bloods.ai Blood Spectroscopy Classification Challenge
Helping Global
$7 500 USD
Completed (~4 years ago)
Classification
1103 joined
265 active
Start
Oct 19, 21
Close
Feb 13, 22
Reveal
Feb 13, 22
This was a very interesting problem. Could we learn from it
Connect · 14 Feb 2022, 08:18 · 5
Dear Zindi,
I think this was a very interesting problem, and unlike many out there, can actually have a tangible effect on health (diagnosis, cost, continuity of care, telemedicine etc) here in Africa.
I request that all the top n solutions are encouraged to share their approaches. Also, given that the final top scores were also not above 95%, we should rethink how we find a better approach to get 95% or more accuracy, validating this test and a good contactless screening/home test.
I agree.. though it will be interesting to see how there is such a big drop between the public and private leaderboards.. > 91 % on public Vs 60% ish on private
As mentioned in other post, I suspect there is distro shift between test and train, especially if you use the data was added later. If you used earlier data, you got better score, so I think there is more similarity between test and earlier datasets.
I think there is a lot we could learn from this competition. From my interaction with the dataset, I would say that the existing features (absorption values) do not possess enough signal to adequately solve the problem with a high degree of accuracy. This could be taken into consideration by the bloods.ai team during the data collection process. Perhaps, related data got from other wearable devices like pulse rate, blood pressure could be integrated.
Congratulations to every participant who were able to crack this very interesting project.
Although no one has final say when it comes to Zindi final decision on this. I think for this kind of challenge and for the benefit of the organizers and the benefit of having best and reliable model for deployment, if @Zindi will just select for every individual automatically the best two submissions like before the true picture of the robust models will be revealed. Unofficially, my team came 3rd position ( with 0.6095 on Private LB) and i can boldly say that the kind of research work my teammate and I put on this project which served as the bedrock of our feature engineering approach based on our medical domain knowledge is more than enough for Ph.D thesis. Though we ended up at 14th position for now. But i will like to know if it's possible to still share this our idea in form of our best notebook with the @Zindi or the Organizers. I must say that this notebook is an asset for the organiser that shouldn't go unnoticed. My take though.
Cheers.
MICADEE
for: Team "blood metabolites"
Thank you for so much for throwing some light into the kind of creativity and hardwork we put into this very difficult challenge over the course of the past 4 months. I personally got obsessed with solving this problem primarily because of its difficult nature and the fact that it is also an interesting health-related problem (given my background as medical student). It's rather unfortunate that we missed out on selecting our best model ,perhaps out of haste :( !!. Maybe this is something @Zindi might want to look more into for future competitions of this nature - although I understand that there might be some computational burden in automatically selecting participants best submissions especially out of the too many submissons per individual or team. Also,as my teammate already mentioned, we are more than glad to share our approaches with the Organizers or @Zindi and perhaps engage in ways to help reproduce our solution in a production setting.