The evolution of machine learning in oncology: an interview with Dr Ofer Sharon
Original source here
Please could you introduce yourself, your scientific background and current scientific interests?
My name is Ofer Sharon, I am the CEO of OncoHost (Binyamina, Israel) and a physician by training. Over the last 20 years, I’ve been building/founding companies that are dealing with the interaction or the intersection between mathematics and medicine, trying to utilize machine learning algorithms and artificial intelligence in clinical scenarios where it may improve the clinical decision making. I used to work as a Chief Medical Officer and researcher, and I was part of the early clinical trials for new immunotherapies that are now being used in several cancer types.
How do you think machine learning has evolved in healthcare over the last 5 years?
Machine learning has been a big promise for many years in medicine, and we have begun to see the widespread use of machine learning technologies. For example, ML is now being used in image processing, which makes a lot of sense, because it’s much easier for the machine to detect anomalies in huge amounts of data than it is for a human being. In the last five years, we have seen an increase of machine learning use in drug development, drug design, and now also in prediction of treatment response and analysis of the biological pathways that are involved in host (patient) response. Its importance to understand the complexity of the biological system – it’s huge, making it very hard to analyze everything at once. I think this is where the promise lies – in the ability to see the big picture, but also all the details. Additionally, in recent years we have begun to see that the anomalies point to places of interest for clinical decision making.
What do you think the current challenges for machine learning are in oncology itself?
I think there’s a lot of confusion about how machine learning can be used in oncology. Some people think machine learning is here to replace the clinician; to replace the oncologist who is treating the patient. I don’t think this is the case. Personally, I don’t think this challenge will materialize – namely, the replacement of human beings with machine learning in the clinical decision-making process. I think one of the challenges we face today is dealing with this misunderstanding of how machine learning can be used. It’s a challenge because many people are afraid of this concept, saying, “I don’t see a machine replacing me as a clinician,” or “As a patient, I don’t want to be treated by a machine.” The correct way to use machine learning and artificial intelligence in medicine is not as a replacement, but as an add-on.
Another big challenge for us as entrepreneurs in the area of clinical research is to develop an interface that would allow the seamless combination between human being and artificial intelligence. We are facing issues of disbelief; we are facing issues of concerns from artificial intelligence, but we are also facing issues that are merely human factor related. How to create an intersection between man and machine in order to produce a better result – I think this is what we are going to see evolving in the coming years.
Could you tell us about OncoHost’s PROphet®?
PROphet was designed to deal with a very interesting question: Why do some patients respond to immunotherapy, and others do not?
There’s been much progress in the way we treat cancer today. We see people who five years ago wouldn’t have had a chance to survive with advanced cancer, now living longer, fruitful lives. However, the challenge remains because many patients are not responsive to treatment. We do not have the ability to identify who will respond and who will not respond to a given therapy, which patient will benefit from the treatment and which patient will not. Finding more accurate ways to predict the success of immunotherapy treatments has been identified as a top research priority by the American Society of Clinical Oncology.
In order for us to be able to identify which patients will benefit from treatment so we can personalize it to them, we require biomarkers, and this is where the challenge lies. We currently do not have good biomarkers that can guide us on how to personalize treatment for patients. This is exactly what we at OncoHost are focusing on. At OncoHost, we combine proteomic analysis (analysis of proteins in the blood) with artificial intelligence and machine learning tools in order to provide two things. First, a prediction of whether the patient is going to respond to their given treatment or not. Second, analysis of the biological pathways and the “resistance to treatment mechanisms” that exist in that specific patient. As such, this provides clinicians with strategies for disease management and potential combination treatment plans.
How does PROphet® help to overcome some challenges in machine learning?
Our approach lies with what I mentioned above – that the right way to use machine learning is to create an intersection between man and machine to produce a better result. We aren’t trying to create a platform that will replace the physician in clinical decision-making. What we’re trying to do is create a platform that supports the clinician in their decision-making process.
When a clinician sits in front of an advanced cancer patient, unfortunately many of their decisions are being made in the dark. We don’t know enough – we have limited modalities to understand what’s going on in the patient’s body, especially considering the complexity of the tumor-body interaction. At OncoHost, we utilize a phenomenon called “host response,” which is the biological reaction of the patient’s body to the anticancer treatment. Together with machine learning tools, we aim to provide clinicians with a map to guide their treatment plans for each individual patient, and his steps moving forward. To reiterate, I believe this is the best way to use machine learning – not as a replacement, but as an enhancement to the clinician’s decision-making capabilities.
What are the next steps for PROphet®?
As with every machine learning based platform, the more patients we test, the better the system becomes. Every patient who joins our platform through our clinical trial initiatives makes the algorithms more accurate. We invest a great deal of effort into our clinical trials and are currently running a multicenter, prospective trial in Israel, the US and soon in the UK. These trials are testing more and more patients in order to find the algorithm capability to predict response. For me, the next steps are following the commercial launch of the first two indications, which are non-small cell lung cancer (NSCLC) and melanoma, and broadening PROphet’s capability of prediction for other cancer indications.
How do you hope machine learning will evolve over the next 5 years?
Let’s focus on how machine learning will evolve in oncology, because machine learning will likely be a part of every aspect of medicine in the coming years.
When it comes to medical imaging, machine learning is already being utilized to a great degree, but it’s starting to make its first steps toward infectious disease, and very soon oncology will be the next big wave. We can anticipate seeing a widespread addition of machine learning-based tools in clinicians’ decision-making processes, and hopefully also in national clinical guidelines.