About me
Bella Caamano
My process
I started by studying 25 randomly selected GBM trials, working out which of their eligibility requirements a computer can actually check — age, performance status, specific molecular markers — versus the ones that come down to a doctor's judgment. The checkable ones became a set of simple rules, about four or five per trial. With the help of Claude Code, I turned those rules into a working tool.
How the tool works
Paste or upload an actual clinical diagnosis. It pulls the text out, then looks for the following details: age, diagnosis, newly diagnosed or recurrent, performance status, IDH, MGMT, prior radiation and chemotherapy, prior treatment lines, measurable disease, three tumor markers, and prior bevacizumab. Anything it can't find clearly, it leaves blank instead of guessing — that's what produces the "not enough information" answers. These details are compared against the rule set built from the 25 trials.
The tool also searches ClinicalTrials.gov for studies recruiting right now. Results are sorted first by confidence, then by location. Confidence here means how many of a trial's requirements the patient actually cleared.
How I tested it
I pulled 25 real anonymous patient cases from public cancer research databases to build an answer key — every patient checked against every trial, 625 decisions in all. Before trusting it, I checked whether each rule actually had patients that could test it. Several didn't — I had no recurrent-disease cases at all. I went back and fixed the data first. A few things the databases don't track, I filled in for testing and marked clearly so they're never mistaken for real measurements.
My tool reproduces all 625 exactly, across both new and recurrent patients. That shows my code follows my rules exactly. Whether those rules are medically correct would take a doctor to review, and that's my next step.
What it can't do
For live trials it checks seven things: age, whether you're newly diagnosed or recurrent, IDH, MGMT, performance status, prior bevacizumab, and prior radiation and chemotherapy. Those are the criteria that screen the most patients in or out, so they're the ones worth checking without the benefit of having specific rules written down and coded. Anything beyond that, the tool can't see — which is why those results are leads, not answers.
If I build a version 2, I'd want the live search to read each trial's requirements the way I did for the original 25 — and then measure how well it does that, the same way I measured everything else.
Artifacts
The research behind the tool
The actual files I built and used, as PDFs, in the order I made them — the 25 trials, the confidence and coverage checks, the patient cohort, and the answer key. Click any one to open it.
The 25 GBM trials
The starting point: 25 randomly selected glioblastoma trials, with each one's checkable eligibility requirements turned into simple yes/no rules.
2 ConfidenceThe GBM eligibility confidence table
Which eligibility criteria a computer can actually check, and how strict each one is — separating how often a rule appears from how hard it screens patients in or out.
3 CoverageThe GBM gate coverage map
A check that every rule actually has patients that can test it — how many of the 25 cases each criterion passes, fails, or leaves as "not enough information."
4 Test dataThe 25 patients (newly diagnosed + recurrent)
25 real, anonymized GBM cases from public cancer-research databases — the people the tool is tested against. Facts the databases don't track are marked, never guessed.
5 Answer keyThe 625-decision gold standard
Every patient checked against every trial — 625 hand-built answers. My tool has to reproduce all 625 exactly.
These open from the site's server. If you're viewing this page as a plain file, the same PDFs are in the project's data/artifacts/ folder.
Sources & research
The databases the tool draws on and the reading behind the project.
Data & databases
- ClinicalTrials.gov — the live trial registry the tool searches for studies recruiting now.
- cBioPortal — TCGA GBM (Firehose Legacy) — molecular and clinical data for the newly-diagnosed patient cases.
- cBioPortal — TCGA GBM (PanCancer Atlas, 2018) — the harmonized TCGA glioblastoma cohort.
- NCI Genomic Data Commons — the underlying genomic and clinical data repository.
- PubMed — glioblastoma multiforme — the literature search behind the eligibility rules.
Background & reading
- A Shot for Life — Mass General Giving — the fundraiser that started this project.
- Forbes (2024): “We Are Failing Those Living With One Of The Rarest And Deadliest Cancers”
- CureToday: “Financial and Educational Disparities Persist in Glioblastoma Care”
- FierceBiotech: “2019’s top 15 clinical trial flops”
- Neuro-Oncology (Oxford Academic) — journal article (vol. 20, 2018).
