At the moment, 44 vaccines targeted at Severe Acute Respiratory Syndrome Coronavirus (SARS-CoV-2), the causative agent of COVID-19 are undergoing clinical trials on human and at least another 91 preclinical vaccines being actively tested on animals. Of which, five had been approved for early or limited use.
As scientists in many parts of the World raced to end the pandemic by next year, Professor Mihaela van der Schaar, Director of the Cambridge Center for Artificial Intelligence (AI) in Medicine at the University of Cambridge believes COVID-19 presented a rare opportunity for the research community to consider and review novel approaches such as machine learning to foster some groundwork for faster and better clinical trials in the future.
Challenges presented by the pandemic on clinical trials
In the article written by Professor van der Schaar and her team scholars from higher institutions in the UK, US, as well as pharmaceutical companies and recently published in Statistics in Biopharmaceutical Research, prolong lockdown had prevented both researchers and subjects from accessing trial sites and created disturbances in the timely data collection process. In some instances, clinical trials were forced to go virtual or put on hold especially if scientists now regard COVID-19 as their new priorities.
Even if the clinical trials were given green light to proceed after the pandemic, the data collected before and after the global health crisis may be of different qualities. All these suggest the standard ways of conducting clinical trials are inefficient, time-consuming and non-flexible. As such, Professor van der Schaar and team proposed to leverage machine learning to extract insights from data of clinical trials that were suspended because of COVID-19 and use that to adjust recruitment plans, sample sizes and treatment allocations thereafter.
Moreover, machine learning may also support the creation of virtual control groups, whereby data-driven methods can be leveraged to integrated hospital data and single out patients to receive standard treatments while other similar patients to receive experimental treatments. Professor van der Schaar and team are confident machine learning including reinforcement learning, causal inference and Bayesian approaches can contribute in trials to repurpose or create new drugs to treat COVID-19 and in trials for drugs unrelated to COVID-19.
For examples, machine learning can support large scale clinical studies such as the ongoing Solidarity trial initiated by the World Health Organization (WHO) and the RECOVERY trial by the University of Oxford, to improve their “design, execution and evaluation”, to minimize the concerns arising from too many small studies that may not be able to generated convincing evidence.
The roles of machine learning
Similarly, machine learning may venture into large bodies of data generated by experimental and compassionate use of drugs to treat COVID-19 and select potential drug target for future clinical trials. Besides, machine learning may look into COVID-19’s biomolecular behaviors, to search for patterns and signatures that can facilitate repurposing or validating the effectiveness of existing drugs.
Ultimately, as new drug therapies get uncover, machine learning methodologies can pick up on the efficacy and toxicity of these new drugs, making it more efficient for future trials by reducing learning time. Professor van der Schaar and team wrote it’s natural for traditional biostatisticians to follow a similar path of relying on methods that have served well in find solutions; approaching each individual trial as a separate challenge, creating small datasets and so on.
What they are proposing in the paper, is a different path, by leveraging artificial intelligence (AI) on the enormous amount of data produced in different areas like epidemiology, operations research, systems biology, and even financing, for important insights to arrive at quicker and more effective solutions. So that at the end of the day, scientists will only have to deal with the challenge of deciding which machine learning approach to use, rather than the many hurdles to overcome along the way.