Collaborative working manuscript · Not peer reviewed
Oralbiome
A retrospective audit of cross-study transfer
A retrospective audit of oral-microbiome diabetes-status models, focused on cohort eligibility, leakage, calibration and cross-study transfer.
Research question
How does apparent within-source performance change across study populations, and what does the available evaluation establish about transfer?
Work and approach
I contributed study design, baseline and CatBoost analyses, interpretation, transfer investigations and manuscript preparation. My collaborator contributed preprocessing, exploratory analysis, feature engineering and a gradient-boosting comparison. We both contributed to methods and experiments, using AI tools throughout the workflow.
- Audit specimen eligibility and related-study overlap before interpreting predictive performance.
- Use nested model selection, disjoint calibration and out-of-fold subgroup reconstruction.
- Report cohort-specific uncertainty and distinguish exploratory transfer from untouched validation.
How to evaluate the work
The retrospective reconstruction excludes mislabeled fecal specimens and related-study overlap. All four displayed external performance intervals include chance, and all external cohorts informed development. Those facts limit what the apparent source-cohort performance establishes.
Evidence and limitations
The corrected public working manuscript is Oral Microbiome Models of Diabetes Status: A Retrospective Audit of Cross-Study Transfer. NHANES labels describe self-reported diabetes, without verified Type 2 subtype. The manuscript is not peer reviewed.
This is exploratory retrospective research, with small heterogeneous external cohorts and unresolved source/linkage limits. It does not establish an untouched confirmatory evaluation, a clinically useful tool or a causal biological mechanism.