- Neuro-Oncology Practice (Oxford Academic) — journal article (vol. 7, 2020).
- PubMed — research article (PMID 31175826)
- PubMed — research article (PMID 35969361)
- PubMed Central — open-access article (PMC6216910)
- PubMed Central — open-access article (PMC7341158)
- eScholarship (University of California) — open-access paper (PDF)
Glossary
Glossary of terms
Plain-language explanations of the clinical terms the tool reads and the coded rules that appear in the artifacts. These are simplified for a general reader, not formal medical definitions.
- Glioblastoma (GBM)
- The most common and most aggressive cancerous brain tumor in adults. It grows fast, which is why finding a matching trial quickly matters.
- Newly diagnosed vs. recurrent diagnosis
- Newly diagnosed means the tumor was just found and hasn't been treated yet. Recurrent (or progressive) means it has come back or started growing again after earlier treatment. Many trials accept only one or the other.
- Performance status
- A simple score for how well a patient can handle everyday activities — a stand-in for overall fitness. Most trials set a minimum. It's recorded on one of two scales, KPS or ECOG (both below).
- KPS — Karnofsky Performance Status performance:KPS
- A 0–100 scale where 100 is normal activity with no complaints and lower numbers mean more day-to-day help is needed. Trials often require KPS ≥ 60 or ≥ 70.
- ECOG performance status performance:ECOG
- A shorter 0–5 scale (Eastern Cooperative Oncology Group) where 0 is fully active and higher numbers mean more limitation. It captures the same idea as KPS — ECOG 0–2 roughly matches KPS ≥ 60–70.
- IDH
- A gene (isocitrate dehydrogenase) tested in the tumor, reported as IDH-mutant or IDH-wildtype. Under current (WHO 2021) rules, glioblastoma is IDH-wildtype; IDH-mutant grade 4 tumors are classified separately. The tool still checks it because a few trials in the set specifically target one status.
- MGMT
- A gene whose methylation status predicts how well the tumor responds to the chemotherapy temozolomide. Methylated usually means a better response; unmethylated usually means less benefit. A few trials are designed for one status only.
- Prior radiation and chemotherapy
- Whether the patient has already completed standard first treatment — radiation plus the chemotherapy temozolomide (the "Stupp protocol"). Some trials require it to be finished; others require that it hasn't been given yet.
- Prior treatment lines
- How many separate courses of treatment a patient has already been through. Some trials cap this — for example, no more than two prior lines.
- Measurable disease
- Whether there's a tumor large enough to measure on an MRI by standard rules (RANO). Trials that track tumor shrinkage often require it. This usually needs a radiologist's read, so the tool leaves it blank rather than guessing.
- Bevacizumab (Avastin)
- A drug that blocks the blood-vessel growth tumors rely on, sometimes used in recurrent GBM. Some trials exclude patients who have already had it; others require that the patient progressed while on it.
- Biomarkers biomarker_CD70 biomarker_B7H3 biomarker_GPC3
- Proteins that can appear on the surface of tumor cells. Certain trials — especially
CAR-T cell therapies — only accept patients whose tumor tests positive for a
specific one, because the treatment is engineered to target that exact protein.
- CD70 — a surface protein targeted by some CAR-T trials.
- B7-H3 (also called CD276) — a surface protein targeted by some CAR-T trials.
- GPC3 (glypican-3) — a surface protein targeted by some CAR-T trials.
- Tumor Treating Fields (TTFields) willingness
- TTFields (the Optune device) is a wearable cap that uses electric fields to slow tumor growth. A few trials require that the patient is willing to use it — a personal choice, not a lab value.
- Coded rule: prior_therapy:post 6000cGy+TMZ
- Shorthand for "has completed standard first-line treatment" — about 60 Gy (6000 centigray) of radiation together with temozolomide (TMZ). Recurrent-disease trials often require this.
- Coded rule: mgmt:unmethylated required
- Shorthand for a trial that only accepts patients whose MGMT is unmethylated (see MGMT above).
Get in touch
Questions, corrections, or ideas? Email [email protected].