Monday, February 22, 2021

Diabetes Meets Machine Learning, Part 1

John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform, wrote this article.

Of all the disorders that have responded well to artificial intelligence and machine learning, diabetes mellitus probably tops the list. The evidence supporting a role for Machine Learning (ML)-enhanced algorithms in managing the disease is persuasive and applies to several components of patient care, including screening, diagnosis, treatment, and prognosis.

Let's start with screening: As most clinicians know, there's a difference between screening for disease and diagnosing it. The former casts a much wider net, which means it will include more false positives. Once this larger cohort has been identified, more precise diagnostic testing can be performed to pinpoint patients who have the disorder.  As we explain in our next book, The Digital Reconstruction of Healthcare, screening tools for diabetes have been in existence for many years, including tools to help clinicians and patients identify the presence of prediabetes. Still, new technology has taken the screening process to a new level.

The American Diabetes Association (ADA) defines prediabetes as fasting plasma glucose between 100 and 126 mg/dl, a 2-hour oral glucose tolerance test reading between 140 and 200 mg/dl, or hemoglobin A1c of 5.7% to 6.5%.1 The Association has developed a risk assessment tool to help clinicians and patients; it asks for a patient's age, gender, family history of diabetes, the presence of hypertension, whether the patient is physically active, their ethnicity, and height and weight. Although this kind of scoring system will detect many persons with prediabetes, it has its limitations and ignores many changes in a patient's medical history over time. An artificial intelligence (AI)-enabled assessment tool that takes advantage of gradient boosting, on the other hand, is capable of factoring in many more risk factors gleaned from a patient’s electronic health record. This type of assessment system can accurately predict which patients will progress to full-blown diabetes, with an area under the curve (AUC) of 0.865.2 The gradient boosting approach has proven more accurate than logistic regression and evaluation of standard clinical cutoffs of blood glucose ≥ 110 mg/dl and HbA1c ≥ 6.0%. An analysis by Avivit Cahn from the Hebrew University of Jerusalem and her colleagues found the said ML algorithm incorporated 69 variables to achieve its superior AUC.

There is also reason to believe that AI-enhanced algorithms can help identify individuals before they become prediabetic. The Centers for Disease Control and Prevention (CDC) has a screening test to help predict prediabetes that asks one’s age, family history of diabetes, the existence of hypertension, race or ethnicity, whether one is physically active, gender, height and weight.

But logistic regression, artificial neural networks, random forest analysis, and gradient boosting have uncovered numerous other risk factors to predict prediabetes. Among the predictors that may help refine the risk stratification process: food security, citizenship, monocyte count, uric acid level, ALT, RBC count, and serum calcium and potassium.3 Whereas the CDC screening tool returned an area under the receiver operating characteristic (AUROC) of 64.4% in a large cohort from the National Health and Nutrition Survey, several AI-based algorithms generated AUROCs at or above 70%.

Several research projects suggest machine learning can improve the treatment of diabetes. In our book on mobile medicine, we discussed DreaMed,*  an Israeli vendor that developed DreaMed Diabetes to provide clinical decision assistance to clinicians managing Type 1 diabetes in patients who are using an insulin pump and continuous glucose monitoring (CGM). Its Advisor Pro algorithms collect several data types, including the patient’s daily blood glucose readings, meal carbohydrate data, insulin dosing records, and physical activity level. The software then uses event-driven machine learning and adaptive technology to create a more individualized regimen, adjusting its insulin dosing recommendations and offering behavior-modification recommendations accordingly. DreaMed Diabetes relies on a clinical AI program called MD-Logic. It uses fuzzy logic and adaptive learning algorithms to approximate how diabetes specialists reason as they adjust an individual's insulin regimen. The company claims that the software provides “faster analysis and deeper insights” than would normally be available with routine endocrinologist care. There is research support for this automated approach to Type 1 diabetes management.4

This coordinated interplay between an insulin pump and CGM sometimes referred to as closed-loop control or an artificial pancreas, has also been a focus of Mayo Clinic researchers. Yogish Kudva, MBBS, along with colleagues Sandeep Gupta and Ayan Bannerjee from Arizona State University, has been analyzing continuous glucose monitoring data from people with type 1 diabetes on the Medtronic MiniMed 670G closed-loop system* to develop software that would help each patient and clinic provider improve closed-loop control. They have also developed a system for measuring heart rate, respiratory rate, and self-report symptoms of viral infections in individuals with diabetes. The aim is to investigate the role of glycemic control on the progression of signs and symptoms of infectious disease in both at-home and hospital settings.

