Curriculum vitae
Ashkan Pirmani
Applied machine learning · Real world evidence · Drug safety methods
Health data is scattered. I work on what we can still learn when it has to stay that way.
45.5% more
records in a worldwide study of multiple sclerosis medicines and COVID-19, through the federated route I built. Its findings informed the global advice for people with multiple sclerosis.
31% higher
AUC-PR from personalizing a federated model to each of 32 countries, on 26,246 patients.
24% higher
AUC-PR with FLoRank, my low-rank personalized federated method, over standard federated learning.
18
registries agreed on one core data set I designed, with the same quality checks on every stream.
On this page
ashkan-pirmani.github.io · github.com/ashkan-pirmani · linkedin.com/in/ashkan-pirmani-3b8a8896
Now, since September 2026
Drug safety methods, with Johnson & Johnson Innovative Medicine as industry partner
An Innovation Mandate: a personal grant from the Flemish government's innovation agency (VLAIO). I am building methods that keep two treatment groups comparable when each data source has few patients: first in multiple sclerosis, against criteria set in advance, then on Optum's US health records and insurance claims.
Read the planSeven years at a glance
Experience
Innovation Mandate, drug safety methods
Hasselt University, with Johnson & Johnson Innovative Medicine as industry partner. Belgium.
- Won a personal Innovation Mandate grant from the Flemish government's innovation agency (VLAIO), with Johnson & Johnson Innovative Medicine as industry partner.
- Developing methods that keep two treatment groups comparable in propensity score analyses when each data source has few patients: first in multiple sclerosis against criteria set in advance, then on Optum's US health records and insurance claims. The plan.
Postdoctoral researcher
KU Leuven, Department of Electrical Engineering, and Hasselt University, Data Science Institute, jointly funded. Belgium.
- Personalizing a federated model to each of 32 countries improved ROC-AUC by 7.2% and AUC-PR by 31% over the same model without personalization, and beat pooling all records, on 26,246 patients. First author, npj Digital Medicine (2025).
- FLoRank, my personalized federated method built on adaptive low-rank factorization, improved AUC-PR by 24% over standard federated learning and F1 by 8% over LoRA, on registry data from over 26,000 patients at 146 centers. Prototype. ECTRIMS personal grant for scientific quality (2025).
- FLkit, an open guide for teams starting studies across centers that keep their own data: 29 contributors, supported by ELIXIR Belgium. I led its design. First author, arXiv (2026).
- Identified five barriers that stop research software from being reused, from scope to institutional trust, by retiring my own framework, FL4E. First author, Medical Informatics Europe (2026).
- Organizing team of OHDSI Europe 2025. Invited talks at Roche, the Flower community and the North American Society of Artificial Intelligence in Multiple Sclerosis.
- Coordinated with data holders, clinicians and legal teams in more than thirty countries. Supervised master's students.
Doctoral researcher
KU Leuven and Hasselt University.
- Added 3,527 records, 45.5% more, to a worldwide study of multiple sclerosis medicines and COVID-19, through a federated route I built for registries that could not send data: a Docker container at each registry turned its records into grouped counts. Its findings informed the global COVID-19 advice for people with multiple sclerosis. First author, JMIR Medical Informatics (2023).
- Designed the collection's data pipeline: one core data set agreed across 18 registries, three routes for data to come in, and the same quality checks on every stream.
- In over 5,500 people with multiple sclerosis, the anti-CD20 medicines ocrelizumab and rituximab were associated with more severe COVID-19 than glatiramer acetate, and interferon-β was not protective compared with other medicines. My part was data acquisition, the pipeline and the federated route. Shared first author, Neurology: Neuroimmunology & Neuroinflammation (2022) and Multiple Sclerosis and Related Disorders (2022).
- Led the disability progression study, from the data access proposal to MSBase and the ethics approval to the experiments. Designed AdaptiveDualBranchNet in Python and PyTorch, and ran it with Flower and Ray on the Flemish Supercomputer Center's cluster.
