Hamza El MassaoudyResearch & engineering

Research

Research

My research examines the reliability of medical AI when data and acquisition conditions change.

I use external validation, dataset audits and failure analysis to investigate the limits of model performance. My current work studies trustworthy atrial fibrillation detection with ECG foundation models, lead selection and deferral.

01 · Paper and code

Liver CT: testing beyond the training data

Wen X., El Massaoudy H.*

Multimodal Classification of Primary Liver Tumors on Multiphase CT: Internal Performance, Acquisition Artifacts, and Limited External Transportability

Journal of Intelligent Medicine and Healthcare (Tech Science Press).

Under revision

Case study Code Release v2.4.2
BibTeX 
@unpublished{wen2026liver,
  author = {Wen, Xuezhi and El Massaoudy, Hamza},
  title  = {Multimodal Classification of Primary Liver Tumors on
            Multiphase {CT}: Internal Performance, Acquisition
            Artifacts, and Limited External Transportability},
  note   = {Manuscript under revision at the Journal of Intelligent
            Medicine and Healthcare (Tech Science Press)},
  year   = {2026}
}

* corresponding author.

Case study · liver CTUnder revision

I built a multimodal classification pipeline using a MONAI 3D CNN and radiomics, then investigated the model's reliance on acquisition shortcuts.

My audit of 1,112 scans and counterfactual tests established its reliance on an acquisition artifact. Evaluation on 203 external patients exposed limited transportability across hospitals.

Corresponding author · tested Python pipeline · CI and release validation · code release v2.4.2.

Read the case study Code
Illustration: the same patients' portal-venous scans have about 3.5 times more slices than their non-contrast scans PORTAL VENOUS NON-CONTRAST
In the public dataset, cHCC-CCA non-contrast scans have about 3.5 times fewer slices than the same patients' portal-venous scans. Illustration, not to scale.

02 · Master's thesis

Trustworthy Lead Selection and Deferral for Atrial Fibrillation Detection with ECG Foundation Models under Distribution Shift

NUIST · supervisor: Wen Xuezhi · September 2026 to present, expected June 2027

Question

How can lead selection and deferral support more reliable atrial fibrillation detection when ECG data or signal quality changes?

Research scope

  • Study ECG foundation models for atrial fibrillation detection.
  • Investigate lead selection and selective prediction, including deferral to human review.
  • Evaluate robustness to distribution shift and signal noise.

Research status: Ongoing thesis; results are not yet public.

03 · Interests

Research interests

  • External validation and distribution shift

    Testing models on data from other hospitals, protocols and devices.

  • Shortcut learning and dataset audits

    Finding acquisition artifacts that models exploit.

  • Trustworthy prediction

    Calibration, selective prediction and deferral to clinicians.

  • Multimodal learning

    Combining imaging, radiomics, biosignals and clinical data.

  • Explainability with a clear scope

    Model and data cards, permutation importance, Grad-CAM.

  • Agentic AI for medical workflows

    A future direction I want to explore through research prototypes that help clinicians find and assess information, with human oversight.