AI is eyeing patient recruitment

The trend of AI-driven pharmaceutical R&D is gaining increasing momentum and expanding its reach across broader scopes.

Previously, the core focus of the market was on the molecular end: who has identified high-potential candidate compounds using AI, and which company has cut the time spent on lead compound discovery by half.

As the source of innovative drugs, molecular discovery is inevitably the most core battlefield. However, pharmaceutical manufacturing is not only about molecular screening, but also includes clinical verification. AI also has huge room for application in the clinical link, and a large number of players have poured into directions such as patient recruitment and clinical trial design optimization.

On October 20, AstraZeneca and Oracle announced a collaboration that once again directs AI to the clinical segment: how to find the right patients and advance clinical trials more efficiently?

01

Opening the Door to Clinical Patient Recruitment

AstraZeneca is no stranger to AI. Its partnership with Oracle is not a supplementary lesson for AI pharmaceutical R&D.

According to public information, its goal is to "leverage modern applications and AI to improve the efficiency and success rate of clinical trials", with indications covering oncology, rare diseases, cardiovascular, renal and metabolic diseases, respiratory and immunology.

In other words, it wants to push AI through the hardest door in clinical practice: patient recruitment. The logic is not hard to understand: clinical recruitment is counted as one of the most complex processes in new drug R&D.

The first issue is efficiency. In traditional clinical practice, a coordinator may need to review 50 to 100 medical records to screen out one qualified participant. This essentially translates to a cycle problem.

The second is the success rate challenge. It is not rare for patients who have signed the informed consent form to eventually be blocked by the inclusion and exclusion criteria. In an audit covering 7481 screened patients at Tata Memorial Centre in India, the screening failure rate was as high as 51.0%.

This means that the stagnation of clinical trials is not only caused by the failure to find enough people, but more importantly, the failure to find the right people. There is a statistic in the FDA's 2025 report related to real-world evidence: among approximately 4,000 clinical trials in the United States, more than 70% of potential patients are locked out by the inclusion and exclusion criteria.

For this reason, Anne-Claire Gerbaldi, Chief Digital, Data and AI Officer of AstraZeneca, stated that too many patients are diagnosed too late, or have to travel a long way to get the care they need.

02

Growing Evidence of Efficiency Improvement and Cost Reduction

Certainly, the attempt of AI to change the efficiency and success rate of clinical recruitment is nothing new. On a global scale, there are not only many players entering this track, but also a large amount of evidence proving its feasibility.

For example, TrialGPT from the NIH team, a large language model-based system for patient-trial matching, was evaluated on 183 synthetic patients and more than 75,000 trial annotations, achieving a matching accuracy of 87.3%. In the paper published by Jin et al. in *Nature Communications* in 2024, the accuracy of human experts ranges from 88.7% to 90%, which is almost at the same level.

More critically, it delivers much higher speed. In the same user study, doctors using TrialGPT spent 42.6% less time on screening than manual review, without any loss of accuracy.

The same logic has been applied to actual products. myTrialsConnect uses generative AI to automatically review medical records, check against inclusion and exclusion criteria, and build a "virtual waiting room" for patients; NLP tools can pre-screen hundreds of candidates in minutes, compressing a task that originally takes a week to just a few hours.

The figures are more specific in hospital settings. A 2026 study published in *Nature Reviews Clinical Oncology* shows that tools like MatchMiner have reduced the screening time for a single patient from 110 minutes to 24 minutes, with some tools improving efficiency by more than 90%.

In MediSearch's 2026 benchmark, the enrollment speed brought by digital recruitment channels is 2 to 3 times that of traditional methods.

Celvion, a real-world data (RWD) platform based in Boston, focuses on using NLP to process unstructured texts that traditional RWD platforms cannot handle, such as free texts in pathology reports, handwritten follow-up notes from doctors, and even self-reported symptoms on patient forums.

According to its disclosed benchmark test, in the same solid tumor patient stratification task, the number of subgroups identified by the model is more than 3 times that of traditional statistical methods, and the separation of these subgroups on the survival curve is statistically significant.

Changes are taking place naturally, and more pharmaceutical companies will inevitably join this battlefield.

03

Achieving Speed, Accuracy and High Quality at the Same Time

Returning to the real-world logic, in the patient recruitment segment, efficiency is not the only factor to consider.

The notification-based intervention of MatchMiner failed to significantly improve the overall enrollment rate in actual deployment; after the Medical College of Wisconsin launched OncoLLM-CW, the working hours of coordinators were indeed shortened, but a large number of candidates matched by AI were excluded after manual review.

The underlying logic is that factors such as travel barriers, quota limits, patient willingness, and doctors' judgment cannot be replaced by algorithms for the time being.

This will inevitably push up direct clinical costs, because the cost of screening failure is very high: the patient's examinations, lab tests and imaging scans have already been done, the time of coordinators and researchers has already been spent, but in the end no enrollable data is obtained. The commonly used industry estimate is that the cost of each screening failure is about 1200 US dollars.

The complete calculation from Tufts CSDD shows that recruiting one subject costs an average of 6533 US dollars, and replacing one subject who drops out midway costs 19533 US dollars.

Of course, compared with screening failure, what is more important is whether the clinical success rate can be improved?

The identification of clinically enrolled patients is very important, and some AI enterprises hope to bring changes. However, some practical results in scenarios such as advanced non-small cell lung cancer and relapsed/refractory multiple myeloma show that although AI tools have improved patient enrollment efficiency and made patients more compliant with enrollment criteria, the hazard ratio of survival outcomes has not changed significantly.

Technology ultimately serves the final results. How to use AI to improve the efficiency and success rate of the clinical segment is clearly a problem that all participants need to continuously consider.

This article is from the WeChat Official Account "Anji Observation" (ID: anjiguancha), written by ZHENG Xiao, and authorized for release by 36Kr.