Medical AI safety researcher: memorisation bias, privacy and rare-pathology detection in clinical imaging
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Not yet scored: the output axis measures publication volume and productivity over time, but this snapshot contains no aggregate publication counts — only two 2026 arXiv preprints in the Evolve catalog. The score is left empty rather than estimated.
Impact—
Not yet scored: the impact axis would aggregate citations and influence metrics for this body of work, but no citation data is present in the input. Evolve therefore reports no impact figure instead of extrapolating from only two 2026 preprints.
Focus100/100
The focus dimension scores 100. The underlying field distribution places the single field-classified paper entirely in 'machine learning for healthcare (medical AI safety)', and both 2026 preprints — memorisation bias and rare-pathology detection — sit inside medical AI rather than scattering across unrelated areas. The basis is narrow, so read this as consistent thematic concentration so far, not an established long-run pattern.
Collaboration50/100
The collaboration dimension scores 50. The snapshot records 11 distinct coauthors, with recurring partners including Ben Glocker, Daniel Rückert (also indexed as Daniel Rueckert), Mélanie Roschewitz, Johannes Kaiser, Martin J. Menten, Laurin Lux, Emma A. M. Stanley, Florian Braunmiller and Moritz A. Knolle. The pattern is compact, team-based medical AI research: he contributes as a middle author on the reviewed memorisation study and as last author on the mixture-of-experts preprint.
Activity—
Not yet scored: the activity axis tracks recent publishing rate and continuity over time, but the snapshot covers only 2026 preprints with no year-by-year history. A meaningful activity score needs a longer publication record than the current data provides.
Reproducibility100/100
The reproducibility dimension scores 100, computed from a single evaluated paper (reproPapers = 1, reproMean = 100). That paper, 'Memorisation bias in medical AI', is rated high on reproducibility readiness in Evolve's review, backed by rigorous large-scale evidence across ECG deep learning and chest radiograph classification. With only one paper evaluated, the perfect score is encouraging but provisional.
Research summary
Georgios Kaissis works at the intersection of machine learning and clinical medicine, with the catalogued output concentrated in machine learning for healthcare and, specifically, medical AI safety. The Evolve catalog holds two 2026 arXiv preprints. In 'Memorisation bias in medical AI' he is a middle author on a study Evolve's review describes as providing rigorous, large-scale evidence that training on patients' past records biases model predictions on their future data, while judging the resulting clinical harm modest; the paper spans ECG deep learning, chest radiograph classification, differential privacy and membership inference, and is rated high on reproducibility readiness. In 'Generalist-Specialist Mixture-of-Experts for Rare Pathology Detection in Multimodal Imaging' he appears as last author, the senior position, on work tackling rare pathology detection across multimodal imaging. The bibliometric snapshot is deliberately narrow: no institution, country or ORCID is attached, citation-based impact and output volumes are unscored, and the reproducibility score of 100 rests on a single evaluated paper. What the data does show is a tightly focused profile — 100 on focus, with the one field-classified paper sitting entirely in medical AI safety — and a compact collaboration network of 11 distinct coauthors, including Ben Glocker, Daniel Rückert and Mélanie Roschewitz, typical of team-based clinical machine learning research.
Collaboration pattern
Records list 11 distinct coauthors for this snapshot. Recurring partners include Ben Glocker, Daniel Rückert (also indexed as 'Daniel Rueckert'), Mélanie Roschewitz, Johannes Kaiser, Martin J. Menten, Laurin Lux, Emma A. M. Stanley, Florian Braunmiller and Moritz A. Knolle — a compact, stable medical-AI team. Kaissis appears as a middle author on the reviewed memorisation study and as last author on the mixture-of-experts preprint, consistent with a senior collaborator role.
Research keywords
Highlights
Evidence that medical AI memorises patients — modest clinical harm so far
Senior-author position on mixture-of-experts rare pathology detection
In 'Generalist-Specialist Mixture-of-Experts for Rare Pathology Detection in Multimodal Imaging' (2026, arXiv), Kaissis is listed as last author — the senior position in this catalog's convention. The preprint has not yet been reviewed by Evolve, so no verdict, keywords or reproducibility rating are attached; it counts here as catalogued output only.
Timeline
- 2026Appears as last (senior) author on 'Generalist-Specialist Mixture-of-Experts for Rare Pathology Detection in Multimodal Imaging' (arXiv preprint).
- 2026Co-authors 'Memorisation bias in medical AI' (arXiv), the first paper in this profile to receive a full Evolve review, rated high on reproducibility readiness.
- 2026The current bibliometric snapshot records 11 distinct coauthors, with recurring partners including Ben Glocker, Daniel Rückert and Mélanie Roschewitz.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Radar scores are computed from a public bibliometric snapshot (OpenAlex-sourced); no institution, country or ORCID is attached to this record, and the profile is name-resolved from public bibliographic data.
- Output, impact and activity are unscored because no publication counts or citation metrics exist in the input data.
- The reproducibility score of 100 (reproMean) rests on a single evaluated paper (reproPapers = 1).
- Only two 2026 arXiv preprints are catalogued here; the raw coauthor list contains the researcher's own name and two spellings of Daniel Rückert, so this is an early-stage, partially deduplicated profile.