Winning Solution Walkthrough: CIBMTR - Equity in post-HCT Survival Predictions by Team Robert Hatch

What this episode is about

The main topic focuses on the development of a predictive model for post-HCT (Hematopoietic Cell Transplant) survival through the lens of equity, as presented by Robert Hatch. He shares his personal connection to the subject due to his daughter's bone marrow transplant experience and discusses the challenges he faced in addressing specific needs in the existing data models, while outlining his exploration and findings during various machine learning competitions.

Key takeaways

• Personal experience in healthcare unveils gaps; data models can't always fill emotional voids for patients.
• Winning competitions depends on robust models; clever strategies won't substitute for raw predictive power.
• In synthetic data competitions, the unexpected becomes an ally; intuitive feature engineering often yields unexpected advantages.
• Custom targets can reveal hidden patterns; sometimes the best insights lie in data's least explored corners.
• Meticulous tuning and model blending can create magic; but beware, the final recipe must balance complexity and clarity.

Who speaks in this episode?

• Robert Hatch (Host) · 29 min speaking

What is discussed, and when

1. 0:00 – 5:00: bone marrow transplant, light GBM, Calg competitions
2. 5:00 – 10:00: EFS
3. 10:00 – 15:00: cat boost
4. 20:00 – 25:00: bone marrow transplant, cat boost, light GBM, CARTT
5. 25:00 – 30:00: bone marrow transplant, EFS
6. 30:00 – 31:47: bone marrow transplant

Notable quotes from "Winning Solution Walkthrough: CIBMTR - Equity in post-HCT Survival Predictions by Team Robert Hatch"

• Nothing to do with my model and more kind of meta talk. as I'll go in my background, I'm actually my daughter two years ago did have a bone marrow transplant. so I have a lot…

  about bone marrow transplant · at 0:09

• Firmware engineer. and I've just kind of self taught this maybe five years ago in terms of when I entered Calg competitions, maybe three, I forget. but I have done well. I also…

  about Calg competitions · at 1:16

• As what I did in auto machine learning is I took a lot of public notebooks and handtuned light GBM in particular which I'm at least very familiar with and can tune very quickly…

  about light GBM · at 3:03

• Was what I'm put calling a in this formula here. and then B was the ranking of the EFS time. But I also did something very different. so I had a custom variant of predicting…

  about EFS · at 4:58

• Truly have all of them, but it was not a relevant difference. cat boost was really good because of the categorical data. but I'm not as good at tuning it, so I didn't ever…

  about cat boost · at 13:38

• Can find data because obviously I am an accomplished data scientist and know what I'm doing. was, hey, can I figure out whether my daughter doing a CARTT and then hoping it…

  about CARTT · at 22:25

Who and what is mentioned in this episode?

Concepts & topics

• bone marrow transplant — Concept · Core topic · 8 mentions · first at 0:19
  Discussed in the context of personal experience and medical treatment options for leukemia.
• EFS — Concept · Key argument · 2 mentions · first at 5:03
  Central to the discussion of modeling patient outcomes and risk scores in medical treatments.
• CARTT — Concept · Core topic · 1 mention · first at 22:38
  Mentioned as a treatment option and its comparison with bone marrow transplants in leukemia cases.
• light GBM — Concept · Framework · 3 mentions · first at 3:07
  Mentioned as a tool used in machine learning models for predicting medical outcomes.
• cat boost — Concept · Framework · 3 mentions · first at 13:45
  Discussed as a machine learning tool used in the context of medical data modeling.

Places & events

• Calg competitions — Event · Referenced work · 1 mention · first at 1:20
  Mentioned as part of the speaker's background in machine learning.

Full transcript with timestamps

0:01 Let's see. Um, so I'll go over um this. I actually I I kind of added a bullet that's nothing to do with my model and and more kind of meta talk. Um, uh, as I'll go in my background, I'm actually um, my daughter two years ago did have a bone marrow transplant. Um so I have a lot of sort of personal invest invested interest in this exact topic and um just you know my what I actually needed wasn't available and for me it's you know worth calling out that this by itself doesn't solve that doesn't doesn't fill the need that I saw. Um and that's fine. You know this may not be for for parents and patients in particular but I want to kind of talk a little about that. um if if we have time um and if uh the host uh you know doesn't mind me going on about that. Um obviously I don't know you guys' use case and and what you're trying to accomplish. So I can only talk so much but I can give a little bit about what I

1:04 see as the need and and maybe ways we can get there. Um unlike I guess every other team um I'm not yet uh working in the machine learning industry. Um I'm a firmware engineer. Um and I've just kind of self taught this uh maybe five years ago in terms of when I entered Calg competitions, maybe three, I forget. Um but I have uh done well. Um I I I also got third place in this auto machine learning grand prix and relevant for that that was actually a synthetic data um competition. So I was kind of half expecting things like feature engineering to not necessarily work. um that there may have been some advantages in my approach given that it was synthetic data.

