Project summary

The International Classification of Diseases, Oncology version 3 (ICD-O-3) is the basis for the communication, billing and treatment of oncology patients as well as public healthcare planning. The documental task of assigning a code is executed over free-text clinical notes and is done by medical oncologists or medical coders, which causes it to be expensive and decreases the amount of time available to actually be with the patient.

Even in developed nations, it is estimated that the task of ICD-10 coding (very similar to ICD-O) takes around 0.5 hour per case, and the problem is even greater in resource-poor settings, where medical oncologists are few and far between.

To try to solve that, I want to take an existing dataset (CANTEMIST) developed and annotated by oncologists, extend it to 17 languagues of both the Global North and South and train a multilingual Small Language Model (SLM).

Concretly, these are the deliverables of this project:

- Multilingual Dataset (huggingface, public domain license)
- Training and Inference Code (github)
- Model Weights (huggingface, public domain license)
- Preprint (arxiv)

What are this project's goals? How will you achieve them?

There are two main goals: create a multilingual dataset and a Small Language Model for automatic ICD-O coding in 17 languages, both released as public domain.

The languages included are: English, Portuguese, Simplified Chinese, Ukrainian, Urdu, Hindi, Swedish, Finnish, Hausa, Swahili, Amharic, Somali, Japanese, Korean, Arabic, Bengali, and Indonesian. This language selection balances Global North and Global South contries to ensure it is usable by billions of people.

Here's the plan, step by step:

Translation (Month 1)
I'll translate the full Cantemist corpus into all 17 target languages using Hy-MT2-30B-A3B. I picked this model because it has SOTA results on translation benchmarks that include low-resource languages as well as on medical text. I'll manually verify a sample of translations in English and Portuguese.

Model selection (Month 2)
Once the multilingual corpus exists, I'll run a zero-shot benchmark across a handful of small (4B–12B parameter) models to find the best base for fine-tuning. Right now I'm looking at Gemma 4 and Qwen 3.5, but I'll stay open to whatever else drops in the next few months.

Fine-tuning (Month 3)
Whichever model performs best in the selection gets finetuned for the actual ICD-O coding task.
This will involve experimentation with finetunning parameters described by Schulman et al. F1 micro will be the primary metric, but several others (Accuracy, Precision, Recall) will be available as well, including macro-averaged values.

How will this funding be used?

All $500 goes to compute. I'll be running everything on Modal, using a container with 4 cores, 32GB RAM, 200GB storage, and an H100 GPU. Based on Modal's pricing:

CPU: $0.19/hr (4 cores, $0.0473/core/hr)
RAM: $0.26/hr (32GB, $0.008/GB/hr)
H100: $3.95/hr
Storage: 200GB falls within Modal's 1TB/month free tier.

That works out to about $4.40/hr or roughly 115 hours of compute. I estimate 100 hours for translation and finetunning and 15 hours as a reserve. If the funding exceeds the $500 minimum, I'll add more languages: Thai, Dutch, Polish.

Who is on your team? What's your track record on similar projects?

It's just me. I'm a Kaggle Datasets Grandmaster, formerly ranked #1 in the world. I have 9 years of experience as a data scientist, 7 of which were in healthcare. Previously I worked as principal data scientist for AI in healthcare, in a project funded by Brazil's Ministry of Health and Gates Foundation.

I'm currently in the process of transitioning into academia so that I can dedicate fully to AI for Global Public Health.

What are the most likely causes and outcomes if this project fails?

The biggest risk is performance on the lowest-resource languages (such as Somali and Amharic). Even a strong translation model like Hy-MT2 has less to work with for these languages, and the same is true for the current SLMs, where their training data mix is unknown and their tokenizers show better fertility rate on higher-resource languages.

There's also a chance that a 4–12B parameter model doesn't have the capacity to handle ICD-O coding at the professional level in all languages: while some studies have shown performance in high-resource languages like Spanish, several of the proposed languages were never attempted.

Even if the model fails entirely, the dataset and code are still released , and the preprint will document what didn't work and why, so the next person attempting this doesn't have to start from zero. Worst case scenario, this becomes a well-documented negative result and a reusable dataset.

How much money have you raised in the last 12 months, and from where?

None. All planning and execution has and will come out of my own pocket. This $500 crowdfunding request is the first external funding I've sought for this project and it is exclusively for GPU access.