Report
March 2026
AI-Powered Clinical Data Extraction
Quantitative Benchmark Results — Oncology Concept Extraction Pipeline
Health Universe's AI extraction pipeline achieves 93.5% precision and 92.9% recall (F1 = 93.2%) across 6 cancer types and 133 pathology documents, extracting over 6,200 structured data fields in under 17 minutes — replacing hours of manual chart abstraction with near-instant, auditable, structured output.
93.5%
Precision
92.9%
Recall
93.2%
F1 Score
6
Cancer Types
Why This Matters
Manual clinical data abstraction is the rate-limiting step in oncology workflows — from clinical trial matching to tumor board preparation to registry reporting. A single pathology report can require 15–30 minutes of expert review to extract structured data. Health Universe's Oncology Concept Extraction pipeline automates this process with AI, delivering structured, schema-validated clinical data in seconds.
The system was validated against expert-curated ground truth across 133 pathology documents from 40 patients spanning 6 major cancer types. Every extracted field — from TNM staging to biomarker results to specimen details — was compared against the reference standard.
Performance by Cancer Type
| Cancer Type | Documents | Precision | Recall | F1 Score |
|---|---|---|---|---|
| Breast | 32 | 95.1% | 95.0% | 95.1% |
| Prostate | 20 | 94.6% | 93.7% | 94.2% |
| Lung | 26 | 93.7% | 91.0% | 92.3% |
| Neuroendocrine | 21 | 93.2% | 93.6% | 93.4% |
| Colorectal | 24 | 89.3% | 92.1% | 90.7% |
| Ovarian | 10 | 94.8% | 91.2% | 93.0% |
| Overall | 133 | 93.5% | 92.9% | 93.2% |
F1 scores exceed 90% across all cancer types, with breast cancer extraction reaching 95.1% F1 and prostate reaching 94.2%. The pipeline demonstrates consistent performance regardless of disease complexity or document structure.
Speed Advantage
7.5s
Per document extraction time
133
Documents processed in 16.6 min
47+
Structured fields per document
120×
Faster than manual abstraction
Manual Chart Abstraction
01
15–30 minutes per document
02
5–15% inter-annotator error rate
03
Not scalable; constrained by workforce
04
Output varies by abstractor
Oncology Concept Extraction AI Pipeline
01
7.5 seconds per document
02
93.5% precision, 92.9% recall
03
Scales horizontally with parallel processing
04
Deterministic, schema-validated output
At scale, the pipeline processes an entire patient's chart history in under a minute — enabling real-time clinical trial matching, automated tumor board preparation, and continuous registry updates.
What the Pipeline Extracts
01
Patient demographics (name, DOB, sex, MRN)
02
Report metadata (accession, dates, pathologist)
03
Specimen details (site, procedure, laterality)
04
Histologic diagnosis and grade
05
TNM staging (pathologic and clinical)
06
Tumor size, margins, and invasion status
07
Biomarkers (ER, PR, HER2, Ki-67, MSI, MMR)
08
Molecular findings (NGS, FISH, PCR)
09
ICD-10 and ICD-O-3 diagnostic codes
10
Disease-specific fields per cancer type
11
Synoptic pathology data
12
Multi-cancer detection per document
Architecture and Differentiators
The Oncology Concept Extraction pipeline uses an agentic, multi-step extraction architecture powered by GPT-4o with chain-of-thought reasoning. Unlike simple prompt-and-extract approaches, the system:
Cancer-type-aware schemas
Tailored Pydantic models for each cancer type ensure disease-specific fields are captured (e.g., Gleason score for prostate, hormone receptor status for breast).
Multi-document aggregation
Patient-level clinical profiles are assembled across multiple reports, resolving conflicts and deduplicating findings.
Structured, auditable output
Every extraction produces schema-validated JSON that can be directly consumed by downstream clinical systems, trial matching engines, and registry databases.
Parallel processing
Documents are processed concurrently, enabling sub-minute turnaround for full patient charts.
Validation Methodology
133 pathology documents from 40 oncology patients across 6 cancer types (breast, lung, prostate, neuroendocrine, colorectal, ovarian) were extracted and compared against expert-curated ground truth. 6,209 total data fields were evaluated, with 4,227 scored fields used for precision/recall computation. Semantic equivalence matching accounts for clinically equivalent representations (e.g., “Grade 1” = “well differentiated”, “negative” = “not identified”).
Precision is defined as correct extractions / total extractions (how often the system is right when it provides a value). Recall is defined as correct extractions / total ground truth fields (how completely the system captures available information). F1 is the harmonic mean of precision and recall. All metrics are computed on the full 133-document validation corpus. ©2026 Health Universe, Inc. Confidential. All rights reserved.