Artificial intelligence – once only found in a Steven Spielberg movie – is being adopted by the healthcare industry more and more from improving diabetes outcomes to reducing unnecessary thyroid surgeries. As these technologies improve, clinicians will soon have yet another ally to individualize patient care.
Artificial intelligence. Machine learning. Those very phrases stir up Philip K. Dick’s paranoid vision of replicants or the eerie, calm voice of Stanley Kubrick’s sadistic spacecraft control system.
But it’s best to leave the more ominous connotations of machine learning where they belong: in the realm of science fiction. Here, at the end of 2018, the world is seeing a rise in new technologies that employ artificial intelligence and machine-learning techniques – apps that use algorithms to predict when someone might be craving pizza; self-driving cars that learn not only what other cars might do on a busy highway, but whether a pedestrian might walk when it’s not his or her turn; robots that can do parkour. Several in the healthcare industry and in clinical settings are already seeing the benefits from this brave new world, and they’re hoping to soon pass these benefits on to patients, especially as personalized, individualized medicine becomes the standard for optimal care.
For example, several thyroid experts recently presented their findings from real-world data they collected using a new version of a test that uses machine learning to determine whether nodules were benign or malignant. The goal here is to increase the identification of benign nodules and save patients from the risk and expense of unnecessary surgeries. This test uses an ecosystem of algorithms to analyze data from deep RNA sequencing so that it can leverage more enriched and previously undetectable genomic information.
“We are employing the same machine learning methods that are being used in other fields such as social media and self-driving cars, but applying them to thyroid cancer diagnosis,” says Giulia C. Kennedy, PhD, chief scientific officer of Veracyte, the company marketing the aforementioned test as the Afrima Genomic Sequencing Classifier. “Our approach uses RNA sequencing to interrogate the entire genome, and takes advantage of newer methods in machine learning to combine valuable features that provide a higher-resolution genomic picture of thyroid nodules.”
Advances using artificial intelligence are being made in diabetes care as well; new devices and apps that help patients control their glucose levels are being developed and brought to market at an ever-increasing pace. One of these new apps is called Sugar.IQ, the result of a partnership between IBM Watson Health and Medtronic. The app works in tandem with Medtronic’s Guardian Connect Continuous Glucose Monitor and is designed to learn a patient’s patterns and habits so that the app can predict a high or low and the patient can make better decisions for self-care.
“Sugar.IQ as an artificial intelligence will get to know your patterns over time,” says Lisa Latts, MD, deputy chief health officer at IBM Watson Health. “We’re basically creatures of habit, and we tend to eat the same things over time. We tend to get the same amount of exercise at the same time of day. And as it gets to know you through the artificial intelligence components it starts to be able to surface insights about your patterns and whether you’re likely to develop a very high blood sugar after something that’s coming up or develop a very low sugar. And then you can take action based on those insights with a goal of spending more time in range.”
Patients with suspicious thyroid nodules and patients with diabetes (especially type 1 diabetes) are seeing positive results using these technologies that can learn and predict and prevent adverse health outcomes. As we close out 2018 and welcome 2019, we’ll take a look at what insights physicians using these products are gaining and what those insights might mean for the future of tailored endocrine care.
Reducing Unnecessary Thyroid Surgery
Researchers at Cleveland Clinic wanted to determine just how well this Afrima Genomic Sequencing Classifier (GSC) performs, or, more specifically, how much of an improvement the GSC is over its predecessor, the Gene Expression Classifier (which didn’t use machine-learning techniques). They compared results of 46 samples tested with the GSC between July 2017 and December 2017 with 182 samples tested with the GEC between December 2011 and July 2017and found that the GSC identified 67.4% as benign, compared to 41.8% with the GEC – an increase of 61%. The overall surgery rate for nodules tested with the GSC was 32.6%, compared to 47.3% with the original test, a decrease of 31%. The results were presented in Washington, D.C., at the annual meeting of the American Thyroid Association.
“I think the biggest thing is to understand how cutting-edge molecular diagnostics are improving care,” Alexander says. “And I think with continual evolution of these tasks, they’re improving. And we are also increasingly individualizing care to each specific patient.” – Erik Alexander, MD, chief, Thyroid Section of the Division of Endocrinology, Diabetes and Hypertension, Brigham and Women’s Hospital, Boston, Mass.
The researchers looked mostly at thyroid nodules in the Bethesda 3 and Bethesda 4 categories, where the risk of malignancy ranges from about 20% to 50%, according Christian Nasr, MD, medical director of the Thyroid Center in the Endocrinology & Metabolism Institute at Cleveland Clinic. “We looked at how the GEC helped us, until we switched to GSC,” he says. “Then we compared our outcomes of GSC compared to GEC. We actually found more benign rates with the GSC, which seems to detect more benign signatures, if you will, compared to the GEC. We’re relying on machine learning technology. Based on our findings and those of others, we believe that this is a useful technology. We are planning to monitor patients who did not have surgery to make sure those nodules are still benign long term.”
