Arkadiusz Sitek

Research Scientist and Physicist

Massachusetts General Hospital | Harvard Medical School

My research focuses on translational healthcare data science, specifically developing AI-driven solutions to improve clinical practice and patient outcomes. I work with diverse biomedical data including medical imaging (X-ray, CT, US, MRI, and SPECT/PET scans), clinical data from electronic health records (EHRs), omics data (genomics, proteomics, metabolomics), physiological time series, and behavioral data.

“Science is the belief in the ignorance of experts.”

— Richard Feynman
Arkadiusz Sitek

Massachusetts General Hospital

55 Fruit Street, Suite 427

Boston, MA 02114, USA

Email

selected recent publications

2025

  1. Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation

    Mamba Goes HoME: Hierarchical Soft Mixture-of-Experts for 3D Medical Image Segmentation

    Szymon Płotka, Gizem Mert, Maciej Chrabaszcz, Ewa Szczurek, and Arkadiusz Sitek

    NeurIPS 2025

    We introduce Hierarchical Soft Mixture-of-Experts (HoME), a novel two-level token-routing architecture for efficient 3D medical image segmentation across diverse modalities (CT, MRI, US). Built on the Mamba Selective State Space Model backbone, HoME addresses key challenges in medical imaging: modeling local-to-global spatial hierarchies, handling modality diversity, and achieving scalability for high-resolution 3D inputs. The architecture combines local expert routing with global context refinement through a hierarchical design that partitions sequences into groups, routes tokens to specialized experts for localized feature extraction, and aggregates outputs via a global layer for cross-group information fusion. Mamba-HoME demonstrates superior generalization and outperforms state-of-the-art models across multiple datasets while maintaining memory and computational efficiency.

    PDFCode
  2. Foundation model of electronic medical records for adaptive risk estimation

    Foundation model of electronic medical records for adaptive risk estimation

    Pawel Renc, Michal K. Grzeszczyk, Nassim Oufattole, Deirdre Goode, Yugang Jia, Szymon Bieganski, Matthew B. A. McDermott, Jaroslaw Was, Anthony E. Samir, Jonathan W. Cunningham, David W. Bates, and Arkadiusz Sitek

    GigaScience, Volume 14, 2025

    We present ETHOS-ARES (Adaptive Risk Estimation System), a foundation model for electronic health records that learns comprehensive representations of patient health trajectories. Using a transformer-based architecture, ETHOS-ARES processes diverse medical data including vital signs, lab results, medications, and clinical events to perform zero-shot predictions of critical outcomes such as hospital mortality, ICU admission, and prolonged length of stay. The model leverages the MEDS (Medical Event Data Standard) format and demonstrates strong performance across multiple healthcare prediction tasks without task-specific fine-tuning, offering a flexible and powerful tool for clinical risk assessment and decision support.

    PDFCodeWebsite