Stuhlmiller TJ, et al. AI in Precision Oncology. 2026;3(6):171–182. Peer-reviewed.
Read the full paper →
The short version
Key cancer details often sit in notes, scanned faxes, and pathology reports, not in structured EHR fields. This study shows how xCures pulls those details out at scale, with 99.0% F1 on cancer diagnosis variables and every answer linked to its source.
What the study did
The study tested two methods that work together. The first pulls specific facts from each document and maps them to FHIR and OHDSI standards. The second answers patient-level clinical questions across the full record, citing the documents behind each answer. Clinically trained reviewers checked the outputs against the source records.
Key findings
- Cancer diagnosis: 98.2% accuracy, 98.7% PPV, and 99.4% sensitivity (99.0% F1) across 3,960 reviewed elements.
- Lines of therapy: 97.0% accuracy, 95.4% PPV, and 99.8% sensitivity (97.6% F1) across 842 reviewed elements.
- Medications, radiation, and surgery: about 95% or higher on accuracy, precision, recall, and F1.
- Real-world deployment across 3,493 patients: cancer diagnosis and date found for 93.5% of patients, compared with 69.8% in structured data.
Results based on a retrospective analysis of a defined historical dataset and do not guarantee future performance.
What this means for your team
- Clinical and trial teams: stage and grade were extracted for 64% and 62% of patients, where structured data had almost none. You get the details trial screening and chart prep depend on without digging through notes.
- Data and RWD teams: 207% more distinct oncology drug ingredients than structured EHR data, mapped to FHIR and OHDSI. Your data arrives analysis-ready.
- Reviewers: a median of 4 cited documents per diagnosis item and 8 per lines-of-therapy item, out of a median of 209.5 per patient. You check a few cited pages instead of the whole chart.
Where this shows up in xCures
This is the methodology behind xCures Evidence Grade Data and Decision-Ready Checklists.
Limitations
The analysis is retrospective. Results depend on document quality, and poor scans can introduce errors. Best response was documented for only 18.5% of therapy lines, so response rates describe that subset only. The authors name a blinded comparison against independent human abstractors as a priority for future work.
Stuhlmiller TJ, et al. AI in Precision Oncology. 2026;3(6):171–182. Peer-reviewed.
Read the full paper →
See clinical clarity in your workflow.
A 30-minute demo with our team. No prep required. Test it out on your own data.
What did the study measure?
How accurately xCures’ validated extractors and checklists pull oncology data from real-world medical records, as checked by clinically trained reviewers.
How accurate was cancer diagnosis extraction?
98.2% accuracy, with 98.7% PPV and 99.4% sensitivity, across 3,960 reviewed elements.
Where can I read the full study?
It’s open access in AI in Precision Oncology, through this link.