Guest Column | October 16, 2026

How New Models Could Improve Trial Enrollment Forecasts

By Uchemadu Nwachukwu

Pharmaceutical industry research, development data-GettyImages-1354838510

Enrollment rarely follows the assumptions made at the start of a clinical trial. Site activation can be delayed, recruitment can accelerate or slow, and early performance at one site may look very different from the broader study trajectory. That makes enrollment forecasting a moving target, particularly when decisions depend on knowing how many patients are likely to enroll and when.

Newer forecasting approaches are designed to account for that uncertainty. Statistical and machine learning models can use historical and emerging site-level data to update enrollment forecasts, quantify uncertainty, and identify changes in recruitment performance.^{1, 2, 4, 5} The goal is not to find a single model that works for every trial but to understand which approaches can produce forecasts that are more useful for real-world clinical trial planning.

Statistical Models For Dynamic Enrollment Forecasting

A natural starting point is a Poisson process. If site i enrolls at an average rate λᵢ, then the count of new participants in interval Δt can be modeled as:

Eᵢ,ₜ ∼ Poisson(λᵢΔt)

The basic Poisson model assumes that the rate is known and constant. Clinical trial enrollment rarely behaves that way: the recent enrollment literature instead uses Poisson–Gamma formulations, site-specific rates, country aggregation, restrictions, and interim updating to represent heterogeneity and uncertainty.^1,2 A Gamma prior for the site rate produces a Poisson–Gamma mixture, commonly represented as a negative-binomial predictive distribution. This introduces overdispersion and produces wider, more realistic prediction intervals.

If:

λᵢ ∼ Gamma(αᵢ, βᵢ)

and observed enrollment over exposure time is eᵢ, the posterior rate remains Gamma-distributed:

λᵢ | eᵢ ∼ Gamma(αᵢ + eᵢ, βᵢ + Tᵢ)

Here, Tᵢ is the site's observed enrollment exposure. The posterior mean is:

E[λᵢ | eᵢ] = (αᵢ + eᵢ) / (βᵢ + Tᵢ)

Non-Homogeneous Enrollment Processes

Enrollment rates are not necessarily constant. The 2023 portfolio planning study explicitly modeled country initiation, site activation, and subject enrollment sequentially rather than treating recruitment as a single fixed-rate process.⁴ Recruitment may accelerate after site activation, decline as the eligible patient pool is exhausted, or vary seasonally. A non-homogeneous Poisson process allows the rate to depend on time:

Eᵢ,ₜ ∼ Poisson(∫ₜ⁽ᵗ⁺Δᵗ⁾ λᵢ(s) ds)

The time-varying rate can be modeled with covariates:

log λᵢ(t) = β₀ + βᵀxᵢ + γᵀzᵢ,ₜ + f(t)

where xᵢ contains baseline site attributes, zᵢ,ₜ contains time-varying operational information, and f(t) captures a nonlinear enrollment trajectory. Relevant features may include indication, country, site type, historical recruitment, time since activation, screening volume, screen failure rate, competing studies, and recent enrollment momentum.

Machine Learning: Where It Adds Value

Machine learning is most useful when recruitment depends on many interacting variables and when the organization has enough historical data to evaluate generalization. Recent studies use real-world patient data and indication-level history for site ranking and combine clinical document representations with tabular features for probabilistic enrollment prediction.^{3, 5} A supervised model can predict either a continuous outcome, such as patients enrolled per month, or a probability, such as the chance that a site will recruit at least one patient during the next four weeks.

A generic site-level model is:

yᵢ,ₜ₊ₕ = fθ(xᵢ, zᵢ,₁:ₜ, c)

where xᵢ represents static site features, zᵢ,₁:ₜ is the time series of observed operational data, and c represents study-level context. Candidate algorithms include gradient-boosted trees, random forests, regularized generalized linear models, recurrent neural networks, temporal convolutional networks, and transformer-based models. The key modeling distinction is between ranking and calibration. Ranking identifies which sites are likely to recruit more patients. Calibration determines whether a predicted probability of 70% actually occurs approximately 70% of the time. Supply decisions require both. A model can rank sites well but still produce unsafe inventory quantities if its prediction intervals are too narrow.