Similarly, Dr. Kudva participated in a recent investigation on the role of a closed-loop system to improve glycemic control in adolescents and young adults, part of the International Diabetes Closed-Loop Trial.5

Unfortunately, until recently, CGM has been out of reach for most Type 2 patients. But with the introduction of new, state-of-the-art technology, it is now available to them as well. We’ll explore these innovative solutions in part 2 of our series on machine learning-enhanced diabetes.

*Products and services mentioned are not endorsements.  Mayo Clinic has no relationship with any of these vendors.



1. American Diabetes Association. Classification and diagnosis of diabetes: standards of medical care in diabetes 2019. Diabetes Care. 2019;42(Suppl. 1):S13–S28. doi .org/10.2337/dc19-S002

2.  Cahn A, Shoshan A, Sagiv T el al. Prediction of progression from pre-diabetes to diabetes: development and validation of a machine learning model. Diabetes Metab Res Rev. 2019; e3252.

3. De Silva K, Jonsson D, Demmer RT. A combined strategy of feature selection and machine learning to identify predictors of prediabetes. JAMIA. 2020;27: 396–406.

4. Nimri R, Muller I, Atlas E, et al. MD-logic overnight control for 6 weeks of home use in patients with type 1 diabetes: randomized crossover trial. Diabetes Care. 2014;37:3025_32.

5. Isganaitis E, Raghinaru D, Amber-Osborn L et al. Closed-Loop Insulin Therapy Improves Glycemic Control in Adolescents and Young Adults: Outcomes from the International Diabetes Closed-Loop Trial. Diabetes Technol. Ther. 2021l Jan 21. doi: 10.1089/dia.2020.0572.  Online ahead of print.

Friday, February 12, 2021

High-Quality Hospital Care — Minus the Hospital

John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform, wrote this article.

In a time of COVID, the current care models just can’t remain in place. We need innovative ways to address the high cost of acute care. One solution that is taking a front seat is telemedicine.  Telecare has accelerated from 3-4% of visits in January 2020 to 90% in April to a new normal of 20% in 2021. Cultural change has modified patient expectations for the care they can get at a distance, which can be more convenient, less likely to result in COVID exposure, and more patient-centric.

Research has demonstrated that hospital-at-home programs for patients with specific acute medical conditions can reduce complications and reduce the cost of care by 30% or more.[1] One of the most progressive programs to focus on this transition was spearheaded by Johns Hopkins Hospital in 1994. Bruce Leff, MD, and his colleagues have tested this program with 455 elderly patients from three Medicare-managed systems and a VA medical center.[2] They found that the home model met the quality of care standards comparable to those expected of in-hospital programs. Also, "On an intention-to-treat basis, patients treated in hospital-at-home had a shorter length of stay (3.2 vs. 4.9 days) (P =0.004), and there was some evidence that they also had fewer complications. The mean cost was lower for hospital-at-home care than for acute hospital care ($5081 vs. $7480) (P < 0.001).”

A meta-analysis of 61 randomized clinical trials that looked at hospital-at-home projects found that among 42 trials, which included almost 7,000 patients, this approach reduced mortality (odds ratio, 0.81). Similarly, they decreased readmission rates by 25% (odds ratio, 0.75) and lowered costs. The same analysis revealed that treating every 50 patients in such a program saved one life.[3] Realizing the potential advantages of hospital-at-home programs, several large U.S. provider organizations have entered this space in the last few years, including Mayo Clinic, Partners Healthcare/Brigham and Women’s Hospital, and Mount Sinai Health System in New York. Across the globe, there are also significant programs in Australia, South Wales, and Spain.

