All articles

Why possible hypotonia should stay possible.

Clinicians hedge for good reasons. A record that strips the hedge and stores a firm finding feeds the ranking something the examiner never claimed.

A paediatrician examines an irritable, crying eighteen-month-old at the end of a long clinic and writes 'possibly mildly hypotonic, difficult to assess today'. The letter goes to the genetics service. The genetic counsellor, filling in the laboratory's request form, has a pick list of HPO terms and no field for 'possibly'. Hypotonia goes on the form.

Two years later the term is still on the record. It has driven a candidate list towards neuromuscular genes in two analyses, it appears in the laboratory's summary as a cardinal feature, and the child, now three and a half, has normal tone on every examination since. No one ever examined her properly for it. The first clinician said as much.

What the field knows

Clinical findings carry degrees of confidence, and clinical writing has always expressed them. 'Possible', 'probable', 'query', 'cannot exclude' are the vocabulary of a profession that examines imperfect patients in imperfect conditions and knows it. The structured vocabularies built for genomics are newer and mostly binary. The Human Phenotype Ontology (Köhler et al., 2021) records a feature as present or, with an annotation flag, excluded; the GA4GH Phenopacket schema (Jacobsen et al., 2022) does the same, and neither carries a native field for the examiner's confidence in the observation.

The ranking methods inherit the binary. Phenotype-driven prioritisation (Smedley et al., 2015; Robinson et al., 2020) treats every present term as a fact and weights it by how specific it is, so an uncertain finding that happens to be specific, as hypotonia is for some neuromuscular conditions, pulls hard on the ranking. The methods have no way to pull less hard for a term the examiner was unsure of, because the input format has no way to say so.

Studies of phenotyping in practice have found both over- and under-coding. When clinicians' free-text notes are compared with the structured terms submitted for the same patient, terms appear that the notes hedged or never asserted, and terms the notes asserted are missing. The direction of the error follows the tooling: a pick list without a hedge produces firm terms from uncertain notes, and a request form with few fields drops the rest. Deep phenotyping guidance has argued for years that the quality of the term set, including its honesty about uncertainty, matters more to the result than its length.

Uncertainty is itself information. A finding the examiner marked as possible, in a child who was crying, is a reason to re-examine, which is cheap. A finding marked as present is a reason to sequence towards it, which is expensive and, if the finding was never there, misleading.

The same applies in the other direction. A feature marked as absent after a hurried look is weaker evidence than one excluded by a specialist examination, and a ranking that treats both as equal exclusions will discard candidates on the strength of a glance. Confidence belongs on absences as much as on presences.

Where practice breaks

The hedge is removed at transcription. The examiner wrote 'possibly'; the person completing the form had a tick box. The removal is invisible because the form looks complete, and from that point on the record states something the only person who examined the child did not.

The hardened term outlives the child's development. Tone changes in the first years; a finding at eighteen months in a distressed child is weak evidence about the child at three. A record that stores the term without the circumstances or the date lets it stand as a feature of the patient rather than of one examination, and later analyses inherit it with full weight.

The term then shapes the search. Neuromuscular candidates rise, the review time goes to them, and the second analysis repeats the first because the input did not change. If the child's real condition does not involve tone, the strongest single input into both rankings was a thing the first clinician said was doubtful.

Removing the term later is harder than adding it was. Once a feature is in the laboratory's record and in the case summary, deleting it feels like altering history. Most services have no state between present and gone, so the term stays, hedged in a comment nobody reads, firm in the field the software uses.

Counsellors are usually the people who notice. They read the letters, they see the child again, and they are the ones who realise that a feature the laboratory treats as cardinal was a single uncertain line. By then the request has been processed and the record belongs to a system they cannot edit.

How Tulparus handles it

The workbench reads the clinical text and keeps the clinician's confidence with the term. 'Possibly mildly hypotonic, difficult to assess today' becomes a finding with the state present, a confidence of possible, the examination date, and the sentence it came from, including the examiner's own caveat. The person confirming the intake sees all of it and can set the state to unknown with one action, which is often the honest one for a finding like this.

The phenotype ranker weights a possible finding less than a firm one, and shows that it did. On the hypothesis page, a candidate raised by a hedged term is marked as resting on it, so a reviewer can see that the neuromuscular genes are there because of one uncertain examination, and can ask for the examination to be repeated before anyone sequences towards them.

Findings carry their date and circumstances, and the workbench reads them against the child's age. An uncertain finding from a single early examination, with no later examination recorded, is listed as a gap: re-examine tone, which is the next most informative action and costs a clinic visit. If a later examination records normal tone, the earlier finding is superseded in the record, kept with its date, and the ranking is recomputed without it.

Nothing is deleted. The first clinician's observation remains in the history exactly as it was written, with its hedge, which is what it should have been all along.

What you can do tomorrow

Most of this is about protecting the hedge on its way from the letter to the form.

  • When the request form has no field for confidence, write the hedge into the free text beside the term: 'hypotonia (possible, single examination, child distressed)'. A reader can use it even if the software cannot.
  • Date every finding on the request, and say how many times it was observed.
  • Before a reanalysis, re-read the original letters against the submitted terms and strike the ones the letters did not assert. It takes half an hour and changes rankings.
  • When a hedged early finding is driving a candidate list, repeat the examination before ordering another test. It is the cheapest experiment in the case.

Sources

  1. Köhler S, Gargano M, Matentzoglu N, et al. The Human Phenotype Ontology in 2021. Nucleic Acids Research, 2021.
  2. Jacobsen JOB, Baudis M, Baynam GS, et al. The GA4GH Phenopacket schema defines a computable representation of clinical data. Nature Biotechnology, 2022.
  3. Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
  4. Robinson PN, Ravanmehr V, Jacobsen JOB, et al. Interpretable clinical genomics with a likelihood ratio paradigm. American Journal of Human Genetics, 2020.
  5. Robinson PN. Deep phenotyping for precision medicine. Human Mutation, 2012.

One hedged finding.See how much weight it carries.

A demo takes 20 minutes, screen shared, on synthetic cases. Bring the question from one of your own.