Abstract
This multi-source retrospective study highlighted key features influencing RCT participant recruitment, offering actionable steps for improvement, including flexible recruitment infrastructure and appropriate participant compensation.
Introduction: The Recruitment Bottleneck
Recruitment is often cited as the "Achilles' heel" of clinical research. Statistics indicate that approximately 85% of clinical trials fail to recruit a sufficient number of participants on time, and nearly 80% fail to meet their initial accrual targets entirely. These failures result in significant financial losses, delays in bringing life-saving treatments to market, and underpowered studies that lack the statistical power to definitively prove a drug's efficacy.
While previous research has often focused on specific diseases or isolated factors, this study adopts a broad, data-driven approach. By analyzing a wide array of trial design features—ranging from funding sources to specific recruitment tools—the researchers aimed to identify the primary drivers of "accrual percentage," defined as the ratio of participants recruited to the target goal.
Methodology: A Machine Learning Approach
The researchers framed recruitment success as a regression problem, evaluating eight different models to predict accrual percentages. These included linear models like Ridge and Lasso regression, as well as various tree-based models such as Decision Trees, Random Forest, AdaBoost, XGBoost, LightGBM, and CatBoost.
CatBoost emerged as the most effective model, likely due to its robust handling of categorical data. To interpret the model's "black box" predictions, the team utilized Tree SHAP (SHapley Additive exPlanations). This game-theory-based method allowed for the ranking of features by importance and the determination of whether a specific feature had a positive or negative impact on recruitment outcomes.
Data Integration
The study employed a multi-source approach, linking two primary databases: RASCAL, an internal electronic research registry from Columbia University Irving Medical Center, and AACT (Aggregate Analysis of ClinicalTrials.gov), which provides standardized global trial data. The final dataset consisted of 393 randomized controlled trials (RCTs) that were interventional in nature, focused on treatment, and had concluded their enrollment phase.
Key Insights: Drivers of Recruitment Success
The analysis revealed several critical factors that influence recruitment:
- The Digital Recruitment Paradox: Surprisingly, the use of digital methods such as websites, radio, and television ads was negatively associated with recruitment success. This suggests a potential selection bias: researchers may resort to these broad methods only when traditional recruitment channels are failing.
- Funding Sources: Government funding (Federal, State, or Local) was a strong positive predictor of success. Conversely, industry-funded trials often showed lower accrual percentages, possibly due to more rigid protocols and tighter timelines.
- Compensation and Equity: Participant compensation was identified as a top-tier feature for success. Beyond serving as an incentive, compensation helps mitigate the "participation burden"—such as travel costs and lost wages—making trials more accessible to economically disadvantaged populations.
- Oncology Challenges: Trials focused on cancer research (Target domain: C04) consistently faced lower recruitment success. This likely reflects the high level of competition for patients in oncology and the stringent eligibility criteria common in these studies.
Results Summary
The CatBoost model achieved the best performance with a Root Mean Square Error (RMSE) of 20.31. Across the 393 trials, the average accrual percentage was only 46.7%, highlighting the prevalence of recruitment challenges. Comparing the most successful trials (≥90% accrual) with the least successful (≤10% accrual) revealed statistically significant differences in 15 of the top 30 features. Notably, longer protocol durations were positively correlated with higher accrual percentages, suggesting that extended timelines may be necessary for reaching recruitment goals.
Limitations and Future Directions
The study's retrospective nature means that identified links are correlational rather than causal. Additionally, the data is derived from a single large academic center, which may limit generalizability. The dataset also lacked information on eligibility complexity and local competition from other trials. Future research could leverage Natural Language Processing (NLP) to analyze the text of clinical trial protocols, allowing for the extraction of more granular features like inclusion and exclusion criteria nuances.
Summary updated