have a doctor then you can shout from my D.M. research and visit us again here for the best lecture on the planning for health care applications and. Doctors showing this research staff member at I.B.M. Research Research Center focus on AI for House care. C.S. published many papers in fact if you look at her publication lists she is a super productive she already published 2 papers this year this is just January 2080 you can imagine haha what's the level of productivity she.

[00:00:49]
And she also is an expert in. Technology since this is also a horse for. Just about she before all this your research she actually worked developing a pipeline to link you. Off to covering this course. Of action. But today she was talking. House OK So let's welcome her so.

[00:01:30]
Good afternoon everyone it's my honor to be here and share my experience. On the topic of the planning for health care and I will talk about opportunities that we are seeing and the innovations that we have made a recent today and also to mention a few critical challenges that we are still facing and hopefully we can conquer that in the near future and 1st thanks Jim for a quick introduction and that's my trajectory for the recent years I got my Ph D. from University of Washington.

[00:02:08]
In December 2016 and after graduation joy I.B.M. Research T.J. was a research center in the world called health knowledge X. That's the 1st like all right January January 2007 thing that's the 1st time I ever got in touch into the topics of the planning for health care and then later that year I wrote okayed to the new I.B.M. Research AI So we are mainly will call for health care specifically mainly people earning for health care with health care pick patients.

[00:02:48]
So. My talk today will be structured follows We 1st of all we will talk about the alternatives that we are seeing right now and then I will elaborate in introduce our recent innovations and merely talk about our most recent 2 publications. Then quickly overview of your challenges in this area so we are seeing over the past a decade we are seeing a growing availability of health data.

[00:03:29]
If you guys can see this bigger clearly this is our Savior borrowed from a wrist into Nature Medicine publication So give us overview of how many different health care data model ality we have been collected during the past decade so we see that these data come from multiple sources in the Manning different model it is in the pain patient information and this information from various aspects so the availability of these health data they provide us a good option other t's to innovate and ultimatum Manning traditional health and knowledge experience with this the help of this data.

[00:04:16]
And meanwhile we also the past a decade we also win this huge advances in machine learning to knowledge ease especially deep learning techniques so here I show 4 different the printing architectures the 1st one is the convolution on your networks. May advantage of convolutional in your networks is it can extract a higher level feature from high dimensional raw data so this good.

[00:04:51]
This cool feature of. The convolution on your networks makes makes did the best model for a lot of computer vision occasion so we have the Syrian transform a lot of vision computer vision have and. The right has either we also have the recurrent neural networks. The recurring neural networks have been hugely applied.

[00:05:19]
Many in natural language processing tasks and we also see a lot of reward at the Acacias Google translation. Other automated automating automatically generate summary writing report is kind of the new innovation and. We also have another type of model which is the new newer model compare with Ian and Oren which are the generator of a serial network so with this that with the help of.

[00:05:56]
This this of us thorough networks because of them we can train data we can train model more robustly so we can learn our motto that can help us distinguish more from these diverse a diverse array situations and the last one. We are seeing is the knowing altering coder so these outlying coterie come from a different family different from the other models such a C.N.N. And aren't the outer encouraged can be trained without getting an ending training label so we have a lot of data we don't have labels so we can apply altering code or with the noising altering code or we train with a corrupted input so the model we get will be more robust these models have chance formed a lot of Haitian area so we just mentioned computer vision or natural language processing so we hope play with the same set of techniques we can also transform healthcare staycation because we are having a huge amount of health data so if we apply and develop novo techniques healthcare placation we hopefully we can see more innovations and more promising results in the area.

[00:07:24]
This is. Soaring. So why we need the steep learning to power health care data and empower patients. In the past we are training doctors and clinical specialist to help us perform these healthcare task for example we give this health care data like this is the medical imaging we get.

