What is AI clinical trial matching in cancer?

AI clinical trial matching means using software to compare a patient's medical details with the eligibility rules of clinical trials. Every cancer trial has a list of inclusion and exclusion criteria, such as cancer type, stage, earlier treatments, biomarker results, blood counts and general fitness. Checking one patient against many trials by hand is slow, and research staff often have to read through long reports to do it.

AI tools try to speed this up. Older systems, such as IBM Watson for Clinical Trial Matching, used natural language processing to pull facts out of records and rules out of trial documents. Newer research tools, such as TrialGPT from the US National Institutes of Health (NIH), use large language models, the same kind of technology behind chatbots, to read a patient summary and explain how it fits each criterion.

The important point is what these tools produce. They produce a shortlist of trials that may fit, with reasons. They do not decide that a patient is eligible, and they do not enrol anyone. A trial team and the treating doctor still have to check every criterion, and the patient still has to agree to take part.

Key facts about the AI clinical trial matching evidence

The main peer-reviewed studies, in brief:

  • TrialGPT (Nature Communications, 2024): a large language model framework from NIH researchers, tested on 183 synthetic patient summaries
  • TrialGPT judged individual eligibility criteria with 87.3% accuracy, against 88.7% to 90.0% for human experts, on 1,015 patient-criterion pairs
  • In a small TrialGPT user study (two doctors, six oncology trials), screening time fell by 42.6%
  • Watson for Clinical Trial Matching, US community cancer centre (JCO Clinical Cancer Informatics, 2020): screening 90 breast cancer patients for three trials took 24 minutes with the system and 110 minutes by hand
  • Watson, US academic centre (JMIR Medical Informatics, 2021): 87.6% accuracy, 81.1% sensitivity and 89% specificity in 318 breast cancer patients
  • AI matching system, Australian lung cancer patients (JAMIA Open, 2020): 91.6% accuracy across 102 patients and ten trials
  • All of these were retrospective or simulated studies; none was a randomised trial of patient outcomes

Key takeaways: AI clinical trial matching

  • AI can read patient summaries and trial criteria quickly and suggest trials that may fit, with reasons for each criterion.
  • In published tests, accuracy was close to, but not better than, human experts (TrialGPT 87.3% against 88.7% to 90.0%).
  • Time savings were real in small studies, such as 42.6% less screening time with TrialGPT and 24 minutes instead of 110 with Watson.
  • Matching is not enrolment. Strong evidence that AI tools increase the number of patients who join trials is still limited.
  • Study authors stress that a clinician or trial team must check eligibility before any decision.

How does AI clinical trial matching work?

Most AI clinical trial matching tools work in three steps. First, they narrow thousands of trials down to a short list of candidates, usually by cancer type, location in the body and key words in the patient's history. Second, they check the patient against each criterion of each candidate trial, marking it as met, not met or not enough information. Third, they combine those judgements into a ranking, so the trials that look like the best fit appear at the top.

TrialGPT follows exactly this design. Its authors describe three modules: retrieval of candidate trials, criterion-level matching and trial-level ranking. It was built on GPT-4 and gives a short written explanation for each criterion, which lets a reviewer see why the tool made a judgement.

Older systems such as Watson for Clinical Trial Matching first turned free-text notes into structured attributes, such as hormone receptor status or tumour stage, and then applied each trial's rules to them. In both designs, the result is only as good as the information fed in. If a report is missing, unclear or out of date, the tool can only mark the criterion as unknown or get it wrong.

What did the NIH TrialGPT study in Nature Communications show?

TrialGPT was published in Nature Communications in November 2024 by researchers from the NIH's National Library of Medicine and National Cancer Institute, with collaborators. It was tested on three public research datasets that together contained 183 synthetic patient summaries (58, 75 and 50 patients) and over 75,000 trial annotations. These summaries were written for research; they were not real hospital records.

The retrieval step found more than 90% of the relevant trials while keeping less than 6% of the starting trial collection. At the level of single criteria, doctors manually reviewed 1,015 patient-criterion pairs. TrialGPT was right 87.3% of the time, while human experts scored between 88.7% and 90.0%. In other words, the AI came close to experts but did not beat them. For ranking and excluding trials, the authors report that TrialGPT outperformed the best competing computer models by 43.8%.

The authors were careful about scope. Patients were represented only by short free-text summaries, without lab tables, imaging or longitudinal notes. Trial location and recruitment status were left out of the datasets, so a real system would have to check separately whether a trial is open and near the patient.

Did TrialGPT save doctors time in screening cancer patients?

The TrialGPT paper included a small user study in oncology. Two medical doctors screened six semi-synthetic patient cases, based on anonymised referral requests, against six oncology trials. Four of those trials were run at the US National Cancer Institute. Each patient-trial pair was screened by one doctor with TrialGPT and the other without it.

