Pathology report data quality: structured vs unstructured biopsy data trust
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.
A pathologist types "moderately differentiated adenocarcinoma, Gleason 3+4=7, margins positive at 0.2mm" into a free-text box. Another pathologist at a different institution enters the same findings into a synoptic template with discrete, coded fields. The clinical information is identical. The data quality is not even close.
This distinction between structured and unstructured pathology data is the fault line underneath every AI system that touches biopsy results. And most organizations building cancer AI have no idea which side of that line their training data falls on.
What is the difference between structured and unstructured data in pathology?
Structured data lives in predefined fields with controlled vocabularies. A synoptic pathology report uses templates defined by the College of American Pathologists (CAP), where each data element (tumor type, grade, margin status, lymphovascular invasion) occupies its own coded field. The data is machine-readable from the moment it is created.
Unstructured data is narrative text. It is the prose paragraph a pathologist dictates or types into a free-text field. It may contain the same clinical facts, but those facts are buried in natural language with variable phrasing, abbreviations, negations, and institutional jargon.
The difference matters because structured data can be queried, validated, and scored without interpretation. Unstructured data requires extraction, which introduces error at every step.
In healthcare more broadly, structured data includes lab values in LOINC-coded fields, ICD-10 diagnosis codes, and medication lists mapped to RxNorm. Unstructured data includes clinical notes, radiology impressions, and pathology narratives. The ratio is roughly 80% unstructured to 20% structured across most EHR systems. Pathology sits at the extreme end of this imbalance because many institutions still generate narrative reports as the primary record.
Why narrative pathology reports break AI pipelines
A 2024 study published in the Journal of Pathology Informatics found that synoptic reporting was 174 times more accurate than generative AI extraction from narrative reports. That number is not a marginal improvement. It represents the difference between a data pipeline that works and one that silently corrupts downstream models.
Here is what goes wrong when biopsy data lives in unstructured text:
Inconsistent terminology. One pathologist writes "negative margins." Another writes "margins free of tumor." A third writes "no carcinoma at inked margins." These mean the same thing clinically. To an NLP extraction pipeline, they are three different strings that must be reconciled.
Negation ambiguity. "No evidence of lymphovascular invasion" and "lymphovascular invasion not identified" are semantically equivalent but syntactically different. Rule-based NLP systems and even large language models frequently misclassify negated findings, flipping a negative result to positive.
Missing data vs. unstated data. In a synoptic report, an empty field is explicitly missing. In a narrative report, the absence of a finding might mean the pathologist did not observe it, did not look for it, or simply did not mention it. There is no way to distinguish these cases programmatically.
Abbreviation drift. "mod diff" might mean moderately differentiated in one institution and something else entirely in another. "LVI" is lymphovascular invasion at most academic centers, but abbreviation dictionaries are not standardized across health systems.
These are not theoretical problems. They are the reason that AI models trained on pathology data produce different outputs depending on which institution supplied the training set.
What happens when pathology data is entered inconsistently
The consequences of inconsistent pathology data entry cascade through every system that touches the results.
Cancer registries receive corrupted data. The National Cancer Institute's SEER program depends on accurate staging and histology data. When registrars must manually abstract unstructured pathology reports, error rates for individual data elements can exceed 20%. Structured synoptic reports reduce abstraction errors to below 1%.
Clinical decision support fails. If an AI system reads a biopsy result to recommend adjuvant therapy, and the extraction pipeline misidentifies margin status or tumor grade, the recommendation is wrong. The clinician may never know the upstream data was corrupted because the AI output looks confident.
Research cohort definitions shift. A retrospective study that identifies "all patients with HER2-positive breast cancer" will return different cohorts depending on whether the query hits structured fields or runs NLP extraction against narrative reports. One approach might miss 15-30% of qualifying patients due to extraction failures.
Billing and coding misalign. Pathology reports drive CPT and ICD-10 coding for biopsy procedures and diagnoses. When the source data is unstructured, coders must interpret narrative text, introducing variability that affects reimbursement accuracy and downstream claims data quality. This connects directly to the broader ICD-10 coding accuracy problem we have written about previously.
Key statistics
These numbers define the current state of pathology report data quality:
Is structured data easier to analyze and store than unstructured data?
Yes. This is true, and it is true by a wide margin.
Structured data fits into relational databases, can be queried with standard SQL, and requires no preprocessing before analysis. A structured pathology report with discrete fields for tumor size, grade, and margin status can be aggregated across 10,000 patients in seconds.
Unstructured narrative reports require NLP preprocessing, entity extraction, negation detection, and normalization before the same query can run. Each of those steps introduces error. The computational cost is higher. The latency is longer. And the results are less reliable.
Storage is also simpler. Structured data compresses efficiently and indexes cleanly. Unstructured text requires full-text indexing, which consumes more storage and makes retrieval slower at scale.
But the more important question is not about storage or speed. It is about trust. Can you verify that a structured field contains what it claims to contain? Yes, because the field has a defined vocabulary and validation rules. Can you verify that an NLP-extracted value from narrative text is correct? Only by going back to the source text and checking manually, which defeats the purpose of automation.