1:51 Um let's see. Um I think it's also worth mentioning like you know my modeling approach um and and how I did this. I took this there was a lot of uh sort of exploration uh that I was that I did. It wasn't where because I was not at the top of the public leaderboard and a lot of people at the top were saying they were seeing pretty good CV to public leaderboard. I kind of knew I was in a sense behind in the competition. I ended up find so I was just looking for the trick and I got fifth place because in the end I was I found you know the kind of try A versus B which I thought of in the past but never tried. um maybe like 11 days before the competition was over.

2:31 So um as such, my models were not as polished. You know, I think first and second in particular, their single models, you know, were were at least on the edge of beating third, fourth place and and and definitely beat mine. You know, even with single models, they could have done that, right? So um I did a lot of things I'm very proud of, but you know, one of the things that will win a three-month competition is getting really good individual models. And I didn't do that to the level first and second did. Um, and I think that's worth kind of calling out. But what I did do as as what I did in auto machine learning is I took a lot of public notebooks and and handtuned light GBM in particular um, which I'm at least very familiar with and can tune very quickly and hand tuned one one of the public models in particular, the Tabm. Um, and then I also ran out of glue on. I got this, you know, what ended up really in

3:23 a sense being hundreds and thousands of models because of the target uh proliferation um and then sort of merged them back down. I different from I think each of the other teams I did merge in uh or at least most of the other teams um I did merge in ones that were trying to predict custom targets in addition to the A plus B, you know, the the classifier and regressor u combination. and I had a formula for that, but I also merged in a little some I saw some benefit in also merging in something that was trying to predict against single labels. Um, and I and I merged them back at the end. Um, so um I did have a few small feature tweaks um but the main important feature was just the target figuring out how to do the target um training time. You know, if you look at what what all of them were that were thrown into this kind of blender ensemble method, it would be large. But if it was only the final models, the

4:23 training time would be much at least a little bit more reasonable, even when it's running auto glue on. Um, oh, and I also apologize. Uh, some of this is not as polished as I would like on the slideshow. I'm on vacation and and even to last two weeks, so I didn't have a lot of time to throw this together. This is all a little bit better in my um writeup. Um and I'll and I'll try to talk to everything, but I may not have it all captured in bullet points. Um so I had the the a the classifier, you know, predicting EFS um was what I'm put calling a in this in this formula here. Um and then B was the ranking of the EFS time. But I also did something very different. Um, so I had a custom variant of predicting the FS label because I noticed that there was surprisingly low overlap. If you found the exact point where there was the least overlap between um EFS0 and EFS one by EFS time like where B essentially you find this kind

5:28 of cut point where as few EFS equals 1 are on one side of that and as few EFS equals zero are on the other side of that exact EFS time. it ended up being this like um you know 13.32 months or something like that. I forget what the number was. Um it's in my write up but um there was only about 2% that were miss that were on the wrong side. And I considered 2% to be, you know, like, well, what if that's noise and I'm actually better off pretending that I'm just trying to predict whether a a patient will reach this 13.3 to 4 months uh mark. Um, so I um so my custom classifier is actually predicting a custom variant of that. And I also said, you know, for then for the censored ones that were before 13.24, 24 like the 200 that didn't reach that mark. I said, let's instead of instead of using them as one or the other, I'm just going to ignore I'm just going to drop those ones from my classifier um for training. So,

6:34 I'll just ignore the ones that were censored crazy early. You know, something weird happened. Let's let's ignore that um and only use the others. And then for time, I just only looked at the ones that were under this artificial cutff point. um a and were actually you know an event happened. Um so my my A and B was a little bit custom because of limited time by the time I was trying this thing. I never had a chance to go back and find out would I have been better off doing both methods only one method or the other in terms of using true EFS versus using this custom label.

