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.
When a patient withdraws consent, most health AI systems have no mechanism to propagate that revocation to downstream models already trained on the data. The gap between consent withdrawal and actual data removal creates legal exposure, trust erosion, and model integrity failures that compound across every system in the chain.
Most consent systems treat authorization as a one-time event. When a patient changes their consent preferences in real time, every downstream dataset, model, and analysis that depends on that data faces an immediate trust problem. Dynamic consent architecture solves the capture problem but creates a data integrity crisis that almost nobody is tracking.
Most health data consent frameworks capture a single moment of agreement, then allow that data to flow through unlimited downstream uses. A 2023 ONC study found that 79% of patients did not know their clinical data had been shared with third parties for secondary purposes. Consent layering, the practice of stacking new data uses on top of original permissions, is the structural failure that erodes health data secondary use trust.
Eviction records exist in over 3,100 county court systems with no standard format, no shared identifiers, and no reliable method for linking to health data. When health systems and Medicaid programs try to integrate housing instability data with clinical records, the result is a trust problem that undermines every downstream AI model, risk score, and care coordination workflow.
Claims-based proxy measures for transportation barriers misclassify up to 40% of patients who actually face access problems. Area-level deprivation indices, ZIP code distance calculations, and missed appointment flags each introduce distinct distortions that compound when fed into population health AI. Building a transportation barrier data trust requires patient-level verification, temporal specificity, and provenance scoring that proxy data cannot provide.
Over 88% of food insecurity screenings captured in EHRs fail to use the FHIR resources required for interoperable SDOH data exchange. Without proper Observation, Condition, and ServiceRequest mappings aligned to the Gravity Project value sets, Z-code data for food insecurity becomes clinically useless and analytically toxic. This post maps the exact FHIR resource requirements, the data quality failures that degrade food insecurity screening data, and what trust scoring reveals about the gap between collection and usability.
Geocoding errors misassign up to 10% of patient addresses to the wrong census tract, silently corrupting every SDOH variable linked to that location. When health AI models consume area-level deprivation indices without verifying the geocoding step, they inherit systematic bias that disproportionately affects rural and minority populations. Address-level SDOH data trust requires scoring the geocoding pipeline itself, not just the social determinant variables it produces.
The AHC-HRSN screening tool captures social risk factor data across five core domains, but most health systems collect it without any quality framework. Without scoring for provenance, recency, and concordance, SDOH screening data fails the moment it enters an AI model or a CMS reporting pipeline.
Up to 30% of race and ethnicity fields in EHR data are either missing or inferred from surnames and geocoding rather than self-reported by patients. When AI models train on this data without distinguishing source method, they replicate and amplify the very disparities health equity programs aim to fix. Understanding the difference between self-reported and inferred demographic data is a prerequisite for any clinical AI deployment.
Patient-entered health data loses clinical reliability at a measurable rate. Within 90 days of collection, self-reported medication lists, symptom logs, and health histories show concordance drops of 20-50% against verified clinical records. Understanding this decay curve is essential before any AI model trains on patient-reported inputs.
Patient-reported outcome measures generate some of the most clinically valuable data in healthcare, yet PROM collection quality failures render up to 40% of submissions unusable for regulatory or AI purposes. Meeting PROM data trust requirements means enforcing provenance, temporal precision, and validation standards at the point of collection, not after the fact.
FDA expects post-market surveillance data to meet specific standards for completeness, timeliness, and traceability. Most device manufacturers collect the data but fail to prove it is trustworthy. Here is what the agency actually requires, what the ISO standards add, and why AI-enabled devices face a fundamentally different surveillance burden.
FDA has accepted real-world evidence in over 100 regulatory decisions since 2018, but most RWE submissions fail on data quality, not study design. The agency's requirements for RWD reliability, relevance, and completeness demand provenance-level trust that traditional data pipelines cannot deliver without systematic scoring.
Up to 80% of clinical data sits in unstructured EHR notes, yet most NLP pipelines never verify the trustworthiness of that text before processing it. Without provenance, recency, and quality scoring on the source notes themselves, clinical NLP outputs inherit every upstream data problem and amplify it at scale.
A single whole slide image can exceed 10 gigabytes and contain billions of pixels, yet most digital pathology AI pipelines have no formal method for verifying where that image came from, how it was scanned, or whether its metadata survived the transfer intact. Whole slide imaging provenance is the missing trust layer between the glass slide and the algorithm.
Multi-omic data integration combines genomics, proteomics, and metabolomics into a single analytical framework, but most integration pipelines ignore the trust problem underneath. When three distinct data types with different provenance chains, quality standards, and temporal profiles converge, the weakest layer determines the ceiling of every downstream AI model.
Biobank samples carry genetic data that can identify individuals, families, and entire ethnic populations. Without verified provenance, consent traceability, and chain of custody documentation, biobank data becomes a liability rather than an asset for genomic research. Sample provenance requirements are tightening across regulatory bodies, and the gap between what biobanks track and what research use demands is growing.
The FDA has issued specific guidance on what patient registries must deliver before rare disease data can support regulatory decisions. Fewer than 7,000 rare diseases have FDA-approved treatments, and registry data quality is a primary reason. This post maps the gap between what registries collect and what the FDA actually requires for accelerated approval, natural history studies, and post-market surveillance.
The ACS NSQIP registry tracks 30-day surgical outcomes across 700+ hospitals, but its benchmarking value depends entirely on the integrity of the data feeding it. Incomplete case capture, variable abstraction quality, and coding inconsistencies degrade the registry's ability to serve as ground truth for surgical AI models. Without a formal trust score applied before data enters quality benchmarking pipelines, surgical outcomes intelligence starts on unstable ground.
Emergency departments generate some of the most chaotic, fragmented, and time-pressured data in medicine. Before any AI model uses ED visit data for triage, prediction, or resource allocation, that data must meet specific trust thresholds across provenance, recency, completeness, and consent. Most ED data fails on at least three of these dimensions.
Between 49% and 96% of clinical decision support alerts are overridden by physicians, and the primary reason is not alert fatigue. It is data trust. CDS systems built on unscored, unstandardized health data generate alerts that clinicians cannot verify, cannot contextualize, and cannot act on safely.
Population health AI models flag care gaps using data that is often incomplete, outdated, or coded inconsistently across systems. Trust scoring each record before it enters a care gap identification engine reduces false positives, eliminates phantom gaps, and produces outreach lists that clinicians actually trust. Without a trust layer, care gap data quality degrades every downstream decision from risk stratification to quality measure reporting.
Pharmacy claims data tells you a prescription was filled. It cannot tell you whether the pill was swallowed, when it was taken, or what happened to the patient between refills. Wearable sensors and smart packaging now capture the signals that claims miss, but the data they generate carries its own trust problem. Closing the pharmacy claims data gap requires scoring both sources before any AI model acts on them.
Sepsis prediction algorithms consume six to dozens of clinical variables, but most validation studies never measure whether the underlying data was trustworthy enough to train on. A sepsis model scoring an AUROC of 0.85 on clean vital signs can collapse to near-random performance when fed late, miscoded, or unstandardized EHR data. This post maps the specific data requirements for sepsis AI and defines the trust score thresholds each input stream needs before any model should act on it.