With TrialGPT, average screening time per pair was 32.7 seconds for one doctor and 37.8 seconds for the other. Without it, the times were 57.5 and 65.4 seconds. Overall, screening time fell by about 42.6%, with savings for individual patients ranging from 35.9% to 51.9%. Accuracy was 97.2% with TrialGPT and 91.7% without, but the authors note this difference was not statistically significant.

This is encouraging, but it is a very small study: two doctors, six patients and six trials. It shows that AI explanations can help a trained reviewer work faster. It does not show what happens when the tool is used across a busy hospital, with real and messy records, or by patients themselves.

What did IBM Watson for Clinical Trial Matching studies find?

Watson for Clinical Trial Matching was studied in several cancer centres before the large language model era. At Highlands Oncology Group, a large US community cancer practice, researchers assessed records of 997 breast cancer patients in 2016 and compared the system with manual review for 239 patients and four trials. Agreement on eligibility ranged from 81% to 96%, and specificity from 76% to 99%. Sensitivity was 91% to 95% for three trials but only 46.7% for the fourth. Manually screening 90 patients for three trials took 110 minutes; with the system, it took 24 minutes. This was published in JCO Clinical Cancer Informatics in 2020.

A retrospective pilot at an academic medical centre, published in JMIR Medical Informatics in 2021, used data from 318 breast cancer patients and four breast cancer trials. The system was 87.6% accurate, with 81.1% sensitivity and 89% specificity. When the team looked at disagreements, some came from the software getting attributes wrong, but others came from human reviewers making mistakes, such as missing a HER2 result that was in the record.

Both studies involved breast cancer only. Three authors of the 2021 study were IBM Watson employees, and the study disclosed Novartis funding, which readers should keep in mind.

How accurate was AI clinical trial matching for lung cancer patients?

A technical evaluation published in JAMIA Open in 2020 tested an AI matching system on 102 lung cancer patients at an Australian cancer hospital, against ten phase I to III trials registered on ClinicalTrials.gov. The system compared 7,252 patient attributes with 11,467 trial eligibility criteria. The median time per patient query was 15.5 seconds.

Clinicians' own judgements were used as the reference, and they agreed well with each other. Against that reference, the system's per-patient accuracy was 91.6%. Its sensitivity, meaning how many truly eligible patients it found, was 83.3%. Its precision, meaning how many of its suggested matches were correct, was 76.5%. Its negative predictive value was 95.7% and specificity 93.8%.

Put simply, the system was better at ruling patients out than at picking the right patients in. About one in four of its suggested matches was not correct, and about one in six eligible patients was missed. The authors called it a promising tool for prescreening large groups of patients, and they wrote that clinician input and oversight are still required.

Key AI clinical trial matching numbers at a glance

Reported results from the main published studies:

  • TrialGPT retrieval: over 90% of relevant trials found using less than 6% of the trial collection
  • TrialGPT criterion accuracy: 87.3% (human experts 88.7% to 90.0%), 1,015 patient-criterion pairs
  • TrialGPT user study: 42.6% less screening time; accuracy 97.2% with the tool and 91.7% without, not statistically significant
  • Watson, breast cancer (2020): eligibility agreement 81% to 96%; sensitivity as low as 46.7% for one trial; 24 minutes instead of 110
  • Watson, breast cancer (2021): accuracy 87.6%, sensitivity 81.1%, specificity 89% in 318 patients
  • Lung cancer, Australia (2020): accuracy 91.6%, sensitivity 83.3%, precision 76.5% in 102 patients

Does AI clinical trial matching increase enrolment in cancer trials?

This is the question that matters most to patients, and the evidence is thinnest here. Most studies measured accuracy or screening time, not how many people actually joined a trial. The authors of the 2021 Watson pilot wrote that further research is needed to know whether faster screening leads to better enrolment.

One retrospective study, published in Therapeutic Innovation and Regulatory Science in 2020, applied an AI tool called Mendel.ai to three oncology trials at a US cancer centre after the fact. For the two trials that had enrolled patients, the AI found 24% to 50% more correctly eligible patients than standard methods and missed none that staff had found. Standard prescreening had taken an average of 19 days to identify eligible breast cancer patients and 263 days for lung cancer patients, while the AI did it in minutes. For the third trial, which had failed to enrol, neither method found suitable patients.

Finding more potentially eligible people is a step towards enrolment, not enrolment itself. Patients may decline, be too unwell, live far from the trial site, or fail screening tests. Matching is not enrolment, and no randomised trial we found has shown that AI matching changes cancer outcomes.

Why does eligibility still need a clinician's review?

Trial eligibility criteria are often long, technical and open to interpretation. A criterion such as "adequate organ function" or "no clinically significant heart disease" needs medical judgement and up-to-date test results. When the TrialGPT authors studied the tool's mistakes, the largest share came from incorrect reasoning (30.7%), followed by ambiguous eligibility definitions (26.9%) and lack of medical knowledge (15.4%).

