Machine learning and artificial intelligence (AI) are prominent buzzwords when the topic of precision medicine comes up, but like many emerging technologies with the potential to make a significant impact, it can be difficult sometimes to separate the hype from the real-world impact.
AI entered that breathless, overhyped territory last year, and luckily there have been some in the life sciences and healthcare who have cautioned against over-promising on the potential. “There is so much hype and noise about the predictive power of the AI,” said Ahmed Ghouri, CEO of clinical and genomic data interpretation company Interpreta. “This saturation of AI claims can make people sour on the field as a whole, so it is good to vet the accuracy of claims and ask organizations if they have any data that can validate their claims.”
In addition, AI still suffers from the fear of the “black box”—the algorithms that are grinding the data fed into them to generate answers or diagnoses, or advice on how to treat patients. “If [AI] is drawing conclusions and you don’t know what the underlying datasets are, that is what physicians will most react to. If they can’t understand how it arrived at that answer, they can’t trust it,” noted Chris Cournoyer, CEO of molecular decision support company N-of-One. “And trust is a big thing with AI right now.”
While care organizations that are looking to technology to help provide precision care are right to cast a critical eye toward AI, that doesn’t mean it is not making contributions today in the clinical setting—and is poised for even greater influence in the future.
Pharmacogenomics is one area where AI-enabled technologies are making a noticeable impact today. Companies like Interpreta are leveraging data from the FDA’s Adverse Event Reporting System and combining it with clinical data and insurance data to help doctors make more precise prescription choices and head off adverse drug events (ADEs).
“Right now most of the drug labeling for genomics tends to be focused on one-dimensional drug-to-gene interactions,” said Ghouri. But the reality of drug prescribing is much more complex than one drug to one specific genetic variant. “If you look at patient’s renal function, or if they have a heart murmur, the risk of an DE goes way up. So we are using the FDA [data] as a baseline, but then we are using AI to say a person with a certain condition has a 97% chance of having an ADE versus a 42% chance in the absence of the co-existing condition.”
Other data leveraged in this process—in real time—include those from the electronic medical records, insurance claims, pharmacy data, genomic data, and information from the National Committee for Quality Assurance (NCQA).
Rare Disease Diagnosis
Whole-genome sequencing has made significant inroads to help doctors sniff out and treat Mendelian diseases quickly in neo-natal care. While the high-mortality rate cases will get the press, AI is also making inroads in the diagnosis of rare diseases that may not be a matter of life and death, and can significantly shorten the diagnostic odyssey that awaits so many patients and families.
Facial analysis company FDNA is leveraging AI for facial patterns, essentially automating the task that geneticists have used for years of looking for specific facial traits as clues to the basis of rare genetic diseases. What FDNA does via its facial recognition engine is to effectively broaden the base of available facial images used to reach a diagnosis.
“Geneticists would look at tell-tale signs in patients’ faces and would take pictures of them, to help them reach a differential, or suspected diagnosis,” said Dekel Gelbman, CEO of FDNA. “We are able to use de-identified photos of patients, classify them into syndromes based on the common patterns in the face—much the way a geneticist would. But we are bringing a much broader set of data and different patterns that they might not have seen in their professional experience.”
The technology, when combined with the genomic data from patients, can also help tie variants of unknown significance (VUSes) to specific rare diseases, essentially linking the phenotype with the genotype. “As we look into the future AI will change the role of the geneticist and instead of being a diagnostician, they will become a treating physician, more involved in therapeutics and clinical development, and I think that is what geneticists want too,” Gelbman added.
At AI-powered genomic diagnostic company Freenome, company CEO Gabriel Otte sees the potential of AI to upend how diseases are diagnosed. He points out that most clinical diagnostics that have been on the market for years experience significant declines in accuracy, a function of the stripped-down approach often taken by companies developing these low-margin tests.
Freenome, which is developing blood-based diagnostics and is currently working to attain premarket approval for its colorectal cancer test, has taken an approach to identify as many biomarkers as possible in the blood and then apply these signatures using AI to the diagnosis of disease.
“We have affected artificial intelligence on the software side to hone in on the signals that are relevant for a specific test,” Otte said. “We do this in the software as opposed to in the laboratory. What this enables is, as we pick up on novel signatures that others haven’t, it allows us to change the test. It is about picking up all the signals in first place to make this a software problem and not a lab problem.”
Hurdles for AI
While AI is beginning to make an impact in clinical diagnosis and care, there still exist roadblocks to adoption. The biggest of these involve availability and accuracy of the data used to help train these systems.
“There is inherently a small data problem in AI applied to genomics especially, because we are never going to have billions of genomes sequenced,” said Otte. “Most of the AI engines out there that are being built assume large datasets. That is just never going to be the case for healthcare. So anyone using existing AI technologies for this purpose is likely not going to have much success.”
At N-of-One the view is a bit different, but it revolves both around the amount of data required and also the accuracy of the data, which is why the company employees its host of Ph.D.s and research scientists to annotate the data it leverages for clinical decision support.
In terms of using AI for highly precise cancer care, Cournoyer isn’t convinced it is ready for prime time.
“We went from AI winning at Jeopardy, to suddenly wanting it to provide information for highly targeted cancer care,” she said. “That’s too far, too fast.”
“I never want to come off as not supporting AI, because I do think there is a way here. We are going to have massive datasets eventually. We just need the right datasets and I don’t think we have the right datasets yet,” Cournoyer concluded.