OBJECTIVES:

Patient recruitment is one of the most challenging steps in running a randomized clinical trial. According to ClinicalTrials.gov, approximately 38.7% of terminated trials were caused by insufficient patient enrollment. One way to achieve successful patient recruitment is through the prediction of enrollment rates at different locations (e.g., country level), so that one can recruit patients in countries that are likely to have higher enrollment rates. In this work, we focus on developing a novel machine learning model that predicts country-level enrollment rates for trials at the early planning stage.

METHODS:

We model the country-level enrollment rate prediction problem as a tensor completion problem, where the tensor consists of partially observed enrollment rates from historical trials and its dimensions specify the information of countries and those from the trials such as phase, therapeutic area, primary indication, etc. We propose a Joint mUlti-linear and nonLinear IdentificAtion (JULIA) approach to solve this problem. JULIA is comprised of two networks: a multi-linear network that learns multi-linear latent components and a nonlinear deep neural network that learns nonlinear latent components from incomplete tensors.

RESULTS:

In this study, we leverage a dataset consisting of nearly 40,000 clinical trials, including more than 100 indications from 60 different countries. The experimental results reveal that JULIA can reduce the Mean Absolute Percentage Error (MAPE) and the Root Mean Squared Error (RMSE) by up to 50% compared with the best human estimations currently available.

CONCLUSIONS:

In this study, we formulate the country-level enrollment rate prediction task as a tensor completion problem and leverage an advanced tensor completion model, namely JULIA, to successfully make enrollment predictions prior to the initiation of clinical trials. Experiments based on a large and comprehensive sample of global clinical trials demonstrate that JULIA can make more accurate enrollment rate predictions compared with the best human estimations currently available.