Investigators at Brigham and Women’s Hospital in Boston conducted a similar study, evaluating results for 583 thyroid nodules tested with either the GSC (n=97) or GEC (n=486) between 2011 and 2018. They found that the GSC identified 64.9% of nodules as benign, compared to 47.9% with the GEC, an increase of 35%.
“We’ve used the GEC for many years and found it to be very useful in reducing a lot of unnecessary surgery,” says Erik Alexander, MD, chief in the Thyroid Section of the Division of Endocrinology, Diabetes and Hypertension at Brigham and Women’s Hospital. “And now the GSC has been out for a bit of time, and we wanted to understand its performance. When we analyzed a large series of patients, we found, first, that the proportion of patients that had a benign result was higher with the GSC.”
Alexander goes on to say that this updated test improved the ability to help analyze and deal with the unique population of hurthle cell neoplasms. “In the past, there has been some difficulty in adequately identifying if those were benign or malignant,” he says. “So this updated version does appear to have significantly improved the ability to detect those cancers.”
Better Diabetes Outcomes
In April 2017, 256 Medtronic MiniMed Connect users were invited to take part in a 90-day Sugar.IQ app pilot program, during which researchers looked at the percentage of time in target range (TIR, 70-180 mg/dL), <70mg/dL, >180mg/dL, and excursions (periods >20 minutes and <70mg/dL or >180mg/dL) 30 days before Sugar.IQ onboarding, and compared to those results to results from 90 days later, in August 2017.
The researchers collected 10,761 unique Sugar.IQ usage sessions out of 11,356 sensor-wear days and identified behaviors associated with excursions. The Sugar.IQ app was used about twice a day, and the researchers found that compared to baseline, participants were in range about 36 minutes more per day and hypoglycemia events were reduced by one per month. The researchers found that users were able to make better decisions about bolusing and carb consumption after gaining insights from the app.
“There is a role for artificial intelligence in managing diabetes and helping to make some of the decisions that an individual has to make a little bit easier. Using artificial intelligence, we’re able to lead to more time in the target range for individual on multiple daily injections, which is the ultimate goal of diabetes care. So fewer episodes of hypoglycemia and fewer time in hyperglycemia.” – Lisa Latts, MD, deputy chief health officer, IBM Watson Health
The researchers, led by Yuxiang Zhong, the data science manager for Medtronic Diabetes, concluded that “[t]imely and personalized insights, such as those provided during the Sugar.IQ pilot, may advance patient understanding of glucose trends, aid in behavioral change that improves therapy adherence, and lead to better outcomes.” The results were presented at the American Diabetes Association’s annual meeting this past June in Orlando.
“[Thirty-six minutes a day] adds up to nine extra days per year where you are totally in range,” Latts says. “The presumption is that by spending more time in range and less time in the highs and less time in the lows, you will develop better glucose control overall and then that again presumably will lead to lower complications, lower episodes of hypoglycemia requiring hospitalization or ER use, and will lead to better overall outcomes in care of an individual with diabetes.”
Improving and Individualizing Care
Patients would certainly prefer to avoid a surgery they don’t need or a hypoglycemic episode like fainting in traffic and applying artificial intelligence to already existing therapies and tests can work to show that an ounce of prevention is indeed worth more than the cost of a trip to the emergency room or time lost while recovering from a lobectomy.
“We don’t want to send patients to have surgery that could be avoided,” Nasr says. “Our surgeons are excellent surgeons. They’re high volume surgeons, every one of them. Still, if someone can avoid surgery from the anxiety to having surgery, to the anesthesia, to the cost, to the possible, although remote risk of vocal cord damage, or even hematoma. I think by using this, for me, for our group here, for the Society, I think this is going to help all of us avoid sending patients unnecessarily to have a lobectomy, just by using this [test].”
Now more than ever, patients should be able to make better medical and lifestyle decisions, especially with the help from a device or a test that is able to rapidly learn and relearn about an individual and tailor a plan that works best for them.
“There is a role for artificial intelligence in managing diabetes and helping to make some of the decisions that an individual has to make a little bit easier,” Latts says. “Using artificial intelligence, we’re able to lead to more time in the target range for individual on multiple daily injections, which is the ultimate goal of diabetes care. So fewer episodes of hypoglycemia and fewer time in hyperglycemia.
“I think the biggest thing is to understand how cutting-edge molecular diagnostics are improving care,” Alexander says. “And I think with continual evolution of these tasks, they’re improving. And we are also increasingly individualizing care to each specific patient.”
–Bagley is the senior editor of Endocrine News. He wrote the October cover story about the effects opioids can have on various endocrinopathies.