Advanced Care at Home, a partnership between Mayo Clinic and Medically Home., tracks heart rate, blood pressure, pulse oximetry, temperature, and respiratory rate in its patient population, using Bluetooth-enabled devices wirelessly connected to the Mayo/Medically Home system. It also uses tablets, a back-up battery system and a Wi-Fi phone. There are, however, critical differences between many home-care programs and the Mayo Clinic system. Many hospital-at-home programs are targeted and designed for low-acuity hospital patients. They use physician house calls as the clinical delivery model. They have a short patient engagement period (2-4 days).

The Medically Home affiliated setup is designed to handle an extended length of stay that includes acute, post-acute and preventative care. It uses a scalable “decentralized” model for high-acuity care and can manage a broad set of diverse use cases and support an extensive patient census. The program uses screening, training, contracting, quality management, and technology and converts "post-acute" community-based supply providers into acute-level providers, bringing goods and services to high-acuity patients at home while focusing heavily on the role of paramedics as the centerpiece of its ability to provide rapid-response capabilities. In practical terms, that means paramedics and other providers go into the home while being virtually connected with a centralized medical command center staffed by physicians who guide the care for decentralized patients and the decentralized providers that care for them.

Advanced Care at Home has made measurable progress within a relatively short period, going from a speculative pilot project about a year ago to a business plan that will likely prove profitable in 2022. During a recent Zoom call with Ajani (AJ) Dunn, Administrator for the program, he emphasized, “It’s a story about volume.  As we looked at the model, we asked ourselves: Will it be effective? And we found that by scaling it up to the point where we can take out the fixed cost of a traditional hospital stay and replace that with the small variable costs of each service we deliver in the home, we can have a sustainable program.” Dunn explained that by working through a centralized command center staffed by physicians and getting buy-in from third-party payors, this approach is slowly turning the corner financially. Since most insurers do not have a hospital at home plan built into their policy, that has required the Mayo team to negotiate with payors one by one, explaining the cost-effectiveness of the new model.

The next step in the program’s growth will require finding ways to reinvent the existing system. “The off-the-shelf system we use to administer Advanced Care at Home is well-calibrated for traditional medical admissions, including COPD, CHF, pneumonia and the like. That approach involves drawing labs, titrating medications, etc. But we need other ways to use the system. Our plan now is to embed teams into other disease states, including oncology, cardiovascular disease, and transplantation, to understand the natural progression of each disease and the necessary interventions. Then we can create clinical protocols and administrative logistics that replicate in-hospital care in the home, but customized for each disease state.”

Undoubtedly, concerns about high costs and unexpected complications will continue to dissuade patients from seeking in-hospital care for many years. But as this new model expands, it's likely more patients will see the advantages of seeking high-quality hospital care — minus the hospital.



1. Klein S. “Hospital at home” programs improve outcomes, lower costs but face resistance from providers and payers. The Commonwealth Fund. Accessed January 15, 2020.

2. Leff B, Burton L, Mader SL, et al. Hospital at home: feasibility and outcomes of a program to provide hospital-level care at home for acutely ill older patients. Ann Intern Med. 2005;143:798–808.

3. Caplan GA, Sulaiman NS, Mangin DA, et al. A meta-analysis of “hospital in the home.” Med J Aust. 2012;197:512–519.

Wednesday, February 10, 2021

Let Your Freak Flag Fly

 John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform, wrote this article.

If Mary Putnam Jacobi were alive today, she would probably embrace artificial intelligence (AI), machine learning, and data analytics. Dr. Jacobi might best be described as the mother of modern scientific medicine  or at the very least one of its founding parents. In 1868, she was the first woman to enroll in the University of Paris School of Medicine. After graduating in 1871, this unconventional thinker arrived in the U.S., where she advocated for the inclusion of laboratory science, experimentation and statistics as the foundation cornerstone of modern medical practice. Equally important, Jacobi “became a powerful advocate for the equal contribution of women to medicine.” Pushing clinicians to buy into the notion that experimentation and statistics were needed for good quality patient care may seem unimpressive today. Still, it was almost heresy in an age when the received wisdom from one’s medical school professor was all that was necessary to “demonstrate” that a treatment protocol was effective. Given her “color-outside-the-lines” approach, it’s not hard to imagine her becoming passionate about the role of AI in medicine — not as a panacea but as an indispensable adjunct to human reasoning and clinical trials.