7:11 I have I didn't have a chance to figure out which was actually better. Uh this was just the one I ended up using um and it worked well. Um, but it may actually not be better. I I I didn't have the chance to check. Sorry. Um, the other thing I did before I got there is I is I did this large search where I was you know using a formula to sort of modify my um my uh you know single regressor uh trained on all the data to to train a single risk score. Um one of the things that worked really well was having sample weights. Um, again, this works less well than splitting the targets, but it is somewhat helpful as a supplemental. Um and and one of the things that worked really well was having a sample weight in the 30 to 35% range depending on like 30 to 40% range uh for the EFS0 samples so that you were mostly looking at the at the ones that mattered the most but you were getting a just enough kind of

8:12 pull for the ones that were zero that the um that the trees were able to make good splits. Um, and I think that was where that's where it had so much trouble, you know, as shown in one of the other graphs of of of P predicting. Um, it it just helps it a lot if it can figure out that a this group means this thing. Um, and that was hard to do both for ranking and at the same time, what about these ones that you need to always rank lower? Um, became a little bit tricky. Um, let's see. Um, I did do a couple features. they were not super important. They definitely seem to help a little but not a ton. Um there was a recalculating HLA sums because of synthetic data. Um but one thing is that in that one I noticed that they couldn't replicate the NMLP feature properly, you know, because because it's synthetic data maybe. Um so I made both. I was like, well, that could help, but let's also keep the original one because that

9:11 was a highly predictive feature compared to most of the HLA. most HLA ones didn't seem to be very predictive in the first place. Um, for gradient Brues, you know, um, the W I never got to trying it, but the winner did um, you know, the one hot encoding numerical is categorical. I've done that in the past before. I just didn't get to it in this one. Um, and the nice thing for grading boosted trees is because of how they work, it's generally best to just add and not subtract anything. So, you know, the first place solution was a great um approach for that. Um for neural nets, however, and even the first place solution, I'd be curious if pairing it down a little bit would help neural nets. Um because sometimes neural nets, if you add uh too many features, it is a little bit influenced by uh random features. Um so, non-predictive features can sometimes be helpful to to remove for neuronets. So I

10:05 did remove just the HLA 6 and HL8 sums because logically they should be pretty encapsulated in HLA 10 and the individual uh ones so the neuronet can figure it out on its own. you don't need to sum every single combination of some um for neuronets it was how it was done in some notebooks and I and I left it and didn't have a lot of time to go back and I just kind of forced all but two is categorical that I the idea also is it's a pretty quick way to have diversity between the main models because the GBTS will usually do much better and the neuron nets um another thing I did when I was that was more because I was worried about my delta between public leaderboard and training data, but I think it's also a just a a good approach is I I slightly rounded the numerical features. One of them was like the age of the donor and obviously that's not very going to be very sensitive to outcome. You know, you don't really care

11:06 if they were 23 in one month versus 23 in three months. There's there's not going to be relevant data there. So um so I rounded that I think to the nearest two years and then for the for the age of the patient um there is going to be quite a bit of difference potentially but it's still not going to be at the level of you know 23 years 2 months and 4 days versus 23 years 2 months and and seven days right so I rounded to 0.1 which would be roughly to rounding to the month um and I actually ended up using both the rounded and not rounded But I mostly from the point that I tried that used mostly the rounded version and that seemed to help very slightly. Um and then sometimes I since I included some public notebooks um sometimes the diversity of having them with features even if those features are kind of noisy and probably not helpful um can be a little bit of help just by injecting diversity. Um I kind of already talked a

12:09 little bit this you know I used a lot of training methods. I use a large ensemble. Um I I use this uh greedy ensemble selection where you basically add one model uh and try to improve the score. Then you add then you pick the model that adds best to that one and just keep going until you don't see improvement. I do that also with the validation thing. So instead of stopping when training is not improving, I stop when validation is not improving. Uh while optimizing the training set, not the validation set. Um so that means that you know what originally I did you know this 1,700 um different target variations but then I ended up only using about six I believe of those target variations based on this ensemble method. So then and then when running those six uh methods on a whole bunch of different things plus running the A and B uh that's when I then ensembled these all um that's how I it's it's very similar to autolun's

