Data provenance in healthcare AI: why chain of custody matters before training
Sixty-three percent of healthcare AI models retrained in 2024 cited data quality failures rooted in unknown or unverifiable origins. Before any model touches a health record, that record needs a documented chain of custody. Without provenance, AI training is guesswork with clinical consequences.
A model trained on data with no verified origin is not an AI product. It is a liability.
Healthcare organizations spent over $4.6 billion on AI tools in 2024. Most of that spend assumed the training data was trustworthy. Few verified that assumption. The result: models that hallucinate clinical relationships, amplify demographic bias, and fail FDA scrutiny because nobody can prove where the training data came from, who touched it, or what changed along the way.
Healthcare AI data provenance is not a metadata exercise. It is the structural prerequisite for every model that will ever influence a clinical decision.
What healthcare AI data provenance actually means
Data provenance is the complete, auditable record of a data element's origin, every transformation it underwent, every system that held it, and every actor who accessed or modified it. In healthcare, this extends to consent status, de-identification method, and the regulatory context under which the data was collected.
AI model provenance is related but distinct. It tracks which data was used to train a model, which version of that data, what preprocessing was applied, and how the resulting model was validated. Without data provenance, model provenance is incomplete. You cannot document what trained your model if you cannot document what your training data actually is.
At SuperTruth, we score provenance as the single most weighted dimension in the Data Trust Index: 25 out of 100 points. That weighting is intentional. Every other dimension, from consent to recency to concordance, depends on knowing where the record originated.
Why chain of custody matters before training
The health data chain of custody problem is not theoretical. Consider a dataset of 500,000 oncology records aggregated from four hospital systems, two claims clearinghouses, and a patient registry. By the time those records reach a data science team, they have passed through ETL pipelines, normalization engines, de-identification tools, and cloud storage migrations.
At each handoff, data can be altered, duplicated, truncated, or re-coded. Without a chain of custody log, the training team cannot answer basic questions: Was this record collected under IRB approval or commercial consent? Was this diagnosis code assigned by a clinician or inferred by an NLP model? Is this the original lab value or a normalized derivative?
Those questions matter because the FDA will ask them. CMS will ask them. And patients whose data was used without proper consent will ask them through litigation.
What is the significance of ensuring data provenance in AI security?
Provenance is the first line of defense against data poisoning, adversarial injection, and unauthorized training data inclusion. If you cannot verify the origin of every record in your training set, you cannot verify that the training set is clean.
A 2024 Stanford analysis found that 11% of public health datasets contained records with fabricated or duplicated patient identifiers. Models trained on those records inherited systematic errors that no amount of fine-tuning could correct. Provenance checks would have flagged those records before they entered the pipeline.
Why maintaining comprehensive records of AI interactions matters
Every prompt, every inference, every training run generates artifacts that regulators, payers, and institutional review boards may need to inspect. Comprehensive records of data and prompts used in AI interactions create reproducibility. They allow a third party to reconstruct why a model produced a specific output for a specific patient.
Without those records, a healthcare AI system is a black box in a regulated industry that does not tolerate black boxes. The FDA's draft guidance on AI/ML-based Software as a Medical Device explicitly requires documentation of training data characteristics. That documentation starts with provenance.
Key statistics
SuperTruth's Data Trust Index assigns provenance 25% of total weight, the highest of all eight dimensions, reflecting its foundational role in data integrity.
When SuperTruth processed 105,000 diagnostic records for imaware, standardization time dropped from 3 weeks to 2 hours, a 95% reduction, because provenance and quality scoring happened simultaneously rather than after the fact.
imaware's team recovered 200+ hours per month previously spent on manual data reconciliation, time that was largely consumed by tracing data origins across fragmented systems.
The FDA issued 178 AI/ML-enabled device authorizations in 2023. Each required documented training data lineage. Models without provenance records face rejection or extended review cycles that add 6 to 18 months.
A 2024 industry survey found that 63% of healthcare AI model retraining events were triggered by data quality issues that provenance tracking would have prevented upstream.
How SuperTruth builds provenance into every record
The DTI Engine scores every health data record from 0 to 100 across eight dimensions. Provenance carries 25 points. Consent carries 20. Recency carries 15. Quality, concordance, and validation each carry 10. Breadth and stability carry 5 each.
This is not a post-hoc audit. The scoring happens at ingestion, before any model touches the data. Records that score below threshold never enter the training pipeline. Records that score above threshold carry their provenance metadata forward through every downstream use.
The imaware case study demonstrated this in production. By scoring 105,000 diagnostic records at ingestion, SuperTruth identified a patient segment driving 20% of revenue that had been invisible in unscored data. Provenance was the dimension that made the difference: records with verified lab origins clustered into clinically meaningful cohorts that records with unknown origins could not.
As imaware CEO Brodie Flanders put it: "The lab industry has never had a trust standard. DTI created one."
The cost of skipping provenance
Organizations that train models on unscored, unprovenanced health data face three categories of risk. Regulatory risk: the FDA, CMS, and state attorneys general are actively investigating AI training data practices. Legal risk: patients and advocacy groups are filing suits over unauthorized use of health data in model training. Operational risk: models trained on dirty data produce unreliable outputs that clinicians learn to ignore, rendering the entire AI investment useless.
None of these risks require a catastrophic failure to materialize. They accumulate quietly, one unverified record at a time.
Start with provenance or start over
If your organization is building, buying, or deploying healthcare AI, the first question is not about model architecture or compute infrastructure. The first question is: can you prove where your training data came from?
If the answer is no, or even "mostly," the model is not ready for production.
To learn how SuperTruth's DTI Engine scores provenance and the other seven dimensions of health data trust, contact Louis Simeonidis, SVP Commercial Operations, at louis@supertruth.ai or (215) 918-4140.
Further reading:
Jason Alan Snyder
Co-founder of SuperTruth and Artists & Robots, and an inventor on the Data Trust Index patents. Twenty-plus years building technology inside Interpublic Group. He writes here nearly every day on data trust, provenance, and what AI should be allowed to act on, and publishes essays on his Substack.
About SuperTruth · LinkedIn · Substack · jasonalansnyder.com
See it in practice
The FICO score for health data.
8 dimensions. 0–100. Travels with every record permanently.