[00:08:00]
Report from. Patients and also these physiological the newer recordings were that if they were to doctor and the doctor had training in training themselves for more than 10 years and a day trained themselves and based on their experience they help do the monitoring task monitoring patient status and that and also has to see whether the patient has for example heart failure that is all ending other disease and to prescribe ending prevention guide a life so this is what usually a doctor will do but when we train or train our doctors so the issue is Dr we need a long time and we don't have so many adulterous and though more issue is each doctor they are subjective so doctors state health disagree with each other how about if we can train machines we can apply this learning techniques to train model and with a huge amount of data we probably came to least the performance of the doctors or maybe like much better than doctors can do so that's what we hope.

[00:09:17]
Here feel realisation that what we currently have this is also a figure from a recent a Nature Medicine publication so we can see actually there are many nowadays there are many. Health care pick Haitian already adopted machine learning all the printing techniques to make to automate a lot of traditional tasks and to be more accurate than a single doctor can do.

[00:09:47]
So that's all these opportunities we are seeing and if what I can say is this is a new area health care so anyone who commit to this area and accountability and if continue years doing this will be pacified or of this area in the future so it's a very promising area right now.

[00:10:15]
So I'm personally very fascinated about this topic and during the past. One year and a half we have done a lot of work in this area so basically we try to perform patient monitoring this diagnosis and Prevention kind lie generation based on different type of health care data and we have published a lot of papers and today I'm going to in the following I'm going to share some of this works to guys.

[00:10:51]
So before talking about these works I'm going to talk about foundational task in the planning for health care this is called a medical imbedding So it's an important building block for. Health Care So imagine you have so many data about how to empower these how to power these data how to leverage them you need to if they don't they're.

[00:11:17]
Noisy they come from different more than it is they are very high dimensional so you need to use us to me you need to have sophisticated a method to embed these data raw data into a lot of dimensional like feature space and then with this low dimensional feature imbedding you.

[00:11:43]
Achieve all the task that you can perform or the task we just mentioned So that's why magical investing is important building block for the all these exciting occasions we just talk about. The definition this is not of are accurate as a nation but basically medical imbedding takes health data input and output vectors of medical concepts we just mention the whole concept that can be.

[00:12:13]
A concept that are over the east of actors that represent a patient so for you with this concept we can can be used in predictive health care. A lot of our work actually focused on your hard data so I'm not sure like how many of you have experience with the electronic health care records data.

[00:12:40]
That actually health care records records data have been corrected also hugely during the past decade or so everyone's like you go to a hospital you see a doctor and there are your records will be recorded in the system so you have for them whole last year if you have 10 visit for each visit you are being diagnosed sister with a few conditions and you are given you're doing you did a few lab test the For Each of it and you also prescribe the with the medications so all these various All these events the wealthy sure show up in your record so you have a sequence a sequence longitudinal sequence of record and each time pointed show one of your hospital visit and therefore each of the of it or your records or your all these medical clinic or activities will be there so this is the H.R. So our committee focus on getting your hard data to perform all these Stevie's these studies in the service medication recommendation and other risk of predation task so we just talk about the definition of medical in batting so we take health data like Broadway concept health data but now we are Consider take patient history represented a by all these medical costs of medication diagnosis and the procedures input and the We also to practice all patients back to our medical concept that can be used in predictive House care so that's.

[00:14:27]
Where we position our research. So let's take a. Take a more. Close look into the 3 components the 1st component is data we're just talking about your chart. That year chart data come from like longitude I know you tried it you have multiple of it and each of it you have multi-core Mary Coco's So you actually we can we can to our heart in coding on about how to in coding for these medical codes and.

[00:15:12]
Here is a simple story this is simple visualization of how you in code draw your H.R. data and these E.H.R. data will be used for learning medical in that ng and how to learn these medical embody since we are talking about the planning for health care so we consider using a deep when you're in a talks to learn about her in that ring of patient all these but actually you can also this is you can also using other types of machine learning models of your choice.

[00:15:56]
So the task the 3rd task is how do we leverage how do we leverage the vector of representation we learned so with this representation we can do sequential prediction to see whether a patient well. Whether a patient will have heart failure on that during a year. Whether a patient classifications whether a patient is that it is patients or do class 3 for example we passed a patient in 2 different today these types.

