A blank is not a finding
Open almost any referral letter and you will find a list of what the patient has. You will rarely find a list of what was checked and came back normal. The letter says the child has seizures and global delay. It does not say whether anyone looked at the eyes, listened to the heart, or measured the head. Those things are not in the letter, so on paper they are not there.
People reading the letter know this. They know a missing line means nobody wrote it down, not that the finding was excluded. Software that has to turn the letter into a table of yes and no usually does not know it. A field with no entry becomes a no, because a table has two kinds of cell and a blank has to go into one of them.
That small shortcut is where this post starts. In a case that has already been through testing once and stayed open, a lot of the remaining work is deciding which genes are still worth a look. Every blank that quietly turns into a no changes that decision.
Five states
The case record keeps five values for a phenotype feature, and they are different everywhere in the product, not just on the screen where you enter them.
Unknown, not tested and not provided are close, and a clinician may reasonably use more than one of them for different features in the same case. They are kept apart because they point at different next steps. An unknown might be settled by asking the family. A feature that was not tested may need an examination or an order. Something that was not provided may only need the referring team to be asked. None of them is a no.
Why matching cares
Matching a patient to a condition is a comparison of two lists: what the patient has and what the condition usually brings. A feature that the condition usually brings and the patient has is support. A feature the condition usually brings and the patient was checked for and does not have is evidence against. Those two outcomes are what the comparison is for.
Take a condition that usually involves the retina, and a patient whose eyes were never examined. If the record says absent, the condition is marked down for something nobody checked. If the record says unknown, not tested or not provided, the condition is not marked down, and the missing retinal finding shows up where it belongs: as something that would settle the question if someone looked.
The first version produces a tidy ranking that is wrong in a way you cannot see from the ranking. The second produces a ranking that is a little less decisive and a list of what is still open. We think the second is the honest one, and it is the one a clinician can act on.
Notes make it harder
Much of what ends up in a case starts as free text: a referral letter, a clinic note, a discharge summary. The text is full of hedges. “Possible hypotonia.” “No obvious dysmorphism.” “Seizures were reported but not witnessed.” A person reads those and knows how much weight each carries. A table has no room for the hedge.
The note reader in the product suggests terms from this kind of text, with the sentence each one came from. When a statement is hedged or unclear, it is flagged instead of being filed under present or absent, and the person chooses the state themselves. Nothing is pre-selected for those, because a default is a decision made on their behalf. Where the text is clear and the term is a precise one, it may be applied automatically. Those terms are marked as read from the notes, and a clinician can review them and reverse any of them.
The reader is a suggestion tool. It will miss things and it will sometimes read a sentence differently from the person who wrote it, which is why the sentence is shown next to the term and why a person confirms.
What it costs
Keeping five states is more work at the point of entry. A clinician who knows a feature is not there has to say so, rather than leaving it empty. A team that has not examined something has to leave it as not tested, which can look like an admission.
We think the work is worth it, and not only for matching. A case record that says what was never looked at is a better document for the next person to open it, whether that is a colleague next month or the same team after a new release. It also makes the gaps visible in a case review, where someone can say that nobody has looked at the heart yet and it should be done.
A short example, start to finish
A six-year-old has had seizures for a few months and some delay in development. The referral letter lists both. Nothing is said about muscle tone, the eyes, or the family. A common way to handle the letter is to enter the two findings as present and leave everything else empty, and then to treat everything empty as absent when matching.
With five states, the clinician enters seizures and delay as present. Tone is not in the letter, so it is unknown. The eyes were not examined, so they are not tested. Nothing came with the referral about the family, so that is not provided, and the parents may be able to answer it. When a few candidate genes are compared, none of them is marked down for features nobody checked, and the case shows a short list of what would help the most to find out.
The ranking in the second version is less sharp. We would argue that the sharpness in the first version was borrowed from assumptions, and that the clinic is better served by a ranking that knows what it does not know.
What the case record shows
On the case page, the states look different and are counted separately. A summary of the phenotype says how many features are present, how many were looked for and not found, and how many are open. Someone reviewing the case can see in a moment how complete the picture is, and where to put the next hour.
The same distinction carries into the reasons given for a candidate. A gene that is held back because a feature was excluded says so, and points at the entry. A gene that is not held back because the feature is unknown says that too.