Data quality is the other big problem. The Watson studies found errors caused by attributes the software derived wrongly, notes that did not contain enough information, and situations the system could not yet handle, such as patients with more than one primary cancer. A report from six months ago may no longer describe the patient today, especially after new treatment.

For these reasons, the TrialGPT authors stated that experts should always be in the loop of medical AI deployments, and the Australian lung cancer study said clinician oversight is still required. The final eligibility decision is made by the trial's investigators after their own screening tests, and the decision to join belongs to the patient, guided by the treating oncologist.

Limits of the AI clinical trial matching evidence

AI trial matching is promising, but the evidence has clear gaps.

  • TrialGPT's main tests used synthetic patient summaries, not real hospital records
  • User studies were very small, such as two doctors screening six patients for six trials
  • Most Watson studies covered breast cancer at single centres; results may not apply to other cancers
  • Accuracy was close to, not better than, human experts, and sensitivity varied widely between trials
  • Few studies measured enrolment, and none we found measured survival or other patient outcomes
  • The published tests used English-language records and trial listings, mostly from ClinicalTrials.gov; we found no published evaluation on Indian hospital records or CTRI listings
  • Some studies had authors employed by, or funding from, companies with a commercial interest

How are cancer clinical trials found in India? The CTRI registry

In India, the main public source of trial information is the Clinical Trials Registry – India (CTRI). It is hosted by the ICMR-National Institute for Research in Digital Health, part of the Indian Council of Medical Research. CTRI was launched on 20 July 2007. Registration was voluntary at first, but since 2009 the Drugs Controller General (India) has made registration of trials in CTRI mandatory.

CTRI is a primary register of the World Health Organization's International Clinical Trials Registry Platform (ICTRP). Registered trials are freely searchable on the CTRI website and through the WHO search portal. Many international cancer trials that recruit in India are also listed on ClinicalTrials.gov, the US registry used in most AI matching research.

A registry entry is a starting point, not a guarantee. A listing may show a trial that has stopped recruiting, has strict criteria, or recruits only at certain hospitals. Families usually need the treating oncologist, or the trial site's research team, to confirm whether a trial is open and whether the patient could qualify. CTRI listings are written for researchers, so a doctor's help in reading them is often useful.

Can AI clinical trial matching be used for patients in India?

The published AI clinical trial matching studies come from the United States and Australia. They were tested on English-language patient summaries and on trials from ClinicalTrials.gov or individual hospitals. We have not found a peer-reviewed study that tested these tools on Indian hospital records, on reports that mix languages or handwritten notes, or on the CTRI registry. So the published accuracy figures may not apply directly to patients in India.

Indian medical records also vary a great deal. Reports may come from several hospitals and laboratories, biomarker results may be missing, and staging may not be written in the format a tool expects. Each of these gaps can cause an AI tool to miss a trial or suggest one that does not fit.

That does not make AI useless. It can help a family organise reports and prepare a list of trials to discuss. But in India, as elsewhere, the shortlist should go to the oncologist, who can check eligibility, explain the risks and benefits, and contact the trial site if a trial looks suitable.

What happens next for AI clinical trial matching research?

Research is moving from tests on simulated cases to real clinical settings. The NIH reported that the TrialGPT team was selected for the NIH Director's Challenge Innovation Award to further assess the model's performance and fairness in real-world clinical settings. The TrialGPT authors also suggested testing open-source language models, longer records and inputs such as lab results and imaging, which their study did not include.

The most useful future studies will measure what matters to patients: whether AI-assisted screening helps more eligible people hear about and join suitable trials, whether it reduces delays, and whether it works fairly for people who are often left out of research. Studies in India, using Indian records and CTRI listings, are needed before anyone can say how well these tools work for Indian patients.

Until then, AI clinical trial matching is best seen as an assistant to the people who screen patients, not a replacement for them. It can save time and widen the search. It cannot decide eligibility.

Questions to ask your doctor about AI clinical trial matching

  • Are there clinical trials, in India or abroad, that could suit my type and stage of cancer?
  • If an AI tool or website suggested a trial, can you check whether I meet its eligibility criteria?
  • Which reports or tests, such as biomarker results, would a trial team need from me?
  • Is this trial still recruiting, and which hospital would I need to visit?
  • What are the possible benefits and risks of this trial compared with standard treatment?
  • If I do not qualify now, could I qualify later, and how will we keep looking?