The false promise of AI extraction from narrative reports
The current SERP landscape shows a heavy emphasis on using generative AI to extract structured information from unstructured pathology reports. Tools like HistoGPT and various LLM-based pipelines promise to bridge the gap between narrative text and structured data.
This approach has value. It is better than leaving narrative data completely unprocessed. But it is not a substitute for generating structured data at the source.
Here is why. Every extraction step introduces a confidence interval. If an LLM extracts tumor grade from narrative text with 95% accuracy, and margin status with 93% accuracy, and lymphovascular invasion with 91% accuracy, the joint probability that all three fields are correct for a single patient is roughly 80%. Scale that across a dataset of 50,000 biopsies and 15 extracted fields per report, and the fraction of records with zero extraction errors becomes vanishingly small.
The 174x accuracy advantage of synoptic reporting over AI extraction is not surprising when you consider this compounding effect. Structured data at the source eliminates the extraction step entirely. There is no confidence interval because there is no inference.
This does not mean AI extraction is useless. For legacy data, for institutions that have not yet adopted synoptic reporting, and for historical research cohorts, NLP extraction is the only option. But the extracted data must be scored and validated before it enters any AI training pipeline or clinical decision system.
How pathology data quality affects AI model training
AI models absorb the biases and errors in their training data. When a cancer detection model trains on pathology data that includes both synoptic and narrative-extracted records without distinguishing between them, the model learns from two populations with different error profiles.
The synoptic records are clean, consistent, and coded. The narrative-extracted records contain NLP artifacts, misclassified negations, and missing values that were imputed rather than observed. The model treats both as equally reliable because no trust score differentiates them.
This is the core problem that data trust scoring solves. A DTI score applied to each pathology record before training tells the model builder exactly how much to trust each data point. A synoptic report with complete fields, verified provenance, and current consent might score 88 out of 100. A narrative report that was NLP-extracted with partial field coverage and no validation might score 41.
Both records contain pathology data. Only one should be training your model without additional review.
The provenance problem in biopsy data
Pathology data has a chain of custody problem that most data platforms ignore.
A tissue sample moves from the patient to the gross examination room, to the histology lab for processing and staining, to the pathologist's microscope, and finally to the report. At each step, identifiers must match. The specimen label must correspond to the patient record. The block and slide identifiers must trace back to the correct specimen. The final report must reference the correct accession number.
When any link in this chain breaks, the downstream data is corrupted regardless of whether the report is structured or unstructured. Provenance, which accounts for 25% of the DTI score, captures exactly this dimension. It asks: can you trace this data point back to its origin, through every transformation, to the source?
For pathology data, provenance means knowing which lab processed the specimen, which pathologist signed the report, which LIS (laboratory information system) generated the record, and whether the data was transcribed, extracted, or natively structured. Without this metadata, biopsy data trust is impossible to establish.
This connects to the broader chain of custody problem that affects all health data, but pathology represents one of the most acute examples because the physical specimen and the digital record must maintain correspondence across multiple systems.
What structured pathology data makes possible for AI
When pathology data is genuinely structured, coded, and scored, entire categories of clinical AI become viable.
Automated cancer staging. With discrete fields for tumor size, node count, and metastasis status, staging algorithms can run without human intervention and produce results that match manual staging in over 98% of cases.
Real-world evidence generation. Pharma companies building RWE submissions for FDA need pathology data they can audit. Structured, provenance-tracked biopsy data with a DTI score above a defined threshold meets this requirement. Narrative-extracted data with unknown confidence intervals does not.
Multi-institutional research. Federated learning across hospital systems requires data that is semantically interoperable. Two institutions using CAP synoptic templates with SNOMED CT coding can merge datasets without reconciliation. Two institutions with narrative reports in different formats cannot. The SNOMED CT mapping trust problem is directly relevant here.
Clinical trial matching. Eligibility criteria often specify pathology findings (e.g., "HER2-negative, hormone receptor-positive, Ki-67 > 20%"). Structured data enables automated matching. Unstructured data requires manual chart review, which is the primary bottleneck in clinical trial recruitment.
What needs to change
The path forward requires action at three levels.
At the source: Every pathology department that has not adopted CAP synoptic reporting should do so. The evidence is overwhelming. Synoptic reporting improves completeness, reduces abstraction errors, and generates machine-readable data without any post-processing.
At the data layer: Legacy narrative reports must be scored before they enter any AI pipeline. A trust score that accounts for provenance, recency, quality, and validation tells the model builder which records are safe to use and which require additional review or exclusion.
At the model layer: AI developers must stop treating all pathology data as equivalent. A training set that mixes synoptic and narrative-extracted records without weighting by trust score will produce models with hidden failure modes. DTI floor enforcement, where only records above a minimum trust score enter the training pipeline, is the minimum standard.
The DTI Engine scores every health data record 0 to 100 across 8 trust dimensions before your AI model sees it. If your team is evaluating pathology data for model training, regulatory submission, or clinical decision support, contact Louis Simeonidis 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.