The evidence shows that machine learning, advanced data analytics, and subgroup analysis all have the power to reinvent the way patient care is delivered and the potential to personalize that care. But as we said in an earlier blog, while there are several well-documented AI-based algorithms now available, there are also several questionable vendors that have rushed to market with little evidence to support their claims. This dilemma begs the question: How do we ensure that well-supported digital tools get the attention they deserve? Several new guidelines have been published to encourage more robust AI-related research, but clinicians also need to develop machine learning literacy. Put another way, we can all benefit from gaining a better grasp of what’s under the hood, explained in plain, jargon-free English. That was one of the goals of our last book, Reinventing Clinical Decision Support.

Our research demonstrates that clinical decision support certainly needs to be reinvented. If you look back over the last several decades, you’ll find that hundreds of clinical research studies have been misinterpreted. A review of 71 randomized clinical trials in 1987, for instance, concluded that numerous treatments were useless; a closer analysis, however,  found that all 71 had generated false-negative results because the populations they studied were too small — a type 2 statistical error. [1] Fast forward to 1994; a JAMA analysis revealed that 383 RCTs, many published in the world’s top academic journals, likewise jumped to the conclusion that several treatments were ineffective, once again because the trials had enrolled too few patients. [2]

More recently, there’s reason to believe that the bedrock upon which many day-to-day clinical decisions rest is somewhat shaky. It’s a foundation that can be made stronger with ML-based tools like convolutional neural networks, random forest modeling, gradient boosting, and clustering. A case in point: The Look AHEAD study that published in 2013.[3] This RCT assigned over 5,000 overweight patients with type 2 diabetes to either an intensive lifestyle modification program or a control group. The investigators' goal was to determine if the lifestyle program would reduce the incidence of cardiovascular events. The study was terminated early because there were no significant cardiovascular differences between the intervention and control groups.

An ML-fueled analysis that used random forest modeling to re-examine the Look AHEAD data turned these results upside down. During random forest analysis, a series of decision trees are created. Initially, the technique randomly splits all the available data — in this case, the stored characteristics of about 5,000 patients in the Look AHEAD study — into two halves. The first half serves as a training data set to generate hypotheses and construct the decision trees. The second half of the data serves as the testing data set.

Using this technique, Baum et al. constructed a forest that contained 1,000 decision trees. They looked at 84 risk factors that may have been influencing patients’ response or lack of response to the intensive lifestyle modification program, including numerous characteristics that researchers rarely, if ever, consider when doing a subgroup analysis. The random forest modeling also allowed the investigators to examine how these variables interact in multiple combinations to impact clinical outcomes. In the final analysis, Baum et al.[4] discovered that intensive lifestyle modification did prevent cardiovascular events for two subgroups, patients with HbA1c 6.8% or higher (poorly managed diabetes) and patients with well-controlled diabetes (Hba1c < 6.8%) and good self-reported health. That finding applied to 85% of the entire patient population studied. On the other hand, the remaining 15% who had controlled diabetes but poor self-reported general health responded negatively to the lifestyle modification regimen. The negative and positive responders canceled each other out in the original Look AHEAD statistical analysis, falsely concluding that lifestyle modification was useless.

AI-enhanced data analyses like this only serve to reinforce Maru Putnam Jacobi’s contention that coloring outside the lines will propel health care in new directions. Her unconventional mindset, and ours, are best summed up in the words of Crosby Still and Nash: “Let your freak flag fly!”


1. Freiman JA et al. The importance of beta, the type II error and sample size in the design and interpretation of the randomized control trial. Survey of 71 "negative" trials. N Engl J Med. 1978; 299:690-