Dr Kamala Maddali looks at how artificial intelligence is currently being utilized across the healthcare spectrum and why the time is now for Big Pharma to adopt this technology more fully.
The beauty of artificial intelligence is that it allows one to assess both forest and tree.
A case in point. Two years ago, researchers from Boston University’s Center for Information and Systems Engineering conducted a study that, in my mind, proves the worth of machine learning. This branch of AI can predict outcomes, once it is told, via algorithms, what to look for among the reams of EHR patient information in its database. Researchers wanted to know how accurately machine learning could predict the percentage of more than 45,000 patients with heart-related diseases who would need hospitalization within a year, based on ten years of EHR charts; 60 percent of the patient data trained the algorithms, 40 percent evaluated their performance.
AI predicted 82 percent would need hospitalization (with 30 percent false alarms). These patients’ physicians, determining individualized treatment on the evidence-based findings from the entire database, began proactive disease management, keeping patients out of the hospital and saving their payers money. To get an idea of just how much money, an AHRQ study reports that USD 30.8 billion in hospitalizations might have been saved in 2006 with just better care – frequent doctor visits, adherence to medications – of and by patients with any chronic disease.
With all this in mind, I keep imagining what more machine learning can do.
The time seems right for pharma to switch business models. Some observers call the existing one fractured
While its relevance in the discovery of gene therapy has already been made evident, artificial intelligence, in theory, could help manage so many chronic diseases. With all patients’ details transferred from an EMR and then put into the database, we could establish an effective, evidence-based approach for patient care. Yes, all details on the chart: researchers found the more information, the better the prediction rate would be.
Certainly, members of the scientific community have begun capitalizing on the value of artificial intelligence, especially those whose specialties and subspecialties are image-reliant, like ophthalmology and cardiology, both of which have patient registries. Even new, tech-focused medical specialties are emerging, like clinical informatics.
But big-name pharma? You will find none of the founding members of the nascent Alliance for Artificial Intelligence in Healthcare to be a household name.
The time seems right for pharma to switch business models. Some observers call the existing one fractured. One author predicts that the Law of Diminishing Returns will wipe out any internal rate of return by next year, as the creation and approval of each new drug, built on the back of its predecessor, becomes more difficult to develop and hence more expensive to do so. The reasons for the huge surge in collaborations, mergers and acquisitions are obvious: It is, for example, more efficient, from both time- and money-saving perspectives, for Novo Nordisk to comb through the genomics information of type 2 patients contained in the interactive networks belonging to drug discovery firm e-therapeutics than to buy the technology and train its own staffers.
The question does arise, however: When would the job descriptor R&D flip positions with biotech matchmaker from first to second?
The time savings can be significant. In one study on diabetic retinopathy, deep learning, a type of machine learning, was used to assess 93,293 fundus photos to estimate disease risk factors and prevalence in many ethnic peoples. Machine, about 1 month: 17 humans, 2 years. A study from Denmark used another type of AI to evaluate diagnostic accuracy. Of 11 studies looked at, the lowest accuracy rate was 78 percent.
And the financial savings could be significant. In 2018, MIT’s Sloan School of Management published its study on clinical trial successes and failures. It found that overall, 14 percent of clinical studies between 2000 and 2015 passed FDA approval. Oncology had the lowest approval rate in the early years of the study, but rose to 8.3 percent in 2015, partly attributable to immunotherapies.
Clinical trials that use biomarkers have a greater chance of succeeding, mainly because of better patient stratification. The study also said investors would likely watch which disease states fare better at the phase 3 stage.
Judging from clinicaltrials.gov, pharma isn’t interested in using AI in its diabetes work, or at least invest in the time it takes to learn about AI. Most of the eight trials listed are non-US based; many of the 19 CVD trials are as well. Other issues include the fact that it would be difficult, as well as exhausting, to extract digitally stored paper diaries for AI use. Misspelt words, jargon, and any other aspect that represents a hurdle for natural-language-processing algorithms will disrupt accuracy. Kitai et al. point out in their 2017 article on precision cardiovascular medicine that “ignorance of the challenges of AI may overshadow the impact of AI on CV medicine.”
We have serious decisions to make about our future, meaning healthcare’s future. We now have a way to become more effective at helping our patients become pain-free, even disease-free, and a way to help the industry become more efficient with drug development. The choice is ours.
Dr Kamala K. Maddali is the President and Founder of Health Collaborations, a Precision Medicine focused Consulting organization on a passionate path to bring change in healthcare strategies via innovation and patient engagement.
