Fine-Tuned Compact AI Models Make Large Gains Reading Cancer Records, Delaware Research Finds

(MENAFN- EIN Presswire) Task-specific training raised two four-billion-parameter models' F1 scores by about 60 %, pointing to oncology tools health systems could run on their own.

WILMINGTON, DE, UNITED STATES, October 22, 2026 /EINPresswire / -- Fine-tuning compact, open-weight AI models on a focused oncology task substantially improved their ability to turn clinical notes into structured cancer data, according to preliminary findings from a yearlong Delaware research project. The work was supported by Delaware's American Rescue Plan Act (ARPA) funding and conducted through Tech Impact's Data Innovation Lab, with Acellus Health serving as a subcontractor to Tech Impact.

Much of what matters in cancer care, including diagnoses, treatments and changes in disease status, is recorded in physician notes and reports rather than structured fields. Moving that information into registries, research datasets and care coordination systems still depends largely on trained specialists reading records and entering details by hand. The project tested whether models small enough to deploy locally could reliably take on that extraction work once trained for it, using data elements defined by mCODE, the HL7 FHIR standard for oncology data.

What the Research Found:
On the same held-out test set and scoring method, fine-tuning raised Qwen3.5-4B's F1 score from 0.41 to 0.66, an improvement of about 60 percent, and nearly tripled MedGemma 1.5-4B's, from 0.22 to 0.62. F1 balances finding relevant information against avoiding incorrect extractions. Results were consistent across three independent training runs for each model.

Both models have four billion parameters, a fraction of the size of leading general-purpose AI systems. Some of the largest gains came on elements that require standardized clinical codes, the format registries and interoperability standards depend on. Together, the results point to task-specific training, rather than model scale alone, as a practical lever for making clinical text computable.

Where the Models Still Fall Short
Performance varied by data element, and training did not improve every category. Medication extraction declined for one model, and some categories had too few test examples to support firm conclusions. The models also differed in whether they tended to miss documented information or supply answers the record did not support. These differences argue for evaluating each data element separately and keeping expert review in the workflow.

Why it Matters:
A compact model trained for a specific clinical task can run within a health system's own computing environment, so patient records need not leave the institution and organizations are less reliant on large, centralized AI services. That could put this kind of tool within reach of rural cancer centers and other programs working with limited staff and resources.

Structured cancer data supports registries, clinical trial matching, quality reporting and care coordination. Reducing the manual effort required to produce it could return specialist time to work that requires clinical judgment and, over time, help care teams move patients toward treatment sooner.

Next Steps:
A forthcoming white paper, drawing on interviews and insights gathered during the project, will examine how healthcare organizations could use this technique and the research model to build practical oncology tools that reduce repetitive data preparation. The next step is testing those applications with oncology providers to measure whether they reduce preparation burden and delays in care.

These are preliminary findings. Full methods and results are being prepared for peer-reviewed publication in the coming months. The work has not yet demonstrated shorter treatment times or reduced workloads in clinical practice. Any clinical application would require validation and human oversight; the models are not intended to make medical decisions.

Elizabeth Levy
Tech Impact
+1 215-557-1559
Visit us on social media:
LinkedIn

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.

You just read:

Fine-tuned Compact AI Models Make Large Gains Reading Cancer Records, Delaware Research Finds News Provided By Tech Impact October 22, 2026, 11:30 GMT Share This Article Distribution channels: Healthcare & Pharmaceuticals Industry

EIN Presswire's priority is author transparency. We do our best to weed out false and misleading content. The content above is the sole responsibility of the author who makes it available. If you have any complaints, kindly contact the author above.

MENAFN22102026003118003196ID1111849749