The Cambridge Centre for AI in Medicine is investigating how to optimise patient recruitment for enrolment in clinical trials and how to conduct more efficient, responsive trials. This work is informing the next generation of adaptive clinical trials, including determining sequential recruitment of patients, arriving at effective drug dosages and much more.

To compare the effectiveness of a new treatment with an existing treatment or a placebo, randomised controlled trials (RCTs) are considered the gold standard. The problem is that such clinical trials are often slow, costly and lack flexibility. In addition, as a result of insufficient analytical power, these trials may fail to uncover specific sub-populations of patients for which a treatment would be most effective.

What’s more, conclusions drawn from RCTs are typically valid only for the types of patients recruited for that RCT. This is an especially significant problem in the case of COVID-19, for example, as elderly patients and patients with comorbidities, who are known to be at higher risk, are typically excluded from RCTs.

Recent work on improving the design of adaptive clinical trials, by members of CCAIM and others, has revealed that the efficiency and effectiveness of many aspects of clinical trials can be significantly boosted through the application of machine learning. Here are just a few examples.

Adaptive, dynamic design

What makes ML-enhanced clinical trials so desirable? While traditional RCTs simply allocate patients to treatment and control groups through uniform randomisation, this procedure can be far from optimal in terms of learning. ML-enhanced trials can recruit patients in cohorts rather than all at once, allowing the treatment effects on each cohort to be observed before deciding who in the subsequent cohort should be assigned to the treatment or control groups, speeding up learning.

With AI/ML technology analysing a wide range of trial-participant data, and their responses to treatment, it becomes possible to identify sub-populations of patients with particular characteristics that are responding in different ways to the treatment. Such techniques have been shown to significantly reduce error and achieve a prescribed level of confidence in their result while requiring fewer patients – a double win.

Another example of where ML techniques can boost adaptive clinical trials is in Phase I dose-finding trials. These trials are increasingly challenging as the relationship between efficacy and toxicity of new compounds or combinations becomes more complex. Despite this, most commonly used methods in practice focus on identifying a Maximum Tolerated Dose (MTD) by learning only from toxicity events. ML techniques can outperform state-of-the-art clinical trial designs by finding the optimal dose with higher success rate and fewer patients.