By Brent Dover, Chief Executive Officer, Carta Healthcare
An error in a registry has a peculiar property. At the moment it is made, it costs almost nothing and makes almost no sound. A value is mis-keyed, a date is misread, a finding is misinterpreted, and the case moves on. The cost arrives later, somewhere else, often when no one is looking at that case anymore. By the time bad data announces itself, it is rarely cheap to fix and sometimes impossible.
Understanding where that cost lands is the key to understanding why data quality in abstraction is not a back-office concern. It is a financial and reputational one.
It is tempting to think of registry submission as an end in itself, a box checked for a quality program. In reality the data flows outward into systems that carry real weight. It feeds reimbursement, accreditation status, public reporting, performance dashboards, quality programs, and population health strategy. Each of these consumes the data as if it were true. None of them re-derive it. The number an abstractor enters becomes the number a benchmark reports, the number an accreditation body reviews, and the number a payment formula trusts.
That is what makes a quiet error dangerous. It does not stay where it was made. It propagates. A single misclassified case can move a benchmark, distort a comparison against peer institutions, and shape conclusions that influence how a service line is run or judged. The error is small. Its blast radius is not.
The most expensive form of downstream cost is the audit. Registry data is auditable, and audits do not announce themselves in time to fix the underlying work. When an auditor arrives, the question is not whether the data felt reasonable when it was entered. It is whether each value can be defended against the source record, months or years after the fact. Data that was plausible but not rigorous tends to unravel under that scrutiny, and the unraveling carries financial consequence, remediation work, and reputational exposure that no efficiency gain on the front end can offset.
This is why throughput alone is a misleading measure of an abstraction program. A program can look productive, moving cases quickly and clearing backlogs, and still be accumulating a liability that only surfaces under audit. The relevant question is not how fast the data was produced. It is how well it holds up when someone with authority examines it line by line.
When abstraction is done to a high standard, accuracy is sustained in the range of ninety-eight to ninety-nine percent, and that range is not a vanity metric. The last percentage points are the hardest and the most valuable, because the cases that fall in them are the ambiguous, judgment-heavy ones that audits and benchmarks are most sensitive to. A system that is accurate on easy cases and shaky on hard ones can post a respectable average and still fail exactly where failure is costly. Defensibility lives in the hard cases, not the average.
There is also a compounding effect to quality. Good abstraction does more than avoid errors. It surfaces them in the underlying documentation. When a clinician notices that a record is incomplete or that a definition is being applied inconsistently, the organization can correct the documentation and tighten its processes. Over time this raises the quality of the source record itself, which makes every future submission more reliable. Poor abstraction misses these signals, and the documentation problems persist unaddressed, quietly seeding the next round of downstream cost.
The economics here point in one direction. The cheapest place to ensure data quality is at the moment of abstraction, when the case is fresh, the source is at hand, and a knowledgeable person can resolve the ambiguity correctly. Every step downstream makes correction more expensive, because the error has by then been built into benchmarks, submissions, and decisions that are costly to revisit. Spending judgment at the point of abstraction is not an added cost. It is the avoidance of much larger costs later.
This reframes how a health system should evaluate an abstraction approach. The right question is not only how much the process costs or how fast it runs. It is how much downstream risk the process prevents. A model built around accuracy, defensibility, and the resolution of hard cases is buying something that does not appear on this quarter's efficiency report. It is buying the absence of an expensive surprise in a future audit, a stable benchmark, and reimbursement that rests on data that holds.
If errors are cheapest to catch at the moment of abstraction, the natural question is how to catch them there reliably. One of the most effective answers is inter-rater reliability, the practice of having more than one qualified person review the same case and comparing the results. Disagreement between two careful reviewers is a signal. It points to the ambiguous cases, the inconsistent definitions, and the documentation gaps that would otherwise slip through unnoticed and reappear, far more expensively, in an audit. Treated as a routine part of the workflow rather than an occasional spot check, inter-rater review functions as an early warning system for exactly the failures that cost the most downstream.
It helps to remember how high the stakes climb once data leaves the abstraction step. Registry data feeds public reporting, where it shapes how an institution is perceived. It feeds accreditation review, where it influences a status that can take years to earn and moments to jeopardize. It feeds reimbursement, where it touches revenue directly. A value that was entered without rigor travels into all of these with the same authority as one that was carefully reasoned, and none of these downstream systems will flag the difference. They trust the number. That trust is precisely why the number has to be earned at the source.
Small errors also compound in ways that are easy to underestimate. A single misclassification is a problem. A consistent pattern of the same misclassification, repeated across hundreds of cases because no one caught the underlying definitional drift, is a systemic distortion that can quietly reshape a benchmark. Quality at the point of abstraction is not only about the individual case. It is about preventing the small, repeated error from hardening into a false picture of how an institution performs.
The uncomfortable implication for leaders is that a clean dashboard is not evidence of clean data. Errors at the point of abstraction do not announce themselves on the report that summarizes them. They are smoothed into averages and rolled into trends, and they look exactly like correct data until something forces a closer look. The only reliable protection is to invest in quality where the data is made, through qualified review and structured cross-checking, rather than to trust that problems will surface on their own. By the time they surface on their own, they have usually become expensive to fix.
Bad data is patient. It waits. The organizations that take data quality seriously at the point of abstraction are the ones who are not surprised by what it was quietly costing them all along.