13:11 builtin although I was not using my own custom version of that. Um so that's how I'm using what ends up being you know dozens or hundreds of uh I think it was like 46 models that made the final solution plus a whole bunch that did not make the final solution. Um and I ensembled them all. Um, I saw this very tiny degradation when I removed all the ones I didn't that weren't in my final because of just how the the algorithm works even slightly better when you truly have all of them, but it was not a relevant difference. Um, um, cat boost was really good because of the categorical data. Um, but I'm not as good at tuning it, so I didn't ever really hand tune it. I just used some public notebooks and autogluent has some good cat boost settings. So autogluin really was the did the line share of the cat boost portion of it although I did a tiny bit of hand tuning of it but almost none. Um tab in I hand tuned but I don't

14:06 know how much I I definitely feel like it was better but I didn't wasn't able to get this huge improvement there. Um but some and then the ODST although I spent time on it wasn't able to improve so I just used it kind of as is but with a new seed to reduce its u overfitting of the public leaderboard. Um, so while I don't, while my models were not, you know, quite at the level state-of-the-art of the first place, um, a couple things that I did that I really like to do, and I've never, I don't think I've ever seen it not work where it made sense is, um, leaving things as logits, if they were that in the first place, um, before converting them to probabilities after ensembling.

14:53 um you know and and I also did for the for the ranking EFS time um I converted that to as if it was a logit um I I I ended up taking I I ended up making it from like 0.03 03 to 0.97 instead of 0 to one ranking. Um and then I uh and then I took the logit of that 0.3 to uh to 0.97 and then I trained the regressor to predict the logit value instead of predicting the uh the raw probability or or the raw ranking. Um, so it was a way of I think another team also saw it better when they did a logistic uh or or an exponential, I forget. Anyway, um I believe it was similar to them just from a different approach. Um but because I did it that way, I also found that for the regressor as well, um leaving it as logits during insembling was better and only transform at the end. Then because it was logits, I found certain um post-processing was really effective.

16:00 Um, and I didn't have a time to really write this into here, but I believe Well, oh, I did a little bit. Yeah, I believe the big difference here, so I I think I did a conservative 1.5 because it wasn't working on public, but it ended up working really well on private and training, which was more data. Um, so I suspect that I could have actually gone a little bit higher. I didn't have a time to go back and test it. U, but I suspect a higher value than what I did was would be even better. And I wouldn't be shocked if it would also improve like second place team to do the post-processing even though they also saw some post-processing do worse on public leaderboard so they didn't do it.

16:38 Um but um but I think that may have been just been noise on the public. Um what I found what what I speculate is I think there could be also this because of the survival GAN because it's synthetic data there's some reasons why um perhaps the split works especially well and the post-processing might be related to that but I actually think it's the concordance index um not because of how the concordance index does you know sensoring or comparisons but really the fact that the concordance index is a relative and not uh you know an absolute um what that means are uh and you're just looking at risk specifically the risk score portion of the concordance index you know you're comparing just these relative values they don't have an absolute meaning and what I speculate is the reason for post-processing is actually because if you already know that all of the models let's say you're 200 most risky out of

17:39 out of 10,000 you're you're you're you know your 1% most risky um um uh most risky uh data points. Well, you no longer care whether those 200 you've already ranked them above the other, you know, 9,900 uh 9,800. So you really you but you want you actually still get a small amount of benefit if you then essentially say well what about within this 200 which one is actually which one should I call the highest risk and which one should I call the second highest risk. It's no longer that hey A is 99% likely to beat the ones that were ranking ahead of it and two is only 94% likely to beat all those other ones. Well, that doesn't matter because you already said it's probably going to beat them. Now, all you care about is will it beat the other 199 within this little group. Um, you can think of it that way, right? And so, I speculate that further emphasizing these logits helps basically zoom in on a kind

18:49 of a local optimized rank. Um, I never had a chance to do other experiments on that. So, it's it's kind of I'd call it just a speculation at this point. Um but I did two different things. I believe that multiplying the logic by a larger value um you know when it's when it's probably EFS0 now you're saying there's less and less importance I'm going to put on its EFS1 ranking. So therefore, with the least risky ones, I'm paying more attention to what's the slight chance that I'm just wrong and it's not an EFS uh zero because that's all that's going to matter for your ranking score. And the other way, if it's an EFS1, I'm emphasizing both of them. Um, and that just means that now I'm just essentially looking at what is the chance that this one beats another one. So you're kind of again, you're emphasizing what matters for that section. And then I also had this little move away from zero the the

19:48 EFS probability which again I speculate helps locally optimize the middle of the range where you're just hey I'm going to pretend to be just a little more confident than I really am which will probably uh slightly help my me figure out um because probably the classifier is has a little bit more signal than the regressor for these unknown uh data points. the ones that you're just not very good at figuring out. Um, so that was kind of what I did there. I suspect that that would see benefit even in first and second place.

