Patient Selection for Clinical Trials

The success of epidemiological studies and clinical trials depends on the selection of the right patients. Rapid and efficient patient selection for clinical trials requires access and interrogation of large electronic medical records which are time-consuming because of the large number of patient records that have to be manually reviewed by investigators. This process is additionally challenging because of variations due to a lack of standards in recording patient information, medical coding mistakes, sparse data or missing details, among other issues. Logical rules that encode trial inclusion and exclusion criteria are widely used to identify patients in electronic medical records; however, they have several limitations including lower accuracy, the need for extensive additional work involving experts and limited reuse in other clinical trials due to the rules being generally trial-specific (Kirby et al., 2016). Automated natural language processing (NLP) methods can alleviate the manual review burden and improve accuracy, but their performance in clinical practice is unclear. Classical machine learning classifiers can automatically learn patterns to identify these patients. However, they still require human expertise to define the most informative feature set for the task.

One example is the NLP- and machine learning-based system - Automated Clinical Trial Eligibility Screener (ACTES) (Ni, Bermudez, Kennebeck, Liddy-Hicks, & Dexheimer, 2019) which automated subject identification for clinical trials and improved the numbers of subjects screened, approached and enrolled by 14.7%, 11.1% and 11.1%, respectively. Although the ACTES achieved an overall performance of 90.3% (micro F-measure), known shortcomings of NLP in understanding language semantics (e.g., word sense disambiguation) and syntax (e.g., assertion detection) resulted in multiple types of false-positive recommendations (Ni et al., 2015).

Consequently, there is a need to explore novel methods that can identify the most suitable set of patients for any clinical trial, independently of the criteria used and with minimal human intervention. Newer deep learning methods promise to improve this process further. One of their greatest advantages is that they can automatically identify the most appropriate features from the raw data and text directly, without the need for expensive manual guidance (Segura-Bedmar & Raez, 2019; Xiong et al., 2019)- Nonadherence and dropout in clinical trials can lower study power, reduce the magnitude of treatment effects, and increase trial cost and duration. Machine learning methods have shown promise in predicting treatment discontinuation (Lutz et al., 2018) and in medication adherence (Lee, Kang, Kim, & Son, 2013; Lo-Ciganic et al., 2015). Similar approaches can be used to predict dropout in clinical trials.

Synthetic Data in Clinical Trials

“Synthetic controls” comes from methods developed for website analytics and economics research (Brodersen, Gallusser, Koehler, Remy, & Scott, 2015) where a number of time series that are unaffected by the intervention are optimally weighted according to their fit to the outcome of interest in the period before the intervention, then combined into a composite time series (Abadie, Diamond, & Hainmueller, 2010). Application of synthetic controls method to nationwide administrative databases in Brazil, Chile, Ecuador, Mexico and the United States to evaluate changes in the burden of hospitalizations for all-cause pneumonia associated with the introduction of pneumococcal conjugate vaccines (PCVs) did not detect a decline in all-cause pneumonia in older adults in any country (Bruhn et al., 2017) which had implications on healthcare policies on more widespread use of the vaccine.

Another potential use of synthetic data is in generating synthetic controls or placebo groups for trials. Although a placebo control arm is crucial in clinical trials to determine treatment effects, participants generally do not like the possibility of being placed in the placebo group. One option being explored is the use of synthetic control arms, which are in-silico placebo arms modelled using information that has previously been collected including historical control data, real-world data or the generation of a companion data set from other sources to serve as a comparator. Indeed, the FDA recognizes clinical trials that use this form of hybrid design where real-world data can be used as a basis for external controls.

The second scenario where machine learning can help by generating synthetic data is to overcome the challenges of sharing individual-level data from clinical trials (El Emam, Rodgers, & Malin, 2015). Data sharing requires formal collaboration and extensive data usage agreements between researchers that are time-consuming, with many ultimately resulting in failure. Generative adversarial networks (GANs) have been very successfully used to generate synthetic images in medical imaging, ophthalmology and dermatology (Chi, Bi, Kim, Feng, & Kumar, 2018; Yi, Walia, & Babyn, 2019; Yu et al., 2019) and subsequently extended to electronic health records (Baowaly, Lin, Liu, & Chen, 2019). GANs are deep neural net architectures comprised of two nets, pitting one against the other which in simple terms allows prediction of features given a label in contrast to discriminative algorithms which map features to labels (Goodfellow et al., 2014). A proof-of-concept application to generate synthetic trial data for sharing used GAN with differential privacy recognizing that traditional GAN could learn to create synthetic data that reveals actual participant data (Beaulieu-Jones et al., 2019). They achieved differential privacy by limiting the maximum influence of any single participant during training and then adding a small amount of random noise (Abadi et al., 2016) on the SPRINT (Systolic Blood Pressure Trial) data set (Wright, Whelton, & Reboussin, 2016). The key advantage of using synthetic data from electronic medical records is that they are artificially created and hence do not correspond to real patients and thus do not pose any danger of re-identification. As synthetic data carry attributes similar to actual data, it has value in public use of the information without the hassle of obtaining real data. This proof of concept is promising, and the next steps are resolving a known weakness of GAN’s inability to deal with discrete data and testing and validation to ensure accurate implementation.