- ATHENA, a public-private partnership with Janssen Pharmaceutica, funded part of the PhD. Worked on the analysis in its machine learning work package.
- Built FL4E, a framework for federated studies, and the degree of federation, tested on classification and on survival analysis. On a public classification benchmark, letting half the centers send their records kept 97.5% of the fully federated score. First author, JMIR Formative Research (2024).
- Taught in the Data Science in Healthcare course at Hasselt University, and began supervising master's theses in 2021.
Research assistant, Decision Support Systems Lab
IranDoc, Iranian Research Institute for Information Science and Technology, Tehran.
- A simulation of Iran Post's national parcel network in R and NetLogo, searched with a genetic algorithm for a better set of hubs, found a layout with 15% lower projected transport cost, in simulation.
- Delivered as a decision support dashboard and a policy analysis for senior executives.
Skills, with proof
Methods
- Federated learning and federated analysis
- Personalization when sites hold different patients
- Privacy-preserving machine learning
- Survival analysis and classification on clinical data
Data
- Real world and observational data
- Cohort building and phenotype definition across registries
- Registry harmonization and data quality
- Data feasibility: whether a question can be answered at all
- OMOP common data model, through OHDSI
- FAIR data practice
Tools
- Python
- PyTorch
- Flower
- Ray
- scikit-learn and pandas
- R
- SQL
- Docker
- Kubernetes and continuous integration
Networks
- OHDSI Belgium, and the organizing team of OHDSI Europe 2025
- ELIXIR Belgium
- The MSBase registry and the MS Data Alliance
Worked with
- Johnson & Johnson Innovative MedicineIndustry partner of my Innovation Mandate
- Janssen PharmaceuticaATHENA partnership, which funded part of my PhD
- RocheInvited talk, to its federated learning team for multiple sclerosis
- MSBaseRegistry data from 26,246 patients in 32 countries
- MS International FederationThe worldwide COVID-19 data collection behind its advice
- OHDSI EuropeOrganizing team of the 2025 symposium
- ELIXIR BelgiumSupports FLkit
- Flemish Supercomputer CenterCompute for the federated experiments
- FlowerInvited talk at Flower Monthly
Education
Dual PhD
Electrical engineering, KU Leuven. Biomedical sciences, Hasselt University. Defended November 2024.
- Thesis: From Centralized to Federated. The Journey of Data in Healthcare.
- Supervisors: Yves Moreau (KU Leuven), Liesbet M. Peeters (Hasselt University). Co-supervisor: Niels Hellings.
MSc, Industrial Engineering
Socio-economic systems. Kharazmi University, Tehran.
MBA, Quality Engineering
Tose'e Institute, Tehran.
Recognition
Innovation Mandate
Flanders Innovation and Entrepreneurship, with Johnson & Johnson Innovative Medicine as the industry partner. Drug safety research on health records and insurance claims.
ECTRIMS personal grant for scientific quality
For FLoRank, at ECTRIMS 2025, Barcelona.
Fully funded PhD scholarship
KU Leuven and Hasselt University, through to the defense in 2024.
Teaching and supervision
Data Science in Healthcare
Teaching member, Healthcare Engineering program, Hasselt University. Two academic years.
Master's thesis supervision
Seven master's theses in artificial intelligence, KU Leuven.
- Predicting admission to care. Federated, pooled and local data compared.
- How splitting the data changes what the model learns.
- Federated boosting in semi-supervised learning.
- Contrastive learning for federated models.
- Personalized models. Architectures and parameter exchange.
- Fine-tuning a global federated model for local performance.
- Early disease detection from shallow whole genome sequencing of cell-free DNA.
Selected publications
All 22 entries are on the publications page. Talks, including Roche, Flower, ECTRIMS and Medical Informatics Europe, are on the talks page.
Let's talk.
If your question depends on real world health data, I probably want to hear about it.