20:22 That's why I'm kind of spending a little more time talking about it, but I don't know that. I didn't have a chance to run that experiment. Um, I found that cat boost was most important for EFS classifier prediction and light GBM worked best for me at least for um the regressor prediction. I did something that probably was not nearly as good as what first place did where they put every thing into the the regressor. I used only the half of the data, but I did have the rest of the data in with sample weight equals zero.

21:01 Um, which weirdly made more difference than I expected. Um, I think it was something about how it does the grouping even though it doesn't affect the weights given it it somehow affects the grouping at least with the regression uh not with regression with the regularization settings that I had. Um for the simple model, I found that this really simple um pairing of cat boost and tabin for classification and light GBM and cat boost for regression uh worked really well with this like kind of doubling the logits um at the end to really emphasize this local confidence.

21:37 Um um so for me it got a 69812 which was within um which would have been like sixth or seventh place and it was very close to my uh full models uh score. Um uh let me ask real briefly um if it's fine for me to I'll just go just real briefly on some kind of off-topic off the main topic stuff. Um, if that's fine, if that's okay with the hosting team. Yeah, that works for me. Okay. Um, so f for me the big question where I was literally looking to find out whether I can find data because obviously I I am an accomplished data scientist and and know what I'm doing. uh was, hey, can I figure out whether my daughter doing a CARTT and then hoping it works? And it often does not. It's only got about a 50% um success rate. Um but the other 50% is almost always relapse. There's very little uh risk um for a for a kid at least of not surviving, of anything happening besides relapse. Um, and if it if it works, um, there's very little,

23:04 um, sort of side effects. You you you're not, um, doing anything too harsh on the body. There is a, you know, a window of time that's dangerous and that is hard on the body, but it's it's small and it's pretty well contained with modern medicine currently. Um, but it's super new. There's not a lot of data. Even if even if we had great models, the models would not be very confident. But that's what I wanted. I wanted the most confident we can get instead of this pure guesswork. Um or the other approach is car is for bone marrow transplants um and with leukemia in particular it's actually very critical that there's no um and I'm sure the hosting team is aware of this but there's it's very critical that there's no um sign of of leukemia. Um, and the way I was told, my understanding of this is essentially one of the dangers is it takes a really long time for the new immune system to to really

23:57 mature. And the goal, what you need is for the new immune system to mature before um any leukemia that may have still been around is able to uh exponentially grow. Um, so CARTT is the best um a bridge to transplant. So it was we were doing CARTT either way, but we didn't know whether to just immediately do a bone marrow transplant no matter what or try to see if the CARTT um would take um and bone marrow transplant has a much uh higher success rate of more like 70% um for for children. Um I think it was around that level and probably a little better because the statistics usually lag the medicine because they're you know such long you know fiveyear studies. Um and um but on the other hand a lot of the u failures um with bone marrow transplant are are more serious.

24:57 you know, relapse is actually a much smaller percentage. Um, and there's much higher chance of u passing away and there's also a lot more side long-term side effects. You know, simply going through the radiation means you have a, you know, small but a slightly higher chance of unrelated cancers or other complications. Um and and but for me it was really not even which one is better but it was does which one impacts the other one. You know you could do a bone marrow transplant and then if that doesn't work you and and and you just relapse well you could just do carti.

25:33 You know if you could do carti and then wait two years and when you finally relapse you can do bone marrow transplant. But it's not as simple as that because if you've relapsed, you have the worry of can I get it back under control again in the first place. If you use CARTT into a bone marrow transplant, well, maybe the fact that you've exposed it to CARTT the first time means you can't use the same CARTT the second time. It's less likely to work. So, um it's all it's also really hard to know and would even be hard to model even if we had the right models. uh the really specific of I did X and now I'm trying to do Y or uh how did that affect my chance of Y um but there's still you know there's still some level extrapolation you can do um if you have a sufficient data and even if you have insufficient data a little extrapolation just not much um so my you know kind of question really is is what are we trying what

26:31 what are you trying to solve who is this for is this for you know doctors and researchers um is this for patients? Um, and you know, almost regardless of that question, I'll just kind of since I'm just writing the slides, you know, I'll just give my my my my my take is that you need not simply uh a risk score, but you really need the set of models that almost is like a flowchart and then you do really the same as this A and B, right? You know, really the flowchart for this uh solution was does an event happen? Yes or no. If no, we're done. We know our risk score. it's super low. If yes, then it was like, okay, and how quickly does it happen? That's our risk score. Um, but instead, we can have what we really want is to what's our chance of hitting no, you know, where you're like, okay, had relapse happened, but that doesn't mean we're we've lost now.

