Artificial Intelligence, which has already had a major impact on industries such as transportation, retail, energy, and banking, is only just beginning to be applied in medicine. Its profound capabilities hold promise for enabling early detection of disease and metabolic abnormalities and hope for empowering doctors and patients.
One advantage of AI is its unique ability to integrate large volumes of data and identify patterns that may be subtle or difficult for humans to recognize. These subtle patterns have a huge potential to alert clinicians to important physiologic changes that need to be addressed.
That is why we turned to the power of AI to invent a new way of detecting fluctuations of blood potassium levels that patients could easily perform at home without drawing blood. We are now preparing to submit the technology to the U.S. Food and Drug Administration for approval.
This initiative to create a “bloodless blood test” gave us insights into the process of creating AI-driven solutions that directly address an unmet patient or clinician need — insights that hopefully others interested in harnessing the power of AI in health care will find useful.
Sponsored by Optum
How technology is changing the design and delivery of care.
Potassium is essential for cellular electrical homeostasis, and the body keeps its levels within a narrow range. Fluctuations can lead to life-threatening arrhythmias and sudden death and can be caused by the medications we use to treat the very patients most at risk for these changes: those with heart or kidney diseases. Due to the high risk of harm to the patient when potassium levels are too high or too low, physicians are reluctant to adjust medications that affect potassium levels without first conducting a blood test. As the prevalence of heart and kidney diseases and their risk factors, high blood pressure and diabetes, grow so will the population of patients at risk of abnormally high and low concentrations of blood potassium (hyper- and hypokalemia).
Traditionally, blood potassium tests have required blood and were only available in doctors’ offices or hospitals — significant limitations for early detection and preventive measures. Our goal was to create a test that people could do at home, without drawing blood. Detecting abnormal blood potassium levels earlier would allow medicines to be more effectively titrated, problems to be addressed before they occurred, hospitalizations to be prevented, and costs to be reduced.
To support our approach, we needed to be able to flag even more subtle changes. AI and machine learning offered us the ability to leverage a massive amount of clinical data compiled at Mayo Clinic over time, which could then be used to generate an automated computer algorithm for the new testing method.
The method we developed utilizes smartphone technology in combination with algorithmic analysis of ECG recordings. Together with our partner AliveCor, we developed a system to attach electrodes to a smartphone and acquire a quality ECG recording. We then developed an algorithm that detects subtle changes in the morphology of the ECG waveform to determine serum potassium levels in near real time. We validated our approach in a series of clinical studies and found that our potassium scores closely correlated with serial blood tests taken from the same patients.
Here are the takeaways from our approach to developing the test:
Assemble a multidisciplinary team. The initial factor that would be key to achieving success was getting the right people working side by side toward a single goal of developing a bloodless blood test that patients could easily use themselves at home. This type of innovation requires a multidisciplinary team of physicians, hardware and software engineers, clinical study coordinators, and experts in commercializing medical technologies. What started as a partnership between experts in cardiac electrophysiology, nephrology, and hypertension quickly evolved to include experts in IT, data analytics, and machine learning as the need for the refinement of the algorithm arose. We then engaged our in-house experts from Mayo Clinic Ventures to help identify a partner, AliveCor, that would help us develop a product that could be brought patients and caregivers everywhere.
Start with known correlations. Since changes in blood potassium levels affect a patient’s electrocardiogram (ECG), we hypothesized that the signal processed by the ECG could serve as a non-invasive, easy-to-apply test if we could develop a way to translate the ECG signal into a measure of a full range of blood potassium levels. The ECG indicators of very abnormal potassium levels, occurring when a patient is close to cardiac arrest, were well-known. But features correlating with potassium levels closer to or within the normal range were not.
Accordingly, we began by examining data from a patient population with highly abnormal potassium levels: people with end-stage renal disease undergoing dialysis. During these dialysis sessions, dramatic reductions in potassium can occur over a few hours. This “supervised” approach involved having our human experts assess changes in ECG features related to potassium changes during dialysis. This led to the development of ECG-based potassium estimates and then associated computer algorithms.
Identify outside partners to go from bench to bedside. In order to transform our test into one that could be widely available, we realized we had to seek an outside partner. With the help of Mayo Clinic Ventures, we chose AliveCor, a company based in Mountain View, California, because it is an AI-driven medtech company recognized for being the first to develop a smartphone-connected ECG that leveraged deep neural networks. These networks, which are designed to function the same way as a human brain in terms of processing and sorting information, make machine learning possible. They make decisions or predictions with a degree of certainty according to the data they receive. We believed AliveCor’s platform and AI-driven networks would permit the potassium test to be performed easily at home and help scale the technology — make it widely available. (AliveCor has licensed the technology.)
The neural networks developed with AliveCor could look at more correlations than humans could to detect subtle pattern changes. The resulting algorithm outperformed the supervised approach — but it wasn’t perfect.
Pay attention to AI’s limitations. Machine learning is not immune to traditional statistical problems such as finding correlations solely from a random occurrence; so large data sets are important with machine learning to minimize these problems. And unlike traditional statistical analysis, we do not know why algorithms used in machine learning reach their conclusions. Consequently, the process requires rigorous human assessment and validation.
For this project, we were able to obtain data containing concurrent ECG and potassium measurements from more than 2 million Mayo Clinic patients. With such a large data set, the assessment and validation by our clinical and IT experts was critical. For example, since potassium declines throughout dialysis, we determined that the algorithm might have used dialysis duration rather than ECG patterns to determine blood potassium. We improved the algorithm by eliminating time from the AI data.
Address usability. The traditional ECG is typically performed in a medical environment. It involves placing 12 skin electrode leads in precise locations on the chest and limbs, needs specialized equipment, and requires training. In order to make the algorithm useful outside of clinical settings in hospitals, we had to simplify the interface that captures the ECG. We will be utilizing AliveCor’s FDA-approved physical device that patients could use at home.
AliveCor has an ECG sensor that attaches to the back of a smartphone. The signal is obtained by holding the sensor with both hands. AliveCor also has an ECG sensor incorporated into an Apple Watch band. The signal is obtained by placing a thumb on the sensor. The FDA has approved both devices for capturing an ECG and detecting an abnormal heart rhythm called atrial fibrillation. Our bloodless blood test technology can now leverage AliveCor’s FDA-approved device to capture ECG data and use these data to predict potassium levels. (Again, the potassium application requires FDA approval, which we will be seeking.)
Rely on commercialization experts. Mayo Clinic Ventures, our in-house team with expertise in intellectual property, early stage investments, and strategic business development, was involved in commercializing potassium technology. It assisted in the development of the intellectual property and, as we said above, helped us identify an appropriate partner (AliveCor) with complimentary expertise and a similar vision: to harness the power of artificial intelligence to advance the practice of medicine for doctors and patients. AliveCor has provided the much-needed expertise in engineering, product development and support, and accessing the larger health care market
Machine learning, which so far has had only limited applications in medicine, can be a powerful tool in the development of reliable solutions for unmet patient and clinical needs. However, fulfilling the potential of AI in medicine — and specifically machine-learning applications — also requires the human perspective. As the development of our test for blood potassium levels shows, it requires experts in medicine and technology working together to contextualize conclusions drawn through artificial intelligence and adjust systems and tools for optimal accuracy. AI and machine learning guided us, but people powered our success.
Disclosure: Mayo Clinic and the authors have financial interests in AliveCor. Mayo Clinic uses any revenue it receives to support its not-for-profit mission in patient care, education and research.
Source: Harward Business