I'm a medical doctor and data scientist passionate about using AI to transform healthcare. I currently serve as Medical Manager of Innovation and AI at Hapvida NotreDame Intermédica and am a PhD student in Machine Learning for Public Health at FSP-USP, researching multicalibration and fairness in predictive models to ensure AI is equitable across all populations. 🌱
- 🏥 Medical AI Manager · Hapvida NotreDame Intermédica
- 🎓 PhD Student · Machine Learning in Public Health · FSP-USP (LABDAPS)
- 🔬 Research in fairness, multicalibration and dataset shift in health
- 🧪 Builder of IA.med · open tools and content on AI in medicine, in Portuguese
- 📍 São Paulo, Brazil · Vegan 🌱
- ✉️ drfabianofilho@gmail.com
The work I'm shipping now follows a few convictions that came out of the clinic, the lab and the health system:
- Local first. Clinical text should not leave the machine. My recent tools run on a self-hosted LLM, offline, with no API key required to try them.
- Brazilian public data as a first-class citizen. DATASUS, InfoDengue, Anvisa, CFM and Conitec are rich, open and under-tooled. I build the plumbing so others can skip it.
- Evaluation beyond accuracy. Variance between identical runs, overconfidence and fragility to semantic perturbation are what accuracy hides. Measure them.
- Honest scope. Every tool states what it is not: not an official source, not a medical device, not a substitute for clinical judgment. The official link comes with every answer.
- Agents that plug into the clinician's workflow. MCP servers, Claude skills and reproducible pipelines instead of yet another chat window.
Languages & Frameworks
Machine Learning & AI
Data & Infrastructure
Health & Interoperability
🔌 IA.med MCP suite · iamed.cc/mcps
Model Context Protocol servers that give any LLM client (Claude, Claude Code, others) grounded access to Brazilian health regulation and literature. Local, auditable, every answer links to the official document.
| Server | What it does |
|---|---|
| anvisa-mcp | Drug registration status and recent medical devices from Anvisa open data, with a heuristic flag for AI-based devices · public connector available |
| radar-cfm-mcp | Searches and monitors CFM (Federal Council of Medicine) resolutions on AI, telemedicine and electronic health records · public connector available |
| protocolos-pcdt-mcp | Queries the Ministry of Health clinical protocols (PCDT/Conitec) and summarizes the recommended conduct with a verified literal citation |
| radar-papers-mcp | Watches PubMed and medRxiv for configurable topics and summarizes new papers with a local LLM, deduplicated by DOI |
| revisor-notas-mcp | Checks SOAP clinical notes for missing fields and incoherence between complaint, exam and plan · 100% offline |
| Project | Description | Stack |
|---|---|---|
| lab-ai-prediction | Open predictive-modeling lab: run the same pipeline on DATASUS microdata, your own dataset or international benchmarks · live app | Python · Streamlit · scikit-learn |
| datasus-outbreak-prediction | Integrated epidemiological surveillance: outbreak alerts (z-score, CUSUM, LightGBM), CID-10 co-occurrence graph, supply planning and sanitation × disease map · live app | InfoDengue · SIM · SNIS · IBGE |
| datasus-ai-prediction | Predictive ML models trained on DATASUS data for clinical and epidemiological outcomes | Python · scikit-learn · XGBoost |
| datasus-ai-search | Natural language queries over Brazilian epidemiological data, no SQL needed | Python · LLMs · DuckDB · Streamlit |
| Project | Description | Stack |
|---|---|---|
| iamed-harness | Minimal, auditable harness to evaluate LLMs on clinical multiple-choice tasks: run-to-run variance, calibration and semantic perturbations. First command runs offline, no API key | Python · uv |
| cardiovascular-timeseries-prediction | Monthly cardiovascular mortality in São Paulo state (SIM, 2010-2023): SARIMA vs Prophet vs TimesFM vs boosting, rolling-origin backtesting with measured uncertainty | skforecast · TimesFM · SIM |
| dengue-timeseries-skforecast | Dengue forecasting for 8 Brazilian capitals (InfoDengue, 2010-2024): seasonal naive, SARIMA, Prophet, LightGBM/XGBoost/CatBoost/RF and TimesFM 2.5 zero-shot | skforecast · InfoDengue |
| amie-agents | Conversational medical AI assistants inspired by Google DeepMind's AMIE, multi-agent architecture | MedGemma · FastAPI · LangChain |
| Project | Description |
|---|---|
| awesome-medical-ai | Curated list of Medical AI resources: papers, datasets, tools, pre-trained models, code and competitions, with weekly automated updates |
| awesome-health-datasets | Open datasets in healthcare, medical imaging, clinical data and bioinformatics |
| labdaps/labskills | Collaborative Claude Code skills for the LABDAPS lab (FSP-USP): ML in health and scientific writing, distributed as plugins |
| iamed | IA.med clinical suite: 8 structured clinical-reasoning templates as a Claude skill |
- 📄 A bibliometric network analysis of coronavirus during the first eight months of COVID-19 · IJERPH, 2021 · 106 citations
- 📄 Strategies for detecting and mitigating dataset shift in ML for health predictions · Journal of Biomedical Informatics, 2025
- 📄 Predictive modeling of gestational weight gain: a ML multiclass classification study · BMC Pregnancy and Childbirth, 2024
- 📄 Multicenter comparative analysis of local and aggregated data in COVID-19 outcome prediction · PLOS Digital Health, 2024
- 🧪 IA.med · Curated papers, podcasts, lessons and open tools on AI in medicine, in Portuguese
- 📰 Substack: Medicine and Artificial Intelligence · Newsletter on fairness, LLMs and generative AI in medicine
- ✍️ Medium: dadosesaude · Educational articles on health data science
- 🌐 portfolio.fabianofilho.com · Full portfolio with career, projects and publications
"Evidence-based medicine, combined with data science, has the power to transform patient care in a personalized and equitable way."