Insights
Data trust, provenance, and what AI should be allowed to act on. Written by Jason Alan Snyder, nearly every day. Longer essays live on his Substack.
Prescription Drug Monitoring Programs report data to 49 states, but latency, format mismatches, and missing dispensation records undermine every AI safety tool that relies on them. Without trust scoring at the record level, opioid prescribing algorithms inherit the errors PDMP databases have carried for years.
Co-located behavioral health integration fails when PHQ-9 scores, substance use disorder records, and psychiatric diagnoses live in separate EHR instances with no shared identifier. Fewer than half of integrated primary care sites achieve reliable data exchange between behavioral health and medical records, and the downstream cost hits quality measurement, reimbursement, and patient safety.
Safety net hospitals lose an estimated $5.7 billion annually in disproportionate share hospital payments they cannot fully justify because the underlying data fails on provenance, recency, and concordance. The Medicaid fraction and low-income utilization rate calculations that drive DSH allotments depend on data elements that are systematically incomplete at the institutions that need the funding most.
HRSA requires 1,400+ health centers to submit Uniform Data System reports annually, covering clinical quality, demographics, financial performance, and social determinants. Each UDS table depends on data integrity standards that most EHR systems do not enforce natively, creating a trust gap between what gets reported and what actually happened at the point of care.
There are 1,363 critical access hospitals in the United States, and the data they report carries structural problems that most quality frameworks ignore. Small denominators, legacy EHR systems, and staffing gaps make CAH data fundamentally different from urban hospital data. Using it without adjustment produces misleading benchmarks, flawed AI models, and policy decisions that accelerate the closures they claim to prevent.
Since 2010, more than 150 rural hospitals have closed across the United States, and over 700 more face financial risk. The data systems meant to track their community health impact are fragmented, stale, and missing the dimensions that matter most for intervention planning.
Housing and health data sit in separate systems governed by separate laws, creating a trust gap that no data-sharing agreement alone can close. Cross-sector integration requires scoring every record for provenance, consent, and recency before it enters a model. Without that foundation, social determinants housing intelligence remains unreliable.
This is not a stunt with a fly. It is a test of what intelligence is made of, and what happens when you hand any kind of intelligence bad data. We ran medical records through a fruit fly's brain and through four big AI models. Nature won. Here is why that matters for anyone whose health depends on a record being right.
Food as medicine programs generate data across clinical encounters, community-based organizations, and grocery delivery platforms, but fewer than 20% of programs track outcomes with structured, linkable data. Without trust scoring on nutrition intervention records, the field cannot prove what works, for whom, or at what dose.
SDOH Z-code capture rates sit between 1% and 2.4% of claims nationally, despite screening programs running in thousands of facilities. The gap between screening a patient and producing a coded, trustable record is where most SDOH data programs fail. Understanding what Z-code capture rates actually measure, and what they miss, is the first step toward building social determinants intelligence that AI systems can act on.
Care management outreach fails when the targeting data is wrong. Roughly 30% of outreach attempts never reach the intended patient due to stale contact information, misattributed risk scores, and incomplete social determinant records. Population health targeting requires data trust infrastructure that most health plans and systems do not yet have.
Nearly one in five hospital discharges results in an adverse event within 30 days, and the root cause is rarely clinical. It is data: missing medication lists, stale provider directories, unverified post-acute capacity, and consent records that never followed the patient. Discharge planning data quality determines whether care transition intelligence works or fails silently.
Concurrent review decisions depend on data that changes hour by hour, yet most utilization management systems treat clinical records as static inputs. When 30% of UM denials trace back to incomplete or stale data rather than clinical judgment, the integrity of the underlying record becomes the central problem. This post maps what concurrent review data actually requires across provenance, recency, and concordance.
Prior authorization AI systems approve or deny care based on data they never verified. When 50% of provider directories contain errors and claims data lags 30 to 90 days, automation without data quality controls produces automated harm. The question is not whether AI can do prior auths. The question is whether the data feeding those decisions meets a minimum trust threshold.
CMS audits consistently find error rates approaching 50 percent in Medicare Advantage provider directories. That means half the records a member sees when searching for a doctor may list the wrong phone number, wrong address, or wrong acceptance status. The problem is not a technology gap. It is a data trust gap, and fixing it requires scoring every provider record before it reaches a directory.
CAQH holds credentialing data on more than 1.6 million providers, yet health plans routinely find that profiles are incomplete, outdated, or contradicted by primary sources. The database was designed to reduce paperwork, not to guarantee data integrity. That gap costs the credentialing system billions in rework, delays, and downstream errors.
Most health data onboarding pipelines copy records into a new environment before scoring them. Onboarding Studio does not. It scores, validates, and governs data at the source, producing a DTI trust score and a hash-chained receipt without ever moving the underlying record. That distinction determines whether your AI pipeline meets regulatory scrutiny or fails it.
Health AI systems that cannot trace every record back to its origin are making clinical decisions on faith. Explainable data provenance assigns a verifiable chain of custody to each data point before it enters a model, and the 25% weight provenance carries in the DTI framework reflects its foundational role. Without it, explainability stops at the model and never reaches the data.
Hospital Compare publishes quality data on over 4,000 hospitals, but the underlying data suffers from coding inconsistencies, risk adjustment gaps, and reporting lags that distort the outcomes consumers see. CMS public reporting data trust depends on layers of transformation that few users understand, and fewer question.
NCQA requires health plans to meet strict data quality thresholds across HEDIS measures before star ratings are calculated. A single missed mammogram record or misaligned value set can drop a plan's rating by half a star, costing millions in Quality Bonus Payments. The gap between collecting data and trusting data is where most health plans fail.
Forty-two states now contract with managed care organizations to administer Medicaid benefits, and every one of those states imposes reporting requirements that assume the underlying data is trustworthy. Most MCO data fails that assumption before it reaches CMS. State Medicaid reporting requirements create a compliance surface that cannot be satisfied by volume alone; MCO data quality depends on provenance, recency, and validation that most plans never measure.
Medicare Advantage plans receive an estimated $12-25 billion annually in excess payments tied to risk adjustment coding practices that inflate patient severity. When AI systems train on this data, they inherit the financial incentives baked into every diagnosis code, producing models that confuse billing optimization with clinical reality.
ACO REACH participants must report on 3 quality measures and satisfy CMS data submission requirements across claims, clinical, and beneficiary-level records. But meeting minimum reporting thresholds is not the same as having trustworthy data. Without a trust infrastructure that scores every record for provenance, recency, and completeness, REACH model participants are building risk-bearing contracts on data they cannot verify.
The CMS Innovation Center has launched over 50 value-based care models since 2010, and every one of them requires participating organizations to submit clinical, financial, and operational data that meets specific quality thresholds. Most participants underestimate what those data requirements actually demand, and the gap between what CMMI expects and what health systems can deliver is where model performance collapses.