[00:16:35]
So that's a major task of how we leveraging this medical in that it. And a toy example the 1st step like we just mentioned we feed the data into neural networks. We just talk about the. Audi if I see D I see the code I think the code is the this diagnosis code code is the procedure code likely you're doing a lab test of the procedure and our US Code is a medication called.

[00:17:15]
Like aspirating or other kind of medication and we embed the east into one hot in coding and then we input this. Data into a neural network so here is the same whole banner a classification task with the input data we just we just see like here in our goal is to detect whether given the importance given given the history of this patient whether this patient has onset of heart failure so this is the prediction target and our.

[00:17:56]
What. All tool for this detection it is multiple hidden layer here we will simply demonstrate. Multilayer perceptual. So given this imported. We transform we simply transform this input for prediction and we apply and activate activation function to generate and output the out the outcome is of course where the score ranging from $0.00 to $1.00 indicating the probability of that is these are set and this same home although well be efficiently trained with backpropagation so not sure if you guys covered this in class but OK so yeah you learned later in the class and but this is just I think he did a poll where we simply.

[00:19:02]
Embed all concatenate concatenate these medical codes from each of it together and without considering their temporal information so however the chart did are like we just mention it as sequential Lyda sequential data so you have multiple evaded and each bit you have this code if we just do such a simple treatment without considering the longitudinal relation like this event happened 1st then the other the events happen later so this order and temporal relation really matters in deciding whether the patient has some kind of condition so if we do the simple treatment who lose this information how do we cover how do we model this capture this relation.

[00:19:56]
In that case we will need the recurring neural networks the one we just mentioned at the beginning of this talk so it shows it's a powerful model for modeling. Natural language or a processing task so hear what you. What's the difference between your occur and your networks in the same poll feeder for the neural networks like the one we just show which is so this is the 1st of it that's 13 minus one tons of it this is the current bit so when you learn the imbedding learn the car embodying for the current time hidden state you're not only considered the current input which is this part we also consider the previous time hidden state which we learned from all the previous.

[00:20:53]
Bit So it's like you not only consider your current But you also consider all your previous. Status so by the way we tried to we actually capture all these temporal information and then we used by knowing embodying to perform these diagnosis and here is the. Procedure for doing that so this is your 1st observation right like we have we have.

[00:21:32]
The 1st bit so we have 1st the opposite of something from your 1st of it and then we'll learn imbedding given your 1st information collected from patients 1st of it and then we open serve one more patient visit. And with the with the model I just mentioned we not only consider the new observation from the 2nd of it but also the embedding the latent representation we learn from we learn from 1st of it so so in the so forth we always serve 10 total bit for the patient and learn our final report.

[00:22:24]
Presentation and based on the final representation we perform transformation and predict whether the patient has sounded nice so this is how we can do. A simple example for to illustrate how we use the patient you chart data to perform these diagnosis task of course don't recur in your networks and feel for neural networks in the only model template that we can use to model health data including your chart data we can also apply the CEA and I'll tell you in quarter they also they have their own benefit compare with our model for example for this year models like we just mentioned it can help extract the high level features also this model has parameter sharing so the model that.

[00:23:30]
Is rightly a bit lighter weight than to our model and. That the noise in altering quarter motto If we learn his European nation using auto included then we no longer need labels for the this data and it's also robust against noise in patients. However there are. A lot of works had been down in the area of applying aren't altering code or to learn patient representation medical imbedding and perform these diagnosis the risk of prediction however there are still you some challenges remain which is how to consider the specific challenge of multi Mado imbedding that's the focus of this topic and.

[00:24:30]
The 2 recent publication we all try to address some of the challenges. Some way and. For multi model medical in that there are a few concrete challenges the 1st one is. These patient data come from small time of arity. Some of them are genius typed data for example one is come from policy where President human knowledge about the relation between.

