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.
Claims data reaches AI models 30 to 90 days after the clinical event it describes. That reporting delay introduces survivorship bias, truncates longitudinal records, and trains predictive models on a version of reality that no longer exists. For health plans and AI developers building on claims, the lag is not a nuisance. It is a structural defect in the training data.
Roughly 98% of patient-generated health data never enters a clinical workflow. The gap is not technology. It is trust. Without a measurable threshold for provenance, recency, and validation, consumer health data remains invisible to the systems that need it most.
Synoptic pathology reporting is 174 times more accurate than AI extraction from narrative text, yet most health systems still train models on unstructured biopsy data. The gap between structured and unstructured pathology data is not a formatting preference. It is a trust problem that determines whether AI outputs are clinically safe or silently wrong.
DICOM metadata carries over 2,000 possible data elements per imaging study, yet most radiology AI pipelines never verify whether that metadata is accurate, complete, or internally consistent. When provenance breaks in medical imaging, every downstream model inherits the error. This post maps where DICOM data trust fails and what scoring it before training actually requires.
RxNorm is the backbone of standardized drug data in the U.S., but inconsistent concept mapping, deprecated codes, and source system fragmentation introduce errors that silently corrupt medication reconciliation. When AI models train on these records without measuring data integrity first, the result is not just bad predictions. It is patient harm at scale.
Between 6% and 19% of laboratory tests cannot be accurately mapped to LOINC codes, according to published research. That mapping failure rate means any AI model trained on lab data inherits systematic misclassification at the source. LOINC code trust is not a terminology problem; it is a data integrity problem that compounds across every downstream system that consumes lab results.
SNOMED CT contains over 350,000 clinical concepts, but mapping errors between systems silently corrupt the data that health AI models train on. When a 'heart failure' concept maps to the wrong hierarchy or a retired code persists in a production system, clinical terminology stops being a foundation and starts being a liability.
CPT codes drive billions of dollars in claims decisions annually, but the procedural code data feeding AI models carries systemic quality problems that most teams never measure. Upcoding, unbundling, modifier misuse, and stale code mappings corrupt training data before a model ever sees it. CPT code integrity is not a billing problem; it is a model accuracy problem.
ICD-10 codes were designed for billing, not clinical truth. When health AI systems train on claims data without verifying diagnostic coding accuracy, they inherit financial incentives as clinical signals, turning reimbursement artifacts into model predictions that affect real patients.
More than 90% of hospital interfaces still run on HL7 Version 2, a standard first published in 1987. Every one of those messages carries structural ambiguity that degrades downstream analytics, AI training, and clinical decision support. Scoring HL7 message quality is how trust gets attached to legacy EHR data before it causes harm.
FHIR R4 contains 157 resources and remains the normative baseline for US healthcare interoperability. FHIR R5, published in 2023, introduces 31 new resources and structural changes to subscriptions, evidence handling, and requirements tracking that directly affect how data trust architectures validate provenance, consent, and recency. The version gap is not just a technical migration question; it is a trust architecture question.
Federated learning keeps raw health data local, but model weights transmitted between nodes still leak patient information. Gradient inversion attacks, membership inference, and property inference can reconstruct individual records from weight updates alone. The privacy promise of federated AI in healthcare requires more than architecture; it requires scored, governed data at the source.
Differential privacy adds mathematical noise to health datasets so that no single patient's record can be reverse-engineered from query results. But noise does not fix provenance failures, consent gaps, or the re-identification risks that arise when multiple de-identified datasets are linked. Understanding what differential privacy actually protects requires separating the math from the marketing.
Parents of children with spinal muscular atrophy produce a distinct behavioral data trail in the months before gene therapy decisions. These search patterns reveal urgency, fear, insurance confusion, and treatment comparison needs that clinical systems never capture. SMA data from parent search behavior creates a real-time map of where families are in the decision window and what support they lack.
Ehlers-Danlos syndrome takes an average of 10 to 12 years to diagnose. During that window, patients generate thousands of behavioral signals across search, community, and clinical data that no one is reading systematically. EDS data, connective tissue disorder behavioral signals, and hypermobility behavioral intelligence represent one of the most undertapped sources of pre-diagnostic insight in rare disease.
POTS patients wait an average of 4 to 7 years for a correct diagnosis, cycling through cardiology, psychiatry, and primary care before anyone names the condition. Behavioral data from search patterns, community forums, and wearable signals reveals a consistent pre-diagnosis fingerprint that the clinical system currently ignores. Understanding postural orthostatic tachycardia behavioral signals can close the autonomic disorder diagnosis gap before patients lose years to misdiagnosis.
The median delay in cardiac amyloidosis diagnosis is 22 months, and in some subtypes it stretches past 34 months. Behavioral data from patient search patterns reveals a pre-diagnostic window filled with fragmented symptom queries, specialist-hopping signals, and misdiagnosis frustration. This gap is not just a clinical failure. It is a data infrastructure failure.
Nearly one in three women with PCOS meets criteria for metabolic syndrome, yet the average diagnostic delay is 2 to 4 years. Behavioral data reveals that patients search for individual symptoms like weight gain, fatigue, and hair loss years before anyone connects them to polycystic ovary syndrome. PCOS metabolic risk intelligence can close this gap if we build the data infrastructure to capture it.
Approximately 300 children are diagnosed with diffuse intrinsic pontine glioma (DIPG) in the United States each year, and fewer than 1% survive five years. The behavioral data generated by their parents during the search for answers follows a pattern unlike any other cancer type: compressed timelines, extreme urgency, and a rapid pivot from symptom queries to clinical trial desperation. Understanding these pediatric brain tumor behavioral signals reveals both the cruelty of the disease and the structural failures of the systems meant to help families find options.
Stage IV cancer patients with bone metastases generate distinct behavioral signals months before clinical milestones like hospice referral or treatment discontinuation. These signals cluster around pain management, life expectancy by cancer type, and end-of-life planning. Bone metastasis data reveals a patient population searching with urgency, specificity, and fear that current oncology content fails to address.
BRCA1 and BRCA2 mutation carriers search for fundamentally different things online, and the divergence maps to real clinical differences in cancer risk, treatment options, and psychological burden. BRCA2 behavioral data reveals a population that searches later, worries about different organs, and engages with clinical trial content at lower rates than BRCA1 carriers. These behavioral signals represent an untapped intelligence layer for oncology teams, pharma, and research organizations.
KRAS mutations appear in roughly 25% of all human cancers, yet most precision oncology platforms treat patient search behavior around KRAS as background noise. KRAS mutation data, when mapped against behavioral signals, reveals where patients are in their treatment arc and what they will do next. This post covers what precision oncology patients actually search, what those signals mean for pharma and clinical trial teams, and why the data layer underneath matters more than the search itself.
Amazon Clinic collected health data from millions of consumers before quietly shutting down in 2024. The retail health data trust gap it exposed remains unsolved: consumer health records generated outside traditional clinical settings lack provenance, consent governance, and quality scoring. Until health data has a trust layer, retail health platforms will keep creating records that cannot be safely used for AI, research, or clinical decision-making.
UnitedHealth Group's AI claim denial controversy exposed a systemic failure that most coverage missed: the underlying data the algorithm trained on was never independently scored for trust, provenance, or recency. When a payer deploys AI on unverified data to deny care at scale, the problem is not just the model. It is the data supply chain that no one audited.
