An estimated 80 percent of the clinical information that matters in oncology lives in narrative text: physician notes, pathology reports, treatment summaries. Today, certified registrars and abstractors read those notes by hand and re-enter the findings into cancer registries. The work is slow, expensive, and does not scale with cancer incidence. The pandemic made it worse, leaving longer, more fragmented patient histories just as trained abstractors became scarce nationally.

General-purpose AI foundation models, used out of the box, have not closed that gap reliably enough for clinical use. Acellus Health took a different path.

Rather than using a large general-purpose model as delivered, the team spent a year building expert-adjudicated training data from real clinical notes and fine-tuning a compact, open source, four-billion-parameter model specifically for oncology. That model is the engine behind the ACE Engine™. The training corpus was deliberately balanced with negative examples, so the model learns not only what to extract but when there is nothing to extract.

The fine-tuned model scored 0.78 F1 across 3,548 held-out extraction tasks covering fifteen essential oncology data elements, scored against ground truth adjudicated by clinical experts. That is roughly 2.5x the same foundation model without fine-tuning, and it matched a fine-tuned model more than twice its size.

Because the model is small, it runs inside a hospital's own firewall on modest hardware. Patient records never leave the institution, and a community cancer program does not need a national vendor contract to benefit. Retraining as clinical standards evolve is inexpensive, so the investment keeps compounding.

Hospitals and health systems do not have to wait for ever-larger general models to make their data usable. A focused investment in clinical fine-tuning can make cancer records computable today: for registries, clinical trial matching, quality measurement, and care coordination.

You know the drill: the answer to "what treatment did this patient get in 2023?" is in there somewhere: a discharge summary from an outside hospital, a scanned PDF, a med rec note. Finding it is your time. The ACE Engine™ reads the full note set and returns the fifteen mCODE elements registries, quality programs, and trial screeners actually query, coded and structured, traceable back to the source text.

These are the fields that carry registry abstraction and CoC/NCDR quality reporting, and they are where the model is strongest.

The training corpus was deliberately balanced with negative examples, so the model learned when a note contains nothing to extract. False abstention, silently returning nothing when information was present, ran at 0.1%. When it doesn't know, it shows you an empty field, not a hallucinated one.

The model is four billion parameters, small enough to run on a single on-prem GPU server. Notes are not sent to a third-party API, and no PHI sits in someone else's cloud. It is a first pass for humans, not a replacement: every extraction is reviewable, and the model makes no medical decisions.

When you're being treated for cancer, your story gets written down in doctor's notes: visits, scans, biopsies, treatments, how you responded. But those notes are written for humans to read, not computers. So when a new doctor needs the whole picture, or a hospital needs to report information to a cancer registry, a person has to read through everything page by page and re-type the important parts. It's slow, it's expensive, and things get missed.

We call it the ACE Engine™. It's an AI that was taught, using thousands of real cancer notes that clinical experts carefully marked up, to find the fifteen things that matter most in a cancer record: the diagnosis and its code, treatments and why they were given, surgeries, cancer spread, other health conditions, and more.

Teaching it what not to write down mattered just as much. When something isn't in your record, the system leaves it blank instead of guessing.

The system is small enough to run on a computer inside your hospital. Your notes are not sent to a big tech company's cloud, and the hospital doesn't need to share them with anyone else to use it.

Everything public health knows about cancer runs through registries: incidence, stage at diagnosis, treatment patterns, survival. Everything registries know arrives through manual abstraction: certified specialists reading narrative notes and re-keying findings. The pipeline is accurate but narrow, backlogged by months in normal years, and it is losing trained abstractors faster than it replaces them, just as patient histories grow longer and more fragmented.

The system extracts directly to mCODE, the minimal Common Oncology Data Elements standard built on FHIR, not to a vendor's private schema. The underlying model is open-weight. The output is 99.9% schema-conformant. A health department or research network can adopt this without handing its data pipeline to a single company, and the data it produces is immediately poolable with other mCODE implementations.

The fine-tune took 4.5 hours on one GPU and roughly sixteen dollars of compute; the 4B model matched a fine-tuned model more than twice its size (0.78 vs 0.78 F1). The result matters beyond one benchmark: extraction at this level no longer requires an academic medical center's budget. Community cancer programs, where most cancer care happens, can run it on one server inside their own walls.

Oncology is one of the most documentation-intensive service lines you run, and a meaningful share of that cost is clerical: trained staff reading narrative notes and re-keying findings into registries, quality programs, and prior-auth packets. That labor market is shrinking while your case volumes are not. The ACE Engine™ takes the first pass of that work, reading the note set and producing coded, structured output, so your certified staff review and confirm instead of hunt and type.

The engineering behind the budget math: a four-billion-parameter model, fine-tuned for oncology, matched a fine-tuned model more than twice its size (0.78 vs 0.78 F1 on 3,548 held-out tasks). Small means it runs on a single on-prem GPU server: a capital item measured in thousands, not millions, often hostable on hardware you already lease. Small also means retraining for your templates, specialties, or an updated standard is a ~$16 compute job, not a vendor change order.