[00:25:06]
Between some medical entity for example the relation between these and the medication like you have is someone that got this disease so should be prescribed with some medications so it is kind of knowledge and some data come from like the year chart data the questions that are just show you guys and so they have had a genius type How do we incorporate these data together and to make.

[00:25:35]
Empower like each other and another challenges another a lot of. Internal structures within these health data especially your trial data were not exploited. For example. This year trying to figure I'd just show most the previous work consider diagnosis procedure and treatment to consider these calls from the same moderately so they are given the homogeneous treatment when modeling them like nothing they did they do not have any difference from each other but this is not.

[00:26:15]
Accurate So how do we leverage and model these internal structure of the data that's another challenge away are facing and the 3rd one is for Between across different model it is. Like one provider one view of information and another moderate you provide a complementary information and how do we feel those 2 types of information together to make complement each other so this is the 3rd challenge we are still facing and in the following I'm going to share a few recent works how we try to address these challenges from.

[00:27:01]
One way or someway So this is some of our so my. Last year's publication the 1st to go in to talk about today but. The 1st one the 1st 2 are about heterogeneous type health data so the 1st one multi view. I'll tell you coder for integrated drugs similarity is 2.

[00:27:40]
Use graph convolutional in your networks to construct a graph model to model the relationship between different drugs different medications and the graph each node is one medication and connecting each medication there are similarities and this graph the graph is attributed a graph which means for each node the node A has a list of features and the R.T. features that's the key for this work so this is a multi view which means we have multi mortalities to all feature and for these features for example one view of feature can be the drag indication and another view of the feature can be the a diverse drug reaction of the future or the chemical structure or the features like like different the we have around the 6 or 7 different deals so how do we integrate this information take Adderall and the puter improve the drug drug network and after we build to improve the truck network we can learn improve the imbedding for each drug like and then with this improved embedding we can make prediction for new drug properties like we can predict new indication new drug or drug adverse drug reaction or new drug drug interaction for each a drug or so that's what we do with this paper if you guys are interested you can go to my website and get the link for this paper and the 2nd the one is heterogeneous hyper network publishing.

[00:29:24]
2018 so this one we just talk about model medical data so for this data we also Bude a large graph but different from other conventional modeling techniques. We consider the note in this graph can come from different more data for example we have patient the one node we also have a medication than the other node or a 3rd type of know this so we build them together and by perform a traditional task called the link protection we can make a prediction whether a patient who will have a day these some days east where there are a drought can cure a new type of these we can perform that kind of task so this is how we how we modeled the head of genius type the health status of the 1st challenge I just mentioned if you guys are interested in this work and go to my website I'm not going to cover the paper here because it's to the mass the I'm a specific there are a lot of knowledge about drug medication I.

[00:30:37]
Don't think it will be good torque here so. You know what I'm going to introduce. The other 2 paper. The one paper is for. Paper so it's called a mini we viewed we leverage our internal structure of medical code thing each our data and the we learn more accurate medical embody and the with this embodying we also learn these things bearing on supervised fashion so with this embodying we can learn we can make this prediction with limited training data and so another talk I am going to share today is called a game at which we will pretend that Interpol AI next week and for this one we are providing a personalized and the Safe Medication recommendation so imagine like you give me you you gave me your health record and you today you visit a clinic in the U.K. me your health record and.

[00:31:51]
Also diagnosis today you thought you were having some kind of condition and how do I prescribe medication to you which is both acrid and also if they if you will not see anything kind of allergy or any kind of drug drug or reaction because you're also taking other type of medication so this is the thing this paper is about.

[00:32:15]
OK so. First going to briefly introduce our own it's peeper So this paper is. An operation with your mom and dad over there is a previous student. Who is now works in Google praying and. In another carburetor in solid house so this paper is about multi-level medical imbedding of your chart data for predictive health care.

[00:32:53]
Now we see a similar a similar figure here remember the last time you see these things are we have all these diagnosis treatment these kind of cults stuck together but there is no we treat them homogeneously there is is no. Different treatment for different types of codes but for this paper we always serve internal structure among these codes which is for each visit the diagnosis is firstly prescribed and then your medication.

