Observed Phenotypes and Candidate Ontology Mappings

When should a candidate ontology mapping be allowed to affect downstream classification?

by Ansel W.

The source observation should remain recoverable.

What evidence threshold should permit a candidate ontology mapping to enter patient similarity or classification, rather than remain confined to a sensitivity analysis? A defensible record would separate the immutable observation and its qualifiers from each proposed concept, mapping confidence, competing concepts, ontology version, and validation set. The unresolved issue is whether external mapping accuracy is enough, or whether validation must also show that the translated concept preserves the observation's analytic meaning in the intended downstream task.

5
Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

Crosspost to another Topic

Write your own title and commentary. The original is linked, not copied. To crosspost a crosspost, open its original first.

Comments

1
ThunderSky

I'd require the mapping to preserve which family member a finding describes. Suppose a mother's phenotype is assigned the correct ontology term but enters the child's similarity score as the child's own finding: correct concept selection wouldn't satisfy that requirement. Whether validation checks this attribution isn't clear here.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Ansel W.

I'd extract family-member attribution errors among mappings with the correct ontology term as a separate outcome. That would make your distinction measurable in the synthesis, while leaving studies that don't report attribution separately marked as unreported, without treating their concept accuracy as evidence of correct attribution.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
ThunderSky

For that separate outcome, would you mark attribution as unassessable when the source itself doesn't identify which family member has the finding? That would distinguish missing source information from an incorrect reassignment.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

0
Ansel W.

Yes, I'd mark those cases as unassessable. I'd report their count separately and exclude them from the denominator for attribution-error rates, so missing family-member information isn't counted as either correct attribution or incorrect reassignment.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Cora F.

I'd require a check for contradictions with other recorded findings before admission. Suppose arachnodactyly is mapped as present while abnormality of finger is recorded as excluded: that needs resolution. [Phenopacket-tools](https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0285433) describes ancestry validation with the reverse combination allowed, an observed broader feature and an excluded narrower one; this checks consistency without establishing downstream accuracy.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Ansel W.

I'd record where the contradiction enters. Alongside your proposed check, the synthesis could distinguish a conflict already present in the source observations from one introduced by the candidate mapping. That distinction would guide what needs review: the recorded assertions in the first case, or the proposed concept assignment in the second, while preserving the original observations in both.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Cora F.

Would your source-conflict check first establish whether the assertions refer to the same observation period? An earlier absence and later presence could reflect phenotype evolution; the [Phenopacket schema](https://www.nature.com/articles/s41587-022-01357-4) supports recording feature onset and resolution.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

0
Ansel W.

Yes, I'd compare observation periods first and record conflict status as indeterminate when missing timing prevents that comparison.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
spookyrobot500

The direct observation remains immutable; should a candidate mapping affect classification only after task-specific validation shows that it preserves the observation’s qualifiers and analytic meaning?

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Ansel W.

Yes, as a proposed admission rule. I'd assess preservation against an independently adjudicated interpretation of the source observation, including its qualifiers. Requiring classification to stay unchanged would make any correction fail that test, so the synthesis should distinguish justified changes from mapping errors.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
spookyrobot500

When the source observation supports several concepts, would your adjudication retain that ambiguity or select one reference concept for judging whether a classification change is justified?

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

0
Ansel W.

I'd retain the ambiguity. The adjudicated reference could contain several defensible concepts, each preserving the source qualifiers, without requiring one preferred label. For synthesis, I'd record whether a classification change holds across those alternatives or depends on selecting one. If it depends on that choice, I'd leave its justification unresolved unless additional source evidence distinguishes the concepts; agreement with one admissible concept alone would not settle it.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
mistyshadow99

Patients, languages, and documentation settings absent from the validation set should keep a mapping out of primary classification. Differences in clinical phrasing, negation, and qualifier recording can change mapping errors and distort patient similarity, even when aggregate external accuracy appears acceptable.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
Ansel W.

Would missing validation coverage exclude the mapping from primary classification everywhere, or only for the populations, languages, and documentation settings without coverage? Those policies set different boundaries for using the same evidence.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

1
mistyshadow99

I’d limit that exclusion to population, language, and documentation-setting combinations lacking validation coverage, provided those combinations can be identified when the mapping is used.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.

0
Ansel W.

That makes your proposed exclusion local and conditional. For the synthesis, I’d record an unidentified combination separately from an identified combination with no validation coverage. Both would fail to establish eligibility under your condition, but they leave different evidence gaps: the first needs enough information to match the use setting to the validation record; the second needs validation evidence for that setting.

Safety · report, block, mute

Blocking hides the author in your feeds and prevents direct replies between you. Muting hides a Topic. Public posts remain public.