27:24 We've got to try a new thing. If we say one more relapse and throw into this you know some other model uh you know that a leukemia doc you know group has made for trying experimental drugs you know what are our chances and you know I feel like there's this theoretical ability to for each kind of or a group of owners to start publishing open source models or you know at least open models that do a small piece of this, you know, a CARTT group can be like, well, you know, our model's not super accurate. It's it's only so great because we only have so much data, but here is the model, you know, and and I and you can feed data into it. And you can say, what if they've tried what if they'd done a carti therapy before as a part of bone marrow transplant, you know, you that can be one of the features of the model.

28:15 So, another thing that I think is really difficult and I and I really noticed when looking at the data is it's really hard to get all the features that would help this meta model um into your model and and I'm not sure if they're all there today or not. Um and but then the biggest the bigger one I think is that target that EFS equal one or zero. you know, can we get something that's not simply an event happened, but a lot more stratified on, okay, at this point, you know, chronograph versus host happened at this point, uh, relapse happened versus survival, you know, they they passed away, all these other possibilities. Um, and of course, if we can then model confidence intervals, you know, that'll be really helpful for patients and doctors in terms of if they're trying to weigh two very different um uh two very different um plans of attack, you know, knowing that okay, the obviously the cartoon model it's

29:20 confidence is is all over the place. You know, it could be it doesn't really know. just has at least a speculation for how long because there's so much less data than for bone marrow transplant. Um I think everyone's already mentioned the synthetic data. I'll just you know add there's you know it makes it harder to figure out um what the best model would be. It it definitely allows you to do a good model. Um, it's possible that, you know, some of the people here could have made even more amazing models with with with access to real data. Um, the metric um, you know, I I have a bit of a concern with the concordance index because even when you split into this classifier and regressor, you still are basically forced to combine it into one at the end. And it's this ranking where local optimization matters. And that maybe isn't really the most important thing. You know, it's possible that actually

30:24 training that actually having a metric of the EFS classifier and a metric of ranking, you know, when it'll happen for individual events is a better, you know, like actually literally breaking up that might work better um for sort of real world scenarios. Um and as I mentioned, you know, the the target columns having more granularity of what's really happening again helps with this sort of uh that concept. Um and you know, so I like I just kind of briefly, you know, obviously there's well if no adverse outcomes happened within certain time range u versus you know, not surviving versus relapse versus an unusual relapse, right? where you know you can get a a mutation you can get 20 years later an unrelated cancer or whatever.

31:14 Um and then I'm not really sure what the other possibilities are. I know there's chronic vasop and I think there might be other things that are you know a form of of rejecting the bone marrow transplant that can that would also need to be considered an event. Um but that's all I have. I just kind of wanted to highlight that because that's um near to my heart. Um, I hope you guys, you know, I love that this competition happened and I hope you guys are able to uh kind of further um find ways to help the the patient community.

Questions about this episode

Who speaks in "Winning Solution Walkthrough: CIBMTR - Equity in post-HCT Survival Predictions by Team Robert Hatch"?

Robert Hatch (Host).

What is "Winning Solution Walkthrough: CIBMTR - Equity in post-HCT Survival Predictions by Team Robert Hatch" about?

The main topic focuses on the development of a predictive model for post-HCT (Hematopoietic Cell Transplant) survival through the lens of equity, as presented by Robert Hatch. He shares his personal connection to the subject due to his daughter's bone marrow transplant experience and discusses the challenges he faced in addressing specific needs in the existing data models, while outlining his…

What is mentioned in this Kaggle episode?

6 people, works and ideas are mentioned, including bone marrow transplant, EFS, CARTT, light GBM, cat boost, Calg competitions. Each one on this page links to its own page with every episode that mentions it.

How long is "Winning Solution Walkthrough: CIBMTR - Equity in post-HCT Survival Predictions by Team Robert Hatch"?

The episode runs 32 min; the transcript is about 5,558 words in 35 timestamped sections.

More from Kaggle

• All analyzed episodes
• People mentioned on Kaggle
• Topics mentioned on Kaggle