Hulstaert and colleagues reported a 2024 machine learning approach that combined indication-level historical recruitment with real-world data to predict enrollment at the site level. In inflammatory bowel disease and multiple myeloma, the nonlinear model outperformed baseline and linear approaches on the test set and improved site ranking.³ This supports using real-world patient availability to improve the initial demand map, especially before trial-specific enrollment data exist.

A 2025 deep learning study combined language model representations of clinical documents with tabular trial features and used a probabilistic Gamma layer to produce uncertainty estimates. The authors reported improved prediction of the number of patients enrolled across sites compared with baseline methods.⁵ The technical lesson is not that deep learning should replace simpler models. Rather, unstructured documents and tabular operational data can be combined when there is sufficient data, but the output must retain uncertainty and be evaluated prospectively.

For many sponsors, gradient boosting with carefully engineered features may be preferable to deep learning because it is easier to audit, retrain, and explain. A practical architecture could use a survival or Poisson–Gamma model for the probabilistic backbone, a machine learning model for nonlinear covariate effects, and conformal prediction or calibrated quantile regression for prediction intervals.

Model Validation And Implementation Risks

Machine learning models add risks of leakage and poor transportability. If a feature is recorded only after an operational decision has been made, using it at prediction time can create leakage. Random train-test splits can also exaggerate performance when sites or investigators recur across studies. Time-based validation, study-level holdouts, calibration assessment, and prospective silent deployment are more credible.

Enrollment forecasting is not about finding one model that works for every trial. It is about using the information available as a study progresses to produce forecasts that can adapt to changing recruitment conditions and make uncertainty visible. Statistical and machine learning approaches can provide that flexibility, but their value ultimately depends on how well they perform on new studies and under real-world conditions.

References:

Anisimov V, Austin M. “Modeling restricted enrollment and optimal cost-efficient design in multicenter clinical trials [Internet].” 2022 [cited Sept. 18, 2026]. https://www.semanticscholar.org/paper/2fa8ba95d345303f1a76f7900f1650cd66bb8d89
Anisimov V. An analytic methodology for forecasting patient enrolment performance in multicentre clinical trials. In: Proceedings of the 2023 ACM Conference; 2023 [cited Sept. 18, 2026]. https://dl.acm.org/doi/pdf/10.1145/3613347.3613361
Hulstaert L, Twick I, Sarsour K, Verstraete H. Enhancing site selection strategies in clinical trial recruitment using real-world data modeling. PLoS One. 2024;19(3). doi:10.1371/journal.pone.0300109.
Zhong S, Xing Y, Yu M, Wang L. Enrollment forecast for clinical trials at the portfolio planning phase based on site-level historical data [Internet]. 2023 [cited Sept. 18, 2026].
Do T, Masquelier A, Lee NE, Crowther J. Deep learning-based prediction of clinical trial enrollment with uncertainty estimates [Internet]. 2025 [cited Sept. 18, 2026].

About The Author:

Uchemadu Nwachukwu is a public health data analyst, health informatics professional, and clinical data management professional with over eight years of experience across healthcare analytics, clinical research, health informatics, and real-world evidence. He holds a First-Class MSc in applied clinical data analytics from the University of Galway and has expertise in statistical modeling, machine learning, epidemiological analysis, healthcare data management, interoperability, and clinical information systems. His work focuses on applying data, technology, and analytical methods to support clinical research, healthcare improvement, and evidence-based decision-making. His professional interests include digital health, health informatics, clinical research, healthcare analytics, epidemiology, and data-driven approaches to improving healthcare systems and outcomes.