All articles

Similar patients, without sharing patients.

The strongest evidence for a rare diagnosis is often another patient like this one. Finding them can be done without moving anyone's record.

Two laboratories three hundred miles apart each hold an unresolved case: a child with a particular combination of a skeletal dysplasia, a retinal finding and a hearing loss, and a de novo variant in the same gene with no disease association. Each laboratory considers the gene a weak candidate on its own. Neither knows the other exists. Each has filed the variant as a finding of uncertain significance and moved on to the next referral.

Together the two cases would be a publishable new gene-disease association and two diagnoses. Apart they are two negative reports. The information that would join them is a phenotype profile and a gene name, and neither laboratory can send either without a data sharing agreement that would take a year to sign.

What the field knows

Matching patients to patients has been one of the most productive moves in rare disease genomics. The Matchmaker Exchange (Philippakis et al., 2015) connected databases of candidate genes and phenotypes across institutions through a common protocol, so that a submission in one node could be matched against the others without the underlying records leaving their home. Boycott and colleagues (2017) reviewed the discoveries that followed, many of them new gene-disease associations established by two or three families found through a match. The design principle was explicit from the start: the match is made on a minimal profile, the result is a notification that a potential match exists, and the clinicians then decide what to share.

Phenotype similarity between patients is computable without identifiers. Semantic similarity over ontology terms (Köhler et al., 2009) compares two sets of HPO terms and returns a score, and the score can be computed from the terms alone, with no name, date of birth or free text. A gene name and a set of phenotype terms are not anonymous in any strong sense, since a rare combination can identify a family to someone who knows it, which is why the matching services treat the profile as sensitive, limit who can query and log every match.

The governance has matured alongside the methods. The GA4GH frameworks for responsible sharing set out consent tiers, access controls and audit for exactly this kind of exchange, and the matching networks operate under them. The lesson from a decade of operation is that a well-designed match, minimal profile in, notification out, has been acceptable to ethics committees, laboratories and families in a way that pooling records has not.

Where practice breaks

Submission is optional and sporadic. A match requires both laboratories to have submitted, and most unresolved cases are in neither network, because submission is a separate task, in a separate system, done by the scientist who remembers. The two laboratories in the opening case could have found each other through an existing network and did not, because neither had submitted the case.

Similarity within a service is not computed either. A laboratory that has seen a similar patient before, in its own records, usually does not know it, because its cases are documents and spreadsheets, and nobody runs a phenotype comparison across them. The second patient is sometimes down the corridor, seen by a colleague two years earlier under a different working diagnosis, with the same gene sitting in a spreadsheet nobody has opened since.

Where sharing is attempted outside the networks, it is done by email, with more information than the match needs. A clinician describes the patient in a paragraph to a colleague at a conference; the paragraph contains the age, the town and the sibling's condition. The exchange that was too sensitive to formalise happens informally, with worse controls.

The match, when it comes, is detached from the case. A notification arrives in an inbox, months after submission; the case has to be found, the match assessed by hand against a candidate list that has moved on, and the outcome recorded nowhere. Matches are missed because they arrive at the wrong time to the wrong person, and a match that nobody acts on is indistinguishable, in the records, from one that was never made.

How Tulparus handles it

Similar-case retrieval runs within each organisation's own cases as a candidate channel. Every open case is compared, on phenotype terms and candidate genes, against the organisation's other cases, and a strong similarity is surfaced on both cases with the shared terms and genes named. The second patient down the corridor is found without anyone remembering, and the comparison is logged on both cases so that the scientist who picks either one up later can see it was made and what came of it.

Across organisations, matching is an explicit, consented exchange. An organisation that opts in has its open cases matched against other opted-in organisations on the minimum the match needs, the phenotype terms and the candidate gene, within the workbench and without either record leaving its tenant. A match returns the fact of a match, the shared terms and gene, and a route to the other organisation's named contact. It never returns the other case. The two clinicians decide what to share next, under their own governance, and the workbench records that the match was made and what followed.

Submission to the public matching networks is prepared from the case, with the same minimal profile, where the laboratory has chosen to use them, and the submission identifier is kept on the case so that a match from the network arrives on the case and not in an inbox. The match is assessed against the current candidate list, and the assessment is recorded as evidence.

Every match query and every match is logged, tenant-scoped, and available to the organisation's own audit. For the two laboratories above, the match would have been made the week the second case was opened, each would have seen a shared gene and three shared terms with a contact at the other, and the conversation that produced two diagnoses would have started a year earlier.

What you can do tomorrow

The networks exist, and the gap in most services is that they are used for a few memorable cases when they should be used for every unresolved one. None of the steps below needs new software.

  • Submit every unresolved case with a candidate gene to a matching network as a routine step at sign-off, and record the submission identifier in the case note.
  • Run a phenotype similarity comparison across your own unresolved cases once a quarter. Any free tool that computes HPO similarity will do.
  • Agree what you will share after a match, in advance, so that the conversation with another laboratory does not wait on a policy.
  • Route match notifications to a service address rather than to an individual, and review them against the case monthly.

Sources

  1. Philippakis AA, Azzariti DR, Beltran S, et al. The Matchmaker Exchange: a platform for rare disease gene discovery. Human Mutation, 2015.
  2. Boycott KM, Rath A, Chong JX, et al. International cooperation to enable the diagnosis of all rare genetic diseases. American Journal of Human Genetics, 2017.
  3. Köhler S, Schulz MH, Krawitz P, et al. Clinical diagnostics in human genetics with semantic similarity searches in ontologies. American Journal of Human Genetics, 2009.
  4. Knoppers BM. Framework for responsible sharing of genomic and health-related data. HUGO Journal, 2014.
  5. Erlich Y, Narayanan A. Routes for breaching and protecting genetic privacy. Nature Reviews Genetics, 2014.

Two cases, one gene.See a match that moves no records.

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