[00:33:31]
And other treatment will be it will be based on the diagnosis your given. So does some kind of causal relation among this data so we show this causal relation using arrow here and go in this the 1st to go in this paper is it to embed this internal multilevel structure to capture the diagnosis in the treatment relation and to model patient status more accurately and the 2nd way.

[00:34:11]
We also for the for health care placation that we also often have limited the training data however to 40 planning to be super power for we need a lot of 20 there. Which is not easy to get in health care so how do we conquer this challenge how do we train the motto to be more accurate with limited the training data so this is another go in this paper we try to training this model with auxiliary task this is like multitask learning we try to try to train multitask together to help with this insufficient data issue and the 3rd thing is how do we change is also providing it.

[00:35:02]
Why we do want to do that also provide a way because if we trying imbedding with respect to one particular training task for example predict a while predicting how thriller then. Learning bedding will be biased toward a task thought we are not sure whether the learning that thing can be used to.

[00:35:27]
Get. Me can be used to apply on the other task. So to achieve the 1st goal of this paper we embed these multi-level structure the multi-level code in. In that fashion for the ball or level is the treatment we have lots of medication. And procedures here this is the procedure and this is all this medication.

[00:36:05]
And the next level as an officer level diagnosis a level because I got a cough today so I would prescribe a deal with these medications so you can see there's a relation multilevel relation between these codes. So we 1st so we learn these things learning batting for these treatment and then we learn imbedding and we take apart and out fashion and the learning batting for the diagnosis and then we go one level up to learn imbedding for this a bit and then we tell take one more level to learn imbedding for this patient and in terms of our formulation.

[00:36:54]
This is it so we 1st to model all these treatment with like this is that the or this the diagnosis and did this. Medication. The I and this is all the medication for with respect to is that nurses so we model all the treatment to retrieve this treatment a level imbedding and then we go wild level up to model that diagnosis we not only capture all the technology this is self with activated but also tried to capture the interaction the touring diagnosis and the medication and then we go one more level we try to capture the interaction with Think of all the stock Narcisse within one visit.

[00:37:47]
And then we summarize one patient if a patient has attended it there with summarised representation for of it. For a patient to summarize all the patients of it using some function for them PO a simple summation all a more sophisticated. To thumb rides in the learn patient patient in bed.

[00:38:13]
And after learning this in varying we join training with all the theory task. The idea behind it is so we try to predict the diagnosis and the treatment for. For the current for the next visit given the previous given for this visit given the learning batting for of this visit and.

[00:38:44]
And idea behind that is multi task asserting is an effective embedding. And effective embedding should be able to of this village of this object can be should be able to predict what it can represents. The diagnosis and the medication so. So we try to predict them together and with this joint training using the combined across entropy laws we actually got a lot of performance cake compared with outer space lies and our math or the we tested sort of health data and we tried to predict the 1st diagnosis of heart failure given over the vision from 18 month with Asian observation from patient records.

[00:39:44]
All the if the patients in this state are said to merely senior patient age they're from 50 years old to 85 years of age and this is how we the results is we in terms of prediction accuracy we actually achieve 15 percent performance K. in P.R. you see which is crazy metrics Pervis state of the art and the we also have good prediction ability and insufficient data scenario so this is a figure showing when we have more data then the performance gap between our method and our method is this between our method and the base lies become smaller but when we have insufficient training data.

[00:40:38]
Our advantage is higher so. That completes our introduction to to the multi-level medical imbedding paper and. Now I'm going to share our Chipotle 2900 paper this paper is about a different topic how to ultimate medication recommendation with him out how modern information through should. The main flow for this paper is the 2 we want to give in patients who are called impute Serwer is my answer so we want to.

