Fragmented populations and uncoded diagnoses make oncology and rare disease recruitment the hardest to do by hand

Oncology and rare disease trials have to find very few eligible patients among very many records, and the information that identifies those patients is often not stored in a searchable field. In the National Cancer Institute’s NCI-MATCH trial, 5,954 patients with refractory cancers were screened centrally, and 17.8% were ultimately assigned to a treatment arm (Flaherty et al.). In rare disease, an analysis of the Orphanet database identified 6,172 distinct conditions, most of which are individually very uncommon (Nguengang Wakap et al.). Under those conditions, recruitment depends on finding the few eligible patients at the moment they are eligible.

Centrova’s platform works across disease areas. This post focuses on these two because the recruitment problem is most acute there, and because the evidence shows why site-level screening practices matter more as populations get smaller. Post 1 in this series covers general oncology enrollment barriers, so we concentrate here on what is specific to these fields.

Biomarkers divide oncology into small groups

Precision oncology trials recruit patients whose tumor carries a particular alteration and who have reached a particular point in treatment. NCI-MATCH illustrates how quickly that narrows the field. Of the 5,954 biopsies analyzed, an actionable alteration was found in 37.6%, and after clinical and molecular exclusion criteria were applied, 17.8% were assigned to one of the 30 treatment arms. The investigators estimated that 26.4% could have been assigned if every subprotocol had been open at the same time (Flaherty et al.). Actionability also varied widely by tumor type, from more than 35% in urothelial cancers to under 6% in pancreatic and small-cell lung cancer.

Individual arms were smaller still. In the nivolumab arm for mismatch repair-deficient cancers outside the colon and rectum, 2% of 4,902 screened patients had the qualifying deficiency (Azad et al.). The trastuzumab and pertuzumab arm for HER2-amplified tumors enrolled 35 patients (Connolly et al.). A site that screens only for the cancer type will therefore find that most of its patients fail the molecular criterion, and a site that screens only for the molecular result will miss the tumor types and treatment history that complete the criteria.

Eligibility often depends on where the patient is in treatment

The second constraint is timing. NCI-MATCH assigned patients who had advanced cancer and had progressed on prior treatment (Connolly et al.), and half of the patients in the HER2 arm had received three or more prior therapies. Eligibility for such a study opens when progression is documented and closes when the next line of therapy begins. Post 1 described studies in which patients were excluded because another line of treatment had already started by the time eligibility was assessed. In practice, a chart reviewed a month late may describe a patient who is no longer eligible.

Trial closures show the cumulative effect. In a review of 238 phase III trials run by five U.S. oncology cooperative groups between 1993 and 2002, 66% closed with sufficient accrual, and 70 trials (29%) closed specifically because of poor accrual (Schroen et al.). That study predates much of today’s biomarker-based trial design, so the current figure may differ, but it establishes the baseline.

Rare diseases are numerous but individually small

Rare disease is a large category made of small parts. The Orphanet analysis found that 84.5% of the 5,304 diseases with point-prevalence data had a prevalence below 1 in 1,000,000, and it estimated that rare diseases as a group affect 3.5% to 5.9% of the population, or 263 to 446 million people worldwide (Nguengang Wakap et al.). Any single site will therefore see only a handful of patients with a given condition. Most of those patients are also hard to recognize. In a Rare Barometer survey of 6,507 people living with 1,675 rare diseases across 41 European countries, the average time to diagnosis was 4.7 years, and 56% were diagnosed more than six months after their first medical contact (Faye et al.). A patient may be eligible for a trial for years before the chart says what the disease is.

Coding adds a further problem. A Spanish study that mapped 5,775 Orphanet disease codes to the ICD-10 system used in that country, in which rare diseases “often lack a specific code,” found that only 1,055 of the equivalences were one-to-one and that 1,322 equivalences were to ICD-10 codes shared with 50 or more Orphanet diseases. The authors concluded that ICD-10-ES codes lack the specificity needed to identify rare diseases (Rico et al.). A separate validation of aplastic anemia cases identified from hospital discharge codes and mortality records found that, among flagged patients whose medical records could be reviewed, only 15% had the disease (Ruiz et al.). A diagnosis-code query is therefore a weak tool for rare disease, both because it can miss patients and because it can flag the wrong ones. These studies come from Spain, and U.S. coding differs in detail, so we treat them as evidence of a general limitation rather than a U.S. measurement.

Small trials stall when accrual is slow

The consequences are visible in trial outcomes. A cross-sectional analysis of 659 randomized rare disease trials registered between 2010 and 2012 found that 199 (30.2%) were discontinued, and lack of patient accrual was the most common reason, cited in 64 of them (Rees et al.). Rare cancers show the same pattern. The GOG 0241 trial in mucinous ovarian cancer stopped after 50 patients because of slow accrual, and central pathology review confirmed the diagnosis in only 18 of 40 cases with available material (Gore et al.). The SWOG S1701 trial in thymic carcinoma enrolled 21 patients between 2018 and 2022 against a target of 66, and was closed early because of feasibility (Tsao et al.).

Where the identifying information is recorded

The common thread is that the facts that define eligibility are specific, and the structured fields in a record are not. The Spanish coding study shows that diagnosis codes cannot name many rare diseases. In oncology, we expect the details that matter, such as a biomarker result, prior therapy and progression, to appear first in pathology, genomic and clinic reports, because those are the documents in which they are generated. This is our reasoning; we did not find a peer-reviewed study that measures what share of those details is absent from structured fields.

