Jean-François Legourd Manager • about 1 month ago
Healthcare Track (Elfie) • General Channel
Hello to all participants and teams,
You can post your questions here. The Elfie team will try to reply promptly.
This week will be intense and we look forward to seeing your work.
Best regards
JF, Dao, and Duy
Log in or sign up for Devpost to join the conversation.

11 comments
Nguyễn Công Tuấn • about 1 month ago
Will any of the tracks have a provided or sampled dataset?
Jean-François Legourd Manager • about 1 month ago
Indeed, for the healthcare track, we have provided:
1) The Medical Scribe challenge has one sample audio:
https://drive.google.com/file/d/1MUObyQIGQHJ3HsaGtAyOvJ9RUglbvORV/view?usp=sharing
2) The Body Scan challenge has no file because you simply need a image of your body front and profile. However, we can give you more information here about the vendor we will replace (they provided the SDK and we did the design, we want to replace the SDK):
https://www.bodygram.com/en/case-study/elfie
3) The Labs Analyzer challenge has one sample lab report (it is easy for you to get other report in a few hours at Diag in HCMC.
https://drive.google.com/file/d/1FzwtAsDlJclaXPqLScGa_EPEcSgpR0t7/view?usp=sharing
We designed these challenges in a way that UI is not the difficulty (we remove that difficulty by showing you what a UI/UX could look like).
The real difficulty is in building the capability around Qwen AI so you reach high clinical accuracy and medical usefulness. We are assessing your speed of learning (a domain you are unfamiliar), speed of building, and the elegance of your proposed architecture.
We hope this helps. Feel free to ask more questions.
Best
JF
Quoc Cuong Vo • about 1 month ago
Thank you Jean
Jean-François Legourd Manager • 29 days ago
Answer to a question from one participant to the Challenge 1 - The Medical Scribe:
*Evaluation framework*
You should test your solution against the following dimensions:
*1. Clinical correctness*
Can your system accurately capture the medical content of the consultation without dangerous omissions or misinterpretations?
Test yourself on:
* symptoms
* findings
* diagnosis (when present)
* treatment / recommendations
*2. Multilingual robustness*
Can your system still perform well if the consultation is in a language you do not personally understand?
You should test:
* different languages
* different accents
* mixed medical and non-medical speech
* imperfect audio quality
*3. Completeness*
Does your output cover all major parts of the consultation, not just the final diagnosis?
A strong output should include:
* reason for visit
* relevant history/context
* observations/findings
* diagnosis or working hypothesis
* treatment / next steps
*4. Actionability*
Does the patient summary clearly explain what the patient should do next?
Avoid vague outputs. The summary should be practical and understandable.
*5. UX clarity*
Is the final output easy to read and immediately usable?
We recommend a clear structure such as:
* transcript
* structured clinical report
* patient summary
* severity / urgency level
You should assume the live evaluation input will be:
* a real consultation audio recording
* unseen in advance
* potentially multilingual
* potentially noisy or unstructured
So the goal is not to optimize for one perfect example, but to build a system that generalizes reliably in real-world conditions.
Huy Le • 29 days ago
From Q6 on https://docs.google.com/document/d/1Ozzgo3kDQGuDW8Ma9h8AIj45yQX1DgNKeOqaCu7yCsc:
> During evaluation, our team will test your submission by randomly inputting data in one of several languages (Arabic, Vietnamese, English, or French). However, rest assured—if your product is well-built but fails solely because the Qwen model itself struggles to accurately handle a specific language, our team will not penalize you for that.
Does this mean a live hosted app is mandatory? Or is it acceptable (on the pitching day perhaps) to be given those file to run them live locally on my laptop?
Jean-François Legourd Manager • 29 days ago
The Elfie judges will shortlist teams BEFORE the pitchday purely based on the performance of their product.
To do so, we need a live window with the Elfie team can feed the input and receive an output in return.
Huy Le • 29 days ago
Thanks for the clarification JF, still, I'm still quite unsure whether it's possible to coordinate that live evaluation with a screenshare call?
This save us the infra logistics pain to focus on fleshing out the product itself. The hosting pain is real for us since we do use a combination of small models (not all available on Alibaba's DashScope) that wouldn't run great on CPU-only cloud hosting, and wiring to GPU-capable instances further amplified the pain.
Jean-François Legourd Manager • 29 days ago
EXPECTED OUTPUT
We will use your software (simply provide us with a link) to upload an input and retrieve the output.
Let's clarify the format of the input and output for each challenge.
1. The Medical Scribe
1.1. Input: an audio consultation (5 to 20 minutes, .wav or .mp3 file)
1.2. Output: a) transcript, b) structured clinical information (symptoms, findings, diagnosis, treatment), c) structured clinical report for doctors
2. The Body Scan:
2.1. Input: two pictures (front and profile), height in cm, weight in kg
2.2. Output: a) body measurements, b) body fat ratio
3. The Labs Analyzer:
3.1. Input: lab report (pdf)
3.2. Output: a) lab values (tests, units, reference ranges), highlighted abnormal results, patient-friendly explanations and next steps
Anh Trung Nguyen Vu • 29 days ago
Merci Jean-Francois
Nguyen Dinh Thien Phuc • 25 days ago
Hi Jean-Francois,
The account from team: Bunos Team (with my personal email: ndtphuclqd1306@gmail.com) is the same as my project: Bidget Health, for no reason I can not understand why my account on devpost with the email: ndtphuclqd1306@gmail.com is being blocked to submit on the website. Please take a look on this Bidget Health project (with the working email I use here).
Thank you so much !!
Jean-François Legourd Manager • 25 days ago
@Bunos Team: you can use the chat / discussion feature on your submission to address the problem and submit the link to your prototype. Right now, your page is made of:
1) a youtube video of a Vietnamese singer
2) a link to a worm-based nutritional product.
Regards