Projects

This is a page for the different projects I have worked on to date. Currently, a better overview can be found on my Github profile.

• 💊​MEDS: An ecosystem for high-capacity Health AI
• 🧪Yet Another ICU Benchmark: A framework for benchmarking EHR tasks accross (open-access) datasets
• 🥧ReciPys: A simple, declarative framework for defining preprocessing pipelines for ML-based data management systems
• 🧑‍🌾GUIDatafarm: A human-in-the-loop ML-based query optimizer data generator
• 📃Care2Report An automated ontology-matching based reporting tool

💊 MEDS – Medical Event Data Standard

MEDS (Medical Event Data Standard) is a simple, event-based data standard and ecosystem for health AI that targets longitudinal medical record data such as EHRs and claims. Its schema represents streams of medical events in a minimal, ML-oriented format that is intentionally designed for interoperability, computational performance, and ease of use across diverse health systems. The ecosystem includes schema definitions, Python types, and ETL libraries that convert widely used datasets—including MIMIC-IV and OMOP—into the MEDS representation so models and tools can reuse the same interface. Building on this core, MEDS-DEV introduces a decentralized, extensible validation benchmark to make comparisons of ML for health more reproducible and robust across institutions and tasks. Recent publications describe MEDS both as a simple interoperable data standard and as a reproducible health AI ecosystem, positioning it as infrastructure for scaling clinical ML from individual datasets to entire health systems.

🧪 Yet Another ICU Benchmark (YAIB)

Yet Another ICU Benchmark (YAIB) is a framework for standardized clinical machine learning experiments on intensive care and other EHR datasets. It provides a common structure for defining cohorts, prediction tasks, endpoints, preprocessing, and models so researchers can run comparable experiments across multiple databases without rewriting boilerplate for each setting. Out of the box, YAIB supports datasets such as MIMIC-III, MIMIC-IV, eICU-CRD, HiRID and AUMCdb, and includes default tasks like ICU mortality, acute kidney injury, sepsis, kidney function, and length-of-stay prediction. The associated paper presents YAIB as a flexible multi-center benchmark that emphasizes comparability across hospitals and data sources rather than focusing on a single curated dataset. By standardizing the experimental pipeline, YAIB helps illuminate where models generalize, where they fail, and how design choices in preprocessing or task definition influence reported performance in ICU risk prediction. github

🥧 ReciPies

ReciPies is a lightweight, declarative data transformation pipeline designed to make machine learning preprocessing more modular, transparent, and reproducible. Instead of scattering data preparation across notebooks and scripts, it provides a configurable framework where transformations are specified and composed in a structured way, making pipelines easier to share, review, and adapt. The system is built to integrate naturally with Python-based ML workflows and modern data tooling, offering a clear separation between data engineering steps and modeling code. In practice, ReciPies helps reduce one of the most common sources of irreproducibility in applied ML: untracked or inconsistent preprocessing that changes silently between experiments. The JOSS paper “ReciPies: A Lightweight Data Transformation Pipeline for Reproducible ML” formalizes this contribution, framing ReciPys as research-backed infrastructure for reproducible pipelines in settings such as clinical benchmarking and ML-based data management. joss.theoj

🧑‍🌾 GUIDataFarm

GUIDataFarm is a human-in-the-loop interface for the existing DataFarm framework, which generates and labels training data for ML-based query optimizers. DataFarm uses a data-driven, white-box approach to synthesize large, heterogeneous workloads from small initial query sets, input data, and available compute, producing realistic jobs together with labels such as runtime or cardinality. GUIDataFarm adds an intuitive graphical interface on top of this pipeline, allowing users to inspect generated plans, understand feature importance and model explanations, and steer active learning by choosing which queries to execute next. This human-guided process yields higher-quality training data tailored to the user’s optimization problems, overcoming the bottleneck of collecting thousands of labeled query plans in real systems. The approach is documented in the CIDR and ICDE publications on “Farm Your ML-based Query Optimizer’s Food!” and related work, which present GUIDataFarm as a practical path toward learned query optimizers in modern data ecosystems. vldb

📃 Care2Report

This project was conceived as my bachelor project in 2019 by the research group of Prof. Dr. Sjaak Brinkkemper at Utrecht University. The project was supervised by Dr. Kees van Deemter, and I was responsible for the implementation of the system. Care2Report is an automated medical reporting system that integrates speech and action recognition with semantic interpretation to reduce the administrative burden of clinical documentation. The system’s dialogue summarization pipeline transforms multimodal inputs from real consultation sessions into structured medical reports, using knowledge graphs and ontology-based matching to align content with existing electronic health record architectures. By automating both transcription and semantic structuring, Care2Report aims to free clinicians from repetitive data entry while improving consistency and completeness of the resulting documentation. The HICSS 2020 paper “The Care2Report System: Automated Medical Reporting as an Integrated Solution to Reduce Administrative Burden in Healthcare” details the functional and technical architecture and reports initial evaluations on real-world clinical data. As a project, Care2Report showcases how ontology-aware, multimodal processing can be used to turn raw clinical interactions into usable, interoperable medical reports at scale. Check out the HICSS paper for more details. To learn more about the continuation of this research, please visit the Care2Report website and the startup, Verticai, which is based on the continuation of this research in the era of Large Language Models.