[00:41:31]
The cursor Sure OK So we want to give you the patient record of the input and the we try to construct a memory bank a dynamic memory bank so the memory bank consists of 2 components there 1st a component is a call memory bank within the memory bank we have a graph we have a graph augmented in full component this graphic compile all the men to the component has to graph one graph is called the H.R. graph since we are trying to predict perform medication recommendation is special a for these patients with complex conditions the most they will take a lot of medication.

[00:42:20]
Together so for in that case we try to decide which medication of them taken together like combined together so this is a trial graph which is the 1st the little graphic show like the left hand side figure so the year trigraph is derived from your chart data it was used initially Mado the graph with a one sided.

[00:42:51]
One side. And the other side. No 11 side. Medication and the other side and the other type of medication and who will try to decide whether the 2 medications are used to be combined together and the other half they are called are. Of the D.T.I. illicit drug drug interaction is a very severe condition that occur very often which is take I take medication A and I don't that's normal I don't see ending condition but if I take another drug to take Attard then I will feel comfortable and see a lot of adverse drug reaction so that's D.D.I. So we want to.

[00:43:42]
The D.T.I. graph is constructed the pace of domain knowledge. G. 5th so we want to come by this to graph together so that we learn and improve the medication imbedding this improve the medication imbedding not only shows what kind of combination it can happen. For this medication to comply with other medication but also try to avoid the case of occurrence of a T.D.I. so does what will happen in the memory back and for the dynamic memory we want to learn in period.

[00:44:22]
Medication history using the patient's own your chart data and this dynamic memory will be done that may kill a updated which means if we if a patient has a more hospitals' visit then this part of memory will be updated to reflect the change of the patient's health status and then we go to memory will be concatenated together with that improved.

[00:44:51]
Memory information and do we use the patient's current patient representation of Corey we can retrieve you to get a recommendation of a 3rd of medication which are safe present alive and the person alive the 2 this patient's health condition so that overview of what we're doing in this paper and this paper is.

[00:45:18]
Also in cooperation with and. In turn in picking University and. So does the task I just talk about now I want to define the task again so medication become ideation we've given. We've given patient history Audi's code. Of the training data and that we try to learn embodying and then given and then based on based on the actually there should be we should i should try another small block to show we obviously are what we are with serve from the patients current a bit about the diagnosis so given the patients current the diagnosis or kind of by occasion we should prescribe for this current of it so that's the task of by the creation recommendation.

[00:46:29]
But occasional commendation the accuracy of the one conserve the jungle interaction is another concern that's the thing I just mentioned you see about 15 percent of US population affected by a worst drought drug interaction it's also very costly like $177000000000.00 of dollars per year in this management only for the D.D.I. issue so that's why it's a huge concern we need to we need to consider and explicitly motto in this in this work and the challenge we tried to address some arrived here one is dependent companies dependency between the least in America and.

[00:47:18]
How to personal lives to patient history how to consider this issue. We this is the existing work some of the existing work also from Jim Scroope like the 1st model the 1st model. Like the 1st model leap. A publication of chemo from 2017 for the and also to most work in 2016 However all these various work they all of them considered a complex dependency between this and the medication but only one work which are the deep can consider the D.D.I. issue.

[00:48:09]
But the leap this paper is not personalized so for different patients the recommendation will be the thing so that's also something we've tried to address in this work and for this work we try to model with complex dependency try to avoid drug trial interaction and try to be highly personalized.

[00:48:36]
This this is the 1st more of the game that the graph augmented the memory network so we have 3 components the 1st. The 1st component is the medical imbedding and the patient representation modular and the 2nd component is the graph augmented the memory module and the 3rd component or weeper for training and protection which we call inference but it's the same protection so the for the 1st the border we simply input.

[00:49:14]
Multi hot vector of medical calls like the ones we just seeing again and again and then we transform these we were transformed is that back to her of medical codes with. With the in that ing matrix. And then we're after the transformation we apply one recurring your network to the diagnosis and the one to the procedure code separately and we learn to embed it and then we can cut in it is the embedding and the transform with transformation function so for our work this earth transform mission function we simply use our one layer feeder for the neural networks but you can but actually it can be flexible to use other type of transformation function.

