A recent study published in the peer-reviewed European Respiratory Journal details how PAREXEL and its partners leveraged machine learning-based texture analysis to analyze medical images, detect disease and correlate results to a real-world clinical drug trial. PAREXEL Insights sat down with Rohit Sood, Vice President, Scientific and Medical Services, to discuss this study – among the first uses of machine learning in medical imaging for drug development.
First, can you tell us about this study and how machine learning was leveraged?
Rohit Sood (RS): This Phase IIB/III study was in patients with a condition called Idiopathic Pulmonary Fibrosis (IPF), a chronic, progressive lung disease of unknown cause, with high morbidity and mortality if not treated. Medical imaging, especially high-resolution computed tomography (HCRT), plays an important role in the diagnosis of IPF, and can also be used for assessment of response to therapy in IPF clinical trials. This involves expert radiologists (known as readers) qualitatively comparing follow-up scans with baseline HRCT scans. However, this type of assessment is subjective and can lead to variances between readers.
It was clear from PAREXEL’s experience running IPF studies that there was a strong need for objective methods to detect change in disease burden with high sensitivity and repeatability. In addition, the method must not be influenced by variability arising from use of different scanner types in multi-site clinical trials. In 2012, we started work to address this need in collaboration with a world-leading academic lab. Together, we worked to develop an analysis method that could assess change based on differences in textures or features in images and then provide a quantitative measure as an output. However, due to the large amount of data HRCT scans can generate, we knew it would not be possible to use the manual method of having radiologists review the data to detect and evaluate changes. By employing machine learning algorithms, we were able to train machines to automatically detect features by learning from large amounts of generated data, eliminating the need for manual review.
Can you explain the technology and how exactly it was applied?
RS: While Artificial Intelligence (AI) is a topic that has been talked about a lot recently in the clinical research industry and has received a lot of hype, this is one of the first real-world use cases. Machine learning, a type of AI, is the science of getting computers to learn and act like humans do. Researchers improve the algorithms over time by feeding them data and information in the form of observations and real-world interactions. It is important to note that data representation (e.g. features and textures) for machine learning is different from human learning, and a lot of effort goes into defining what the representation should be. A simple analogy of this in our world - people can easily perform arithmetic on Arabic numerals (1, 2, 3…), but find arithmetic on Roman numerals (I, II, III…) much more time-consuming. In this case, the concept of math hasn’t changed, but its representation is very different. Interestingly, while we can teach machines to play chess relatively easily (and also be defeated by them), it is much harder to teach a person’s everyday life or profession to machines. Much of this knowledge is subjective and intuitive, which makes it difficult to articulate. This is a bigger challenge for data representation.
Over a six-year period, between 2012 and 2018, we defined, tested and validated several different approaches to extract textures from HRCT scans, created a dictionary that could be used to teach machines and then actually trained the machines with optimal machine learning algorithms. Once validated on HRCT training datasets and against expert radiologists, we applied the workflow to a real-world clinical trial. The technology was able to recognize, classify and quantify patterns in medical images, analyzing them to detect disease and correlate results to currently accepted clinical and laboratory measures used for evaluation of response to therapy. We also compared the results with currently accepted and regulatory approved endpoints in IPF trials, which were published in the European Respiratory Journal. This was done at a level and speed that could not be accomplished by manual analysis of the scans. The machine analyzed 60,000 images from over 500 subjects in a few hours – a team of radiologists would take days to perform the same task.
How does this technology impact the roles of those conducting clinical research?
RS: Machine learning impacts roles in many ways. First, it reduces manual processing by leveraging the machine’s ability to learn and process data at a significantly faster rate and at a level of analysis previously deemed impossible. In addition, machine learning provides objective assessment of clinical endpoints with high reproducibility. This has significant implications for clinical research, as it can reduce costs and save time in clinical trials, ultimately supporting faster delivery of important therapies to patients in need.
This application of machine learning clearly has a significant impact for those conducting clinical trials. How does it impact the patients?
RS: Applying machine learning to clinical trials has the potential to reduce the number of patients needed to participate in large clinical studies because of the massive datasets now available for fast quantitative analysis. For those enrolled in trials, machine learning can potentially reduce the number of tests needed while on trial. This, in turn, would reduce the trial’s duration, getting new drugs to market faster.
Do you see this approach being adapted and deployed in other types of trials? Is there a type of trial that this application would be best suited for?
RS: The methodology used in this trial - applying machine learning to medical imaging - can be used across different indications and trials by teaching the machines to learn different features and textures. However, the application of machine learning is not restricted to just imaging – any device or method that makes decisions based on data and generates large amounts of data over a period of time can leverage this method. Examples include wearable sensors (e.g., Apple Watch, FitBit) that continuously monitor physiological parameters and tend to generate large amounts of data. This is another area of focus for PAREXEL.
How will this development impact drug development and what does it mean for the industry and healthcare more broadly?
RS: Machine learning is already changing the drug development landscape. Large amounts of complex health data can be analyzed in different ways in real-time and fed back to the treating physician, who can make decisions for the patient’s benefit. Machine learning allows us to do this automatically, faster, and, using an effective model, optimization and cross-validation, with higher precision and sensitivity than ever before, and will ultimately positively impact the patient by supporting enhanced care and delivering important new therapies faster.