Artificial intelligence (AI) has been around for decades, and yet it remains a nebulous subject for many. “During the ‘70s and ‘80s, the field of AI tried to understand how experts make decisions,” says Tod Klingler, Chief Architect, Roche Diagnostics Information Solutions (DIS), “and build expert systems to emulate human decision-making.” MYCIN, a system built in the ‘70s to identify bacteria causing severe infections and recommend antibiotics, is one such early example. To this day expert systems haven’t replaced human experts but many other advances in AI, such as machine learning, are finding applications in medicine.
Machine learning (ML), which can be thought of as a subset of AI, is the dominant form of AI these days, Klingler explains. Using data as the starting point and given a stated goal, machine learning uses computers to derive models that lead to the desired answers. “Machine learning is playing a big role in molecular diagnostics today,” he goes on, “which examines genomic markers to develop a diagnosis or prognosis. An example is the Oncotype DX test, which uses multiple genes and a classification algorithm to predict how likely breast cancer is to return and how aggressive treatment should be.” The AlloMap blood test, which Klingler helped develop, is another example of machine learning in the service of medicine. AlloMap creates a score, based on the expression levels of 11 genes, which is used to identify the risk of organ transplant rejection.
One of the latest applications of AI is deep learning, a subset of machine learning that uses multi-layered neural networks to learn from vast amounts of data. “Deep learning is already being used in image recognition. Radiology-based diagnostic systems today analyze radiological images to diagnose tuberculosis and lung cancers,” Klingler continues. “Compared to expert radiologists, these systems have been shown to have as good or better performance on that task.”
Clinical decision support systems (CDSS) are often powered by machine learning. The NAVIFY Clinical Decision Support applications, which sit on top of the NAVIFY Tumor Board, will increasingly use machine learning and deep learning to assist cancer care clinical teams. “Machine learning applications need data, first and foremost,” explains Klingler. “Here at Roche we have a ton of such data. We have access to clinical databases, digital pathology, genomics and real-world data. And then we have very strong clinical expertise, particularly in cancer. Now we are building our capabilities in clinical data sciences, biostatistics, machine learning and deep learning. This is where all the Roche assets come together in medicine.”