EHR-Based Recruitment Tools Fell Short of a Broad Enrollment Gain
Clinical trial teams have spent years looking for a cleaner way to find eligible patients inside the electronic health record. In this multicenter German evaluation, two clinical trial recruitment support systems—recruIT and KAS+—did not deliver a statistically significant overall lift in weekly enrollment once they were switched on across university hospitals.
The finding is blunt. A digital pre-screening platform can organize data, but it does not automatically translate into more participants at the bedside.
Across the German MIRACUM consortium, investigators used a segmented interrupted time-series design at six university hospitals to examine whether IT-supported recruitment changed actual trial enrollment. Twenty-nine trials contributed 814 observation weeks after exclusions for premature termination, too few valid observation weeks, or no recruitment at all during the study window. The trial mix was broad, spanning neurology, internal medicine, urology, dermatology, otorhinolaryngology, intensive care, neurosurgery, and pediatrics. Of those 29 trials, 14 involved patient interventions, 15 were observational, 14 were randomized, and 11 were blinded.
That heterogeneity mattered because recruitment support tools operate inside real clinical workflows rather than in a controlled lab setting. A system that searches structured data efficiently still depends on how well eligibility criteria can be translated into computable logic, how complete the underlying records are, and how often trial staff act on the generated candidate lists.
Why the Systems Struggled to Move the Needle
The two tools were designed to automate pre-screening from routine care data. RecruIT used the OMOP Common Data Model and HL7 FHIR to support interoperable, automated patient pre-screening, while KAS+ was integrated directly into the EHR and generated pseudonymized screening lists with customizable alerts. Both depended on structured data already captured during care, which makes them attractive in principle and fragile in practice.
Eligibility logic was dominated by simple variables. Condition and age appeared in 82.7% of trials, and patient visit information appeared in 48.3%. Laboratory values and structured clinical observations were rarely available, appearing in only two trials each. In operational terms, that means the recruitment engines were often built around data elements that are relatively easy to extract, but not always rich enough to capture the true nuance of trial eligibility.
The average weekly recruitment number rose from 0.30 during the control phase to 0.54 during the intervention phase. The highest trial-level mean weekly recruitment also increased, from 3.15 to 4.47. Yet the unadjusted comparison did not hold up statistically. The Wilcoxon signed-rank test produced a W statistic of 153, a z value of −1.14, and a p value of 0.26, with a small effect size of 0.22.
The Signal Was Real in Some Trials and Absent in Others
Recruitment effects were highly uneven. Some trials showed marked increases after implementation, while others declined. Eleven trials showed negative intervention effects, and the spread across departments was wide enough to erase a pooled benefit.
The segmented negative binomial mixed models told the same story. In the crude model, the intervention effect was not statistically significant across all sites, with an incidence rate ratio of 1.731 and a 95% confidence interval from 0.922 to 3.251. After adjustment for criteria revision, the intervention estimate remained nonsignificant at 1.344 with a 95% confidence interval from 0.710 to 2.544. The confidence intervals crossed unity in both models, leaving no reliable evidence that turning the tools on, by itself, increased enrollment.
The data point to a familiar problem in digital recruitment: infrastructure is not the same as implementation. A system can identify more candidates on paper than a manual screen, but if the workflow is slow, the alert is ignored, or the eligibility definition is too blunt, enrollment may not change. In that sense, the bottleneck is not the database alone. It is the distance between a computable cohort and a real patient who is approached, consented, and enrolled.
Iterative Criteria Revision Emerged as the Strongest Driver
A more interesting pattern emerged around the revision of computable cohort definitions. The criteria revision covariate was associated with higher recruitment across all sites in the adjusted model, with an incidence rate ratio of 1.601 and a 95% confidence interval from 1.143 to 2.242. Most updates occurred early in the intervention phase, and the number of revisions varied markedly between trials and sites.
In plain clinical terms, the performance of the platform improved when the eligibility logic was refined rather than merely deployed. At the single KAS+ site, the criteria revision effect was especially strong, with an incidence rate ratio of 2.151 and a 95% confidence interval from 1.784 to 2.593. By contrast, the intervention effects for both recruIT and KAS+ were nonsignificant in exploratory stratified analyses.
The pattern is biologically and operationally plausible. Trial eligibility is often built on layered clinical judgment, not just diagnosis codes and age cutoffs. As those rules are rewritten to match real-world documentation patterns, the computable phenotype begins to resemble the clinical phenotype more closely, and the screening list becomes more usable. The strongest signal therefore came not from the launch of the system, but from the repeated adjustment of its logic.
What the Recruitment Data Mean for Trial Operations
This evaluation makes a narrow but important point: IT-supported patient recruitment can help, but only when the local setup is tuned to the trial and the site. A universal switch-on effect was not visible. Instead, benefit appeared context dependent and tied to iterative refinement of eligibility rules.
That has practical implications for research hospitals that hope to reduce manual chart review. Tools like recruIT and KAS+ can reorganize screening work, but they do not remove the need for local calibration, staff engagement, and data-quality review. Trials that rely heavily on laboratory thresholds or detailed structured observations may be especially difficult to automate if those fields are absent or inconsistently recorded in routine care.
The study also suggests that future evaluations should look beyond weekly enrollment counts alone. Identification-to-enrollment latency, workflow adherence, and the number of patients who move from electronic pre-screening to actual consent could reveal benefits that are invisible in aggregate recruitment totals. Trial heterogeneity remained substantial here, and that heterogeneity likely diluted any broader signal.
Residuals showed no major dispersion and model calibration was good, so the negative result does not read as a modeling artifact. It reads as a reminder that recruitment technology is only as effective as the eligibility logic, data completeness, and local clinical process surrounding it. The next step is not a more abstract platform claim. It is a tighter trial-by-trial assessment of how computable phenotypes are revised, who acts on the candidate list, and how quickly those candidates are converted into enrolled participants.