Algo-Model

Research decision record — 2026-08-04: professor feedback on scope and method

Status: accepted, implemented in this repository on 2026-08-04. Scope: MetaboGuard research direction, evidence handling, and analysis method.

These are recollected notes, written after the meeting from memory. They are not a transcript and contain no verbatim quotations. Where a statement is our own inference or engineering consequence rather than something a supervisor said, it is marked (our interpretation). Corrections from attendees are welcome and should be added as a dated amendment below rather than by rewriting this record.

Attendees referenced

• Prof. Helmount — early-stage detection and clinical value requirement.
• Prof. Nada — clustering / phenotype discovery recommendation.

Recollected feedback

1. Prof. Helmount — clinical value depends on early-stage detection

Recollected substance:

• A model only has clinical value if it contributes at a stage where intervention is still possible. Separating people who already have a diagnosis is not clinically useful.
• Early cancer detection realistically depends on panels of markers and interacting features, not on one universal marker that works across cancers.
• Any claim about detection must be tied to the stage and lead time at which the evidence was obtained.

Explicitly not claimed, and not to be written anywhere in this repository (our interpretation of the correct scientific statement):

• It is false to say that “no cancer has any specific biomarker”. Site-specific markers exist and are clinically used (for example CA19-9 in pancreatic disease, with the documented sensitivity/lead-time limits recorded in our evidence catalogue). The defensible statement is that no single marker is universally sufficient for early detection across cancers, and that panels plus interacting features are required.

2. Prof. Nada — use clustering to discover phenotypes

Recollected substance:

• Rather than forcing a supervised outcome the data cannot support, run unsupervised clustering to look for patient/metabolic phenotypes.
• Disease labels may then be used to characterise the resulting clusters, but not to fit or select them.

Consequences (our interpretation):

• Clusters are patient/metabolic phenotypes, never cancer diagnoses, cancer subtypes or cancer sites. Naming a cluster after a disease is prohibited in code, API and UI.
• Label use is strictly post hoc and must be reported as cross-sectional association with an already-recorded diagnosis, suppressed when class counts are inadequate.
• Clustering on survey data will happily rediscover the survey rather than biology, so negative controls are mandatory: any cluster solution dominated by survey cycle, assay availability, missingness burden, age or sex is reported as a data artefact.

Decisions taken

#	Decision	Rationale
D1	Terminology: no causal phrasing. Use risk-associated features, early-development signals, biological pathways. causal is reserved for the DoWhy estimation module, where it names a method, not a finding.	Nothing in the current data supports causal claims.
D2	Early detection is framed as a panel + interaction problem; the repository must never state that cancers have no specific biomarkers.	Scientific accuracy; see above.
D3	Clustering is exploratory phenotype discovery, label-free in fit and selection, with an explicit abstain result when stability criteria fail.	A clustering that is not stable is not a finding.
D4	An evidence catalogue with mandatory provenance (URL/DOI, study design, validation status, evidence grade, limitations) gates any doctor-facing statement. Unknowns are recorded as unknown, never inferred.	Prevents evidence drift and fabricated citations.
D5	A data reliability report with feature eligibility tiers (usable_now, qualified_use, unavailable, prohibited) runs before analysis and fails closed on hard violations.	Reliability must be a gate, not a footnote.
D6	Future-risk and cancer-site outputs stay disabled and fail-closed; NHANES here is cross-sectional and TCGA is post-diagnosis context only.	Unchanged from the pre-existing capability gates.
D7	Clinician-facing explanations must label every statement as data observation, model association, published evidence, or causal claim not established.	Keeps the epistemic status visible at the point of reading.

What this does not change

• No new cancer-site prediction, no future-risk head, no re-enabling the invalidated pancreatic-cancer supervised artifacts.
• The self-supervised encoder remains frozen and label-free; clustering consumes its representation, it does not retrain it.

Blocker that this feedback makes sharper, not smaller

Prof. Helmount’s early-stage requirement cannot be met with the current data at all: there is no follow-up time, no incident outcome and no stage information in the cross-sectional NHANES files, and TCGA is post-diagnosis. Cluster phenotypes are therefore a hypothesis generator for a future longitudinal study, not evidence of early detection.

Amendments

(none yet — append dated entries here rather than editing the notes above)

This site is open source. Improve this page.