There is evidence that text-based screening can extract such details. A system applied to the clinical notes of 229 breast cancer patients extracted eligibility criteria with a recall of 90.9% and precision of 89.7% (Meystre et al.). In a study of neuro-oncology clinic letters, a language-processing model identified brain tumor diagnoses with a macro-recall of 0.964 and linked them to ongoing trials (Booker et al.). An automated approach in a pediatric emergency department reduced the screening workload by 92% in a gold-standard test set (Ni et al.). These are single-site evaluations with different designs, and none directly compares text-based screening with a structured query on the same patients, so they show feasibility rather than a measured gain.

What sites can do

The evidence points to a few practical steps. Sites can screen on the full set of criteria, combining diagnosis, biomarker and treatment line, rather than on cancer type alone. They can re-screen at predictable points, such as a new progression note or a new pathology result, because the eligibility window is short. For rare disease, they can search the notes and problem lists for disease names and synonyms as well as billing codes, and confirm the diagnosis from the record. They can also track which trials compete for the same patients, although we found no study quantifying that overlap. Each of these steps is more work than a code query, which is the reason to automate them.

What we are building

Centrova’s platform reviews the clinical notes and documents a site already holds and checks each patient against the inclusion and exclusion criteria of open trials. Coordinators and investigators review the candidates and make the final eligibility determination. The platform also includes patient tracking, communication and site coordination tools, works across disease areas and with multiple EHR systems, and is free to test during development. Sites can contact us.

Sources

1. Flaherty KT, Gray RJ, Chen AP, et al. Molecular landscape and actionable alterations in a genomically guided cancer clinical trial: National Cancer Institute Molecular Analysis for Therapy Choice (NCI-MATCH). J Clin Oncol. 2020;38(33):3883-3894. https://doi.org/10.1200/JCO.19.03010
2. Azad NS, Gray RJ, Overman MJ, et al. Nivolumab is effective in mismatch repair-deficient noncolorectal cancers: results from Arm Z1D-A subprotocol of the NCI-MATCH (EAY131) study. J Clin Oncol. 2020;38(3):214-222. https://doi.org/10.1200/JCO.19.00818
3. Connolly RM, Wang V, Hyman DM, et al. Trastuzumab and pertuzumab in patients with non-breast/gastroesophageal HER2-amplified tumors: results from the NCI-MATCH ECOG-ACRIN trial (EAY131) subprotocol J. Clin Cancer Res. 2024;30(7):1273-1280. https://doi.org/10.1158/1078-0432.CCR-23-0633
4. Schroen AT, Petroni GR, Wang H, et al. Achieving sufficient accrual to address the primary endpoint in phase III clinical trials from U.S. Cooperative Oncology Groups. Clin Cancer Res. 2012;18(1):256-262. https://doi.org/10.1158/1078-0432.CCR-11-1633
5. Nguengang Wakap S, Lambert DM, Olry A, et al. Estimating cumulative point prevalence of rare diseases: analysis of the Orphanet database. Eur J Hum Genet. 2020;28(2):165-173. https://doi.org/10.1038/s41431-019-0508-0
6. Faye F, Crocione C, Anido de Peña R, et al. Time to diagnosis and determinants of diagnostic delays of people living with a rare disease: results of a Rare Barometer retrospective patient survey. Eur J Hum Genet. 2024;32(9):1116-1126. https://doi.org/10.1038/s41431-024-01604-z
7. Rico J, Echevarría-González de Garibay LJ, García-López M, et al. The interoperability between the Spanish version of the International Classification of Diseases and ORPHAcodes: towards better identification of rare diseases. Orphanet J Rare Dis. 2021;16(1):121. https://doi.org/10.1186/s13023-021-01763-y
8. Ruiz E, Ramalle-Gómara E, Quiñones C, Rabasa P, Pisón C. Validation of diagnosis of aplastic anaemia in La Rioja (Spain) by International Classification of Diseases codes for case ascertainment for the Spanish National Rare Diseases Registry. Eur J Haematol. 2015;94(5):400-403. https://doi.org/10.1111/ejh.12432
9. Rees CA, Pica N, Monuteaux MC, Bourgeois FT. Noncompletion and nonpublication of trials studying rare diseases: a cross-sectional analysis. PLoS Med. 2019;16(11):e1002966. https://doi.org/10.1371/journal.pmed.1002966
10. Gore M, Hackshaw A, Brady WE, et al. An international, phase III randomized trial in patients with mucinous epithelial ovarian cancer (mEOC/GOG 0241) with long-term follow-up. Gynecol Oncol. 2019;153(3):541-548. https://doi.org/10.1016/j.ygyno.2019.03.256
11. Tsao AS, Hsieh MH, Koczywas M, et al. S1701, a randomized phase 2 trial of carboplatin-paclitaxel with and without ramucirumab in patients with locally advanced, recurrent, or metastatic thymic carcinoma. JTO Clin Res Rep. 2024;5(12):100738. https://doi.org/10.1016/j.jtocrr.2024.100738
12. Meystre SM, Heider PM, Kim Y, Aruch DB, Britten CD. Automatic trial eligibility surveillance based on unstructured clinical data. Int J Med Inform. 2019;129:13-19. https://doi.org/10.1016/j.ijmedinf.2019.05.018
13. Booker J, Penn J, Noor K, et al. Utilising natural language processing to identify brain tumor patients for clinical trials: development and initial evaluation. World Neurosurg. 2025;197:123907. https://doi.org/10.1016/j.wneu.2025.123907
14. Ni Y, Kennebeck S, Dexheimer JW, et al. Automated clinical trial eligibility prescreening: increasing the efficiency of patient identification for clinical trials in the emergency department. J Am Med Inform Assoc. 2015;22(1):166-178. https://doi.org/10.1136/amiajnl-2014-002887