Sources

1. Jin Q, Wang Z, Floudas CS, et al. Matching patients to clinical trials with large language models. Nature Communications, 2024. https://doi.org/10.1038/s41467-024-53081-z 2. National Institutes of Health (via ScienceDaily). NIH-developed AI algorithm matches potential volunteers to clinical trials. ScienceDaily, 2024. https://www.sciencedaily.com/releases/2024/11/241118125819.htm 3. Beck JT, Rammage M, Jackson GP, et al. Artificial intelligence tool for optimizing eligibility screening for clinical trials in a large community cancer center. JCO Clinical Cancer Informatics, 2020. https://doi.org/10.1200/CCI.19.00079 4. Haddad T, Helgeson JM, Pomerleau KE, et al. Accuracy of an artificial intelligence system for cancer clinical trial eligibility screening: retrospective pilot study. JMIR Medical Informatics, 2021. https://doi.org/10.2196/27767 5. Alexander M, Solomon B, Ball DL, et al. Evaluation of an artificial intelligence clinical trial matching system in Australian lung cancer patients. JAMIA Open, 2020. https://doi.org/10.1093/jamiaopen/ooaa002 6. Calaprice-Whitty D, Galil K, Salloum W, et al. Improving clinical trial participant prescreening with artificial intelligence (AI): a comparison of the results of AI-assisted vs standard methods in 3 oncology trials. Therapeutic Innovation and Regulatory Science, 2020. https://doi.org/10.1007/s43441-019-00030-4 7. ICMR-National Institute for Research in Digital Health. Clinical Trials Registry – India (CTRI). CTRI website, accessed 2026. https://ctri.nic.in

Frequently asked questions

What is AI clinical trial matching for cancer patients?

AI clinical trial matching uses software to compare a patient's medical details with the eligibility rules of clinical trials. It reads reports and trial criteria and suggests trials that may fit, often with a reason for each criterion.

The result is a shortlist, not a decision. A trial team and the treating oncologist still need to check every criterion, run their own screening tests and discuss the risks and benefits before anyone joins a trial.

How accurate is AI at matching patients to cancer trials?

In published studies, accuracy was close to, but not better than, human experts. NIH's TrialGPT judged single eligibility criteria correctly 87.3% of the time, while human experts scored 88.7% to 90.0%. A Watson study in 318 breast cancer patients reported 87.6% accuracy.

Accuracy also varied between trials. In one Watson study, the system found 91% to 95% of eligible patients for three trials but only 46.7% for a fourth. Most of these results came from retrospective or simulated studies.

What is TrialGPT and has it been tested on real patients?

TrialGPT is a research tool from NIH scientists that uses large language models to match patients to trials. It was published in Nature Communications in 2024 and tested mainly on 183 synthetic patient summaries written for research.

A small user study with two doctors, six semi-synthetic oncology cases and six trials found screening time fell by 42.6%. The NIH reported that the team received an award to assess its performance and fairness in real-world clinical settings.

Does AI trial matching help more patients join clinical trials?

The evidence is limited. Most studies measured accuracy or screening time, not how many people joined a trial. One retrospective study of three oncology trials found an AI tool identified 24% to 50% more correctly eligible patients for the two trials that had enrolled.

Finding more possible matches is not the same as enrolment. Patients may decline, be too unwell, live far from the site or fail screening tests. We found no randomised trial showing that AI matching changes cancer outcomes.

Can I trust an AI tool that says I am eligible for a trial?

Treat it as a starting point for a conversation with your oncologist. AI tools can misread reports, miss recent changes in your health or misunderstand vague criteria. In the TrialGPT study, the most common errors were incorrect reasoning and ambiguous eligibility definitions.

The researchers who built these tools say clinicians must stay involved. Only the trial's own team can confirm eligibility, after reviewing your records and doing their own screening tests.

Which reports are needed for clinical trial matching?

Trial criteria usually depend on cancer type, stage, earlier treatments, biomarker results, blood counts and general fitness. That means pathology reports, scan reports, biomarker test results and a treatment history are often needed.

AI tools are only as good as the information they receive. In the Watson studies, some errors came from notes that did not contain enough information. Missing or old reports can lead a tool to miss a trial or suggest one that does not fit.

How do I find cancer clinical trials in India?

The main public source is the Clinical Trials Registry – India (CTRI), hosted by the ICMR-National Institute for Research in Digital Health. It was launched in 2007, and since 2009 the Drugs Controller General (India) has made trial registration in CTRI mandatory. Trials are freely searchable on the CTRI website and the WHO search portal.

A listing does not guarantee a trial is still recruiting or that you qualify. Your oncologist or the trial site's research team can confirm this.

Do AI trial matching studies apply to patients in India?

Not directly. The published studies came from the United States and Australia and used English-language patient summaries and trials mostly from ClinicalTrials.gov. We found no peer-reviewed study testing these tools on Indian hospital records or on CTRI listings.

Indian records often come from several hospitals and laboratories, and some details may be missing. AI can still help organise reports and prepare questions, but eligibility should always be checked by an oncologist.

This page is general information, not medical advice. OncoVOIS explains and helps you prepare; your oncologist decides about your care.