[00:50:16]
So we can calculate it out as an office in bedding and procedure in bedding and or transform them to receive the patients who have then Taishan this is the output for this medical imbedding and the patient presentation module or after we all put this we start to input the output and to the 2nd module the 2nd modular is called a graph augmented a memory network memory module so this is a memory network that is a very famous model from.

[00:50:51]
Which is cited here you guys can take a look if interested and. We also adopt the 4 components from the memory networks paper so which is the input component there are generalization and output and responds the 4 components so here are more details of the 4 components. For the 1st import of components we import patient to repair the station of the quarry So imagine you are building a large dictionary so to reach your information from the dictionary you need a quarry and then if you have a corridor you can rejuvenated the information so we use the patient for a representation of Cory So this pressure representation is the output of that patient representation module the 1st modular so we use this one.

[00:51:47]
And then for the generalization module. We have a graph walk a mentor the memory. Excuse me memory module which has which comprises all of the memory bank more memory bank component and that dynamic memory component which I just described at the beginning and the one thing to talk about which I didn't mention is here to.

[00:52:24]
Feel with the year. Your child graph and idiotic worth get are we apply our graph cover Lucian or network which is very famous for popular model right now for. Forth you've Matamata information together like to learn improve the inventing So we apply a G.C. And here to learn improve the embedding for memory that.

[00:52:57]
This dynamic a member of. This. Dynamic memory even learned from. If they learn from patient history. Up to up to the ton up to the time of. Patient last visit because we're trying to make prediction for this bit and the output for the output we are just given a patient record and the will perform attention on member retrieval by attentional ritual which will remain.

[00:53:37]
The dynamic memory component. For the spiritual information front and I make a memory will put a higher weight to a more recent part of memory so we will consider that as more important but earlier history less important so we've performed attentional memory retrieval and out who is this recommendation to get a 3rd of medication and the Here is the training and the inference was worth noting for the training is.

[00:54:14]
We have 3 part of last the 1st is a simple plan or a banner across the entropy loss and the 2nd one we use a multi label prediction last we while we add a multi-level prediction was that we try to sharpen the probability or we don't want to see all these predict probability like the probability for this drug to be in the recommendation is poised by one or the other 8.52 we don't want to see all the probability of near 30 Hoda through try to sharpen this probability that's why we include our multi label prediction last year we also did a lot of to control the rate of the so and then we make a prediction so here we simply 3rd point 5.

[00:55:05]
30 showed. And that's all we do this is a recap of the 1st page of this work. And though we conducted experiments using patient record data from makes $3.00 and $4.00 the D.T.I. information it's trotted out from the go the standard the T.D.I. knowledge base called the 2 sides the link here and there we only pick the top 40 severe D.D.I. is.

[00:55:38]
In fact that there are a lot more D.D.'s thumb the idea is actually very mild and so it can be ignored in practice and though we consider include all patients with the modern one they did because we need to ensure the patient has to prove it with each our history so we include the patient with more but one more they want it or other patients with only one bit our job to from the training data and they will try to predict medication during their During the 1st $24.00 hours of the patient for the mission because that's the most critical time of prescribe medication and this is the performance from this performance we can see that we achieve much better accuracy than all the base lies we also can you see the 1st the metric the 1st a metric a record the D.D.I. rate change so they're all means they're grown the truth because the training day that actually included some of that idea is in doctor's take him prescribe a drug to so the truth.

[00:56:54]
So we actually can reduce care identified is existing D.D.I. can reduce that rate from the ground truth. This is the author Marie of this paper we. Desire and to end a different model to generate effective in the safer accommodation of medication recommendation. We integrate we're just talking about Malta moto medical embody here we're trying to integrate more time all the information one is real world evidence from the chart data and the other information is the knowledge for our knowledge base so we try to integrate the 2 type of information to make more accurate information prediction and we also.

