Threshold

Threshold reads the eligibility criteria of every recruiting cancer trial in Massachusetts, flags the rules that keep people out, and shows protocol teams what each rule costs, where, and how other trials already word it more inclusively.

Live: https://patient-eligibility-linter.vercel.app Built for: Hackachusetts 2026, Healthcare & biotech track (clinical research access)

Trial protocols are written once and copied forward. Exclusions such as "known HIV infection" or "must speak English" are often boilerplate, not science, and they shut out patients long before anyone looks at their cancer. Threshold is a linter for those rules. It suggests changes and cites the guidance behind them; the protocol team decides.

What we found

Across all 1,011 recruiting cancer trials with a recruiting Massachusetts site (ClinicalTrials.gov, pulled October 24, 2026):

• 1 in 5 phase 1–3 drug trials (22%, 166 of 748) still exclude people with HIV outright, even though FDA guidance since 2020 says people with well-controlled HIV should generally be eligible. The fix is already in use: about half of the trials that mention HIV (177 of 347) set conditions instead, such as an undetectable viral load on treatment.
• About 3 in 10 supportive-care, behavioral and quality-of-life studies (82 of 263) require a specific language, usually English. In Suffolk County's environmental-justice neighborhoods, 1 in 7 households (14.4%) has limited English.
• Access is concentrated: 945 of these trials recruit in Suffolk County, 24 in Hampden, 1 in Berkshire.

Sources: ClinicalTrials.gov API v2, CDC PLACES 2023, MassGIS Environmental Justice 2020. Guidance citations are listed in docs/sources.md.

What's in the app

Screen	What it does
Landscape	How often each of 13 restriction types appears across the 1,011 trials, with a filterable table of the exact quoted criteria. Every row links to its trial page.
Map	A county map of Massachusetts by restriction type, with CDC PLACES health measures and MassGIS environmental-justice measures for each county. Who gets turned away walks a sample patient through each restriction and shows how many local trials remain at each step.
Linter	Paste draft inclusion and exclusion criteria. Claude reads them with the same prompt and checks that extracted all 1,011 trials (or switch to in-browser keyword rules). Flags are grouped into Fix first, Worth a look and Team's call. Each one shows the quoted sentence, a suggestion with its cited guidance and, for blanket rules, standard inclusive wording you can copy, with a count of Massachusetts trials that already use conditional wording for that rule.
Find Trials	A patient-side prototype. Answer a few coarse questions (cancer type, age, county, HIV, daily activity, language, and more) and see trials sorted into Worth reaching out, Ask about one thing and Likely not yet, each with the sentence that put it there. Not medical advice.
Trial pages	One page per trial (1,011 in total) with the full criteria and every flagged sentence highlighted.

Can you trust it?

We measured our own AI instead of assuming it works.

• A human-labeled answer key. 42 trials: a seeded random sample of 30 (quotas by phase, sponsor type and cancer type), 2 added for BMI coverage, and 10 more drawn after prompt tuning and labeled blind. A person wrote every label; no model did.
• Held-out accuracy (the 10 trials the prompt never saw): the extractor caught 46 of the 50 restrictions the human found (recall 92%), and 46 of its 47 flags were right (precision 98%). F1 0.95, 95% CI 0.89–0.99. With only 10 trials, the interval is wide.
• Weak spots are shown, not hidden. Prior malignancy is the weakest category (recall 72% across all 42 trials). Diabetes, BMI and smoking have only two labeled examples each, so we don't claim a score for them. Full table: eval/results.md.
• Every flag shows its receipt. All 3,839 flags in the dataset are checked to appear in their trial's criteria (ignoring whitespace, letter case and ClinicalTrials.gov's formatting escapes), and each trial page highlights them.
• Citations say how strong they are. Suggestions cite FDA and ASCO–Friends of Cancer Research guidance and mark it final, draft (ICH E21 for pregnancy) or general only (diabetes, BMI, smoking).
• The live Linter matches the measured pipeline. It uses the same model (claude-haiku-5-5) and prompt version as the batch extraction; tests/test_prompt_parity.py fails if the copies drift.

Reproduce the scores with python eval/run_eval.py.

What the numbers don't say

• The tool suggests; people decide. Some exclusions are justified, for example pregnancy with a drug that can harm a fetus.
• County figures are estimates for places, not people. CDC PLACES gives county-level crude prevalence; MassGIS figures cover environmental-justice block groups only. We report each measure on its own and never add or multiply them across criteria.
• BMI is an upper bound. PLACES obesity means BMI 30 or more, so it's exact only for caps at 30.
• A trial's site county stands in for where its patients come from. Patients travel.
• Find Trials is a first matcher. It doesn't check a trial's sex requirement, cancer stage or treatment history.

Privacy

• Find Trials uses your answers only to run the search. They aren't saved, shared or tied to you.
• Linter sends pasted text to Anthropic's API so Claude can read it. There is no database: the server keeps its last 200 results in memory so a repeat run of the same text is instant, and that memory clears when the server restarts. To keep a draft in your browser, choose Keyword rules.

This is the beginning of the README. Read the rest at the original. Read the full README ↗