All articles

The negative finding nobody wrote down.

An examined and normal finding argues against a whole class of candidates. Most request forms have nowhere to put it, so the laboratory never sees it.

A boy of four is referred for exome sequencing with developmental delay, seizures from the second year and an unsteady gait. The request form carries those three terms. His neurology letter, written a month earlier, is more careful. It records that the fundi were examined and were normal, that a hearing test at three was normal, that there is no organomegaly and that head circumference has tracked the fiftieth centile since birth. None of that reached the laboratory.

The exome was negative. When the ranked candidate list was reviewed later, three of the ten genes at the top caused conditions in which retinal dystrophy, hearing loss or an enlarged liver would be expected by his age. Each had been examined for. Each had been found normal. The list had been built without any of it.

What the field knows

The pertinent negative is older than genomic testing. A clinical examination is a sequence of questions put to the body, and a normal answer narrows the differential as surely as an abnormal one, which is why a case presentation lists what was absent beside what was present. What has changed is that candidate ranking is now done by software that compares the patient's recorded terms with the terms annotated to each disease, and software weighs only what it is told.

The vocabulary for a normal finding exists at both ends. The Human Phenotype Ontology (Köhler et al., 2021) annotates diseases with the features they show, and its annotation format allows a feature to be marked as excluded for a disease as well as present. The GA4GH Phenopacket schema (Jacobsen et al., 2022) does the same on the patient side: each phenotypic feature carries a flag that says whether it was observed or explicitly excluded.

Some ranking methods use it. LIRICAL (Robinson et al., 2020) computes, for each candidate disease, a likelihood ratio for every feature the patient has, and also for every feature the patient was examined for and found not to have. An excluded feature that the disease nearly always shows counts heavily against that disease, and the authors show excluded features moving rankings in simulated and real cases. Exomiser (Smedley et al., 2015), the most widely used tool of its kind, scores the features that are present; its similarity measure has no term for a negative, so a disease that predicts a feature the patient demonstrably lacks is not moved down.

There is a reason the methods that ignore negatives became the norm. In most clinical datasets there are almost no negatives to use. Where HPO coding in practice has been described, the recorded terms for a patient are few, general and nearly all positive, because the forms that produced them had nowhere to put a normal finding. When the inputs never contain a negative, a method that could use one has nothing to work with, and a method that cannot loses nothing by it.

Where practice breaks

A normal finding does not feel like a finding to the person who made it. The ophthalmologist who examines the fundi and sees nothing wrong writes 'fundi normal' in a letter to the referring paediatrician, and the matter is closed. Nobody carries it to the request form, because the form asks what is wrong with the child and this is a thing that is right.

The forms make it worse. A free-text box, or a pick list of HPO terms, offers no way to say that a term was looked for and excluded. Some laboratories ask for a clinical summary and receive a paragraph; a scientist reads the paragraph and converts it to present terms, and the normals in it are dropped because there is no field to drop them into. The letter in the case above said four useful things that the laboratory never saw, and the loss happened in the ordinary course of good practice, with nobody doing anything wrong.

The downstream cost is specific. A ranking that cannot see negatives puts genes at the top that a glance at the letter would have removed. The scientist spends the review on those genes. The genes that fit everything the child has and nothing the child lacks sit lower, and in a long list they may not be reached. When the result is negative, the next reanalysis starts from the same positive-only record, and the same genes come back to the top.

There is a second loss, less visible. A negative finding has a date. Fundi normal at four says nothing about the fundi at nine, and some retinal dystrophies declare themselves late. A record that stores 'no retinal dystrophy' as a plain fact, without when it was established, will mislead the clinician who reads it five years on. The negative has to carry its age, which is one more reason a form built for a single moment cannot hold it.

How Tulparus handles it

The workbench reads the letters rather than the form. Clinical text is read for phenotype terms, and a sentence that says the fundi were normal becomes a finding with the state absent, with the sentence it came from and the date of the letter attached. The clinician confirming the intake sees each absent finding beside the present ones and can accept, change or reject it. An absence is never inferred from silence: if no letter mentions hearing, hearing is unknown, and the record says so.

The phenotype ranker scores both directions. A present finding that a candidate's disease predicts raises the candidate. An examined-and-absent finding that the disease predicts lowers it. The three genes at the top of the boy's list would have moved down, with the normal fundus examination named as the reason beside each. A clinician who thinks the examination was too early to exclude the feature can mark it so, and the ranker treats the feature as unknown again.

Absent findings also age. Every finding carries the date it was established, and the workbench shows a negative's age when the candidate it argues against is one whose feature appears late. On reanalysis, a negative older than the window in which the feature usually appears is flagged for the clinician to re-examine rather than carried forward as settled.

What was never examined stays visible as a gap. If a leading candidate predicts a feature and no finding, present or absent, is recorded for it, that gap appears as missing evidence with the examination that would close it. The ranker orders candidates for a person to review, and it never states the chance that one of them is the answer.

What you can do tomorrow

Most of the loss happens before any software is involved, and the clinic can stop it.

  • Write the pertinent negatives on the request, in words. 'Fundi normal, March 2026' is enough. A laboratory that cannot code it can still read it.
  • Date every negative. A normal examination without a date is a fact about a child who no longer exists.
  • Distinguish not examined from normal. If hearing was never tested, say so. The two are opposite pieces of information, and a form that merges them destroys both.
  • When a negative exome comes back, ask the laboratory which candidates were near the top and whether any predicts a feature your examination has excluded. That conversation has reopened cases.

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. Robinson PN, Ravanmehr V, Jacobsen JOB, et al. Interpretable clinical genomics with a likelihood ratio paradigm. American Journal of Human Genetics, 2020.
  4. Smedley D, Jacobsen JOB, Jäger M, et al. Next-generation diagnostics and disease-gene discovery with the Exomiser. Nature Protocols, 2015.
  5. Robinson PN, Köhler S, Bauer S, Seelow D, Horn D, Mundlos S. The Human Phenotype Ontology: a tool for annotating and analyzing human hereditary disease. American Journal of Human Genetics, 2008.

One normal examination.See which candidates it removes.

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