[00:57:48]
Dynamic memory of 4 that I made in the present are either medication or recommendation so. That completes or if gosh I'm recent to work and however so we actually have done a lot of other exciting work in this area however direst it there are still many challenges we are still seeing for building the printing models of for health care pick a ship so we compare the planning with the human learning we still see some challenges some gap that human can do but a model at the printing techniques cannot do the 1st one is adaptive which means how you can be able to generalize by open serving fill some holes.

[00:58:40]
When we open so him a learning even a child you give the child a teddy bear and you bring the child the to the zoo and the child the cam point to a real bear you give a prompt had it there they tell the can point to a polar bear and tell you that that's there so your theory like this is only one of them poll one simple human can learn.

[00:59:09]
Very accurate information. But the planning needs a lot of training data with your training data we cannot achieve that level of accuracy so how do we make future deep learning model with more adaptive. With adaptive to less data to changing environment changing our changing demographics like we learned our model with seeing this population but it can be adaptive to the other population of maybe some disease have a lot of different subtypes we only of his or her show subtypes but we can identify the other subtypes who we have nothing before given some knowledge so this adaptive is something is lacking in current D. printing models for health care occasions and the other thing is explainable.

[01:00:08]
All that if models that we are talking about. Consider a black box so there are still Per of complex and sophisticated you know what input the word output but you don't know what the artist or the reasoning of these decisions like I said this patient can be diagnosed as a with some this but I cannot give you explainable is there a nation of why this model make that prediction so that's another challenge although there are many ongoing work trying to address you not to that.

[01:00:46]
Really like if they were both level so that's the 2 open challenges I hope if you guys are interested maybe we can work together and try to make the planning for health care better try to build better model and great more fascinating products and transform the health care industry.

[01:01:11]
So we're all working together to the future of health care thank you guys and. I mean we've heard from. You mean like the performance of model and human experts to evaluate their performance yet. I think like Jim O. has a few like engine most period publications they always consult the human experts to evaluate whether this makes sense or not like there are some kind of human evaluation being conducted on always.

[01:02:15]
First or some of the. My pleasure Well yeah this is what you meant vision or. Yeah if that was supposed to be I think current to the post to be augmenting dog person afeard because doctors cannot be always be available all the doctors can be like not having that like amount of time to read all these things formation or we can provide some kind of the support to these called thirst.

[01:02:57]
And the future in the future if we can improve the model accuracy and probably with care and there's a doctor like well one last question. Hello really you're involved yeah with certain. Things taking place in this world yeah right yes there's a lot of new research this thing.

[01:03:27]
So there my. Late How do you yeah. Also. Bring. Me. Are someone. Else Yeah the 3rd or of me how do we incorporate is the behave very data genome data and other types of data more data here yeah yeah research right. Even if you really mean this is G.P.S. the road to revisit.

[01:04:07]
Here children took the easy person to say Right yeah. This. Is. So true yeah so I think that's a very good point in fact our current research also limited. In terms of data because we don't have access to all these like genomics or behavior data you just mentioned especially behavior data related to human subjects is not easy to get in the use of the research proposed so that some kind of constraint or we are seeing but I think your point really makes sense because if you we have this kind of data like behavior like genomics we can build a happy model like when you make decision you can consider some of the more more critical risk factor like put a higher power of already like you over there 1st and then if not then we can play more sophisticated a model to tell the difference from the audience to start more thought.

[01:05:14]
Right so I think a hyper model probably or it worth exploring but we are seeing the constraint of data so yeah thank you Rob. OK. You. Know we want to predict the drug interaction for your drug so you do so for that paper I can quickly give you an idea how we took try to predict you know Russian for you drug we model you should drug or the node right and we use the drug similarity here.

[01:06:05]
For this graph and for you Dr is attached the way they are labeled after the label about her is multi heart like one heart each so you have another drug through this target drug you can interact with a lot of other drugs right there you see this like one hot back to the label we tried to predict the do for heart label that thought it's.

[01:06:32]
Different we can charge this off if you will. I will next week. OK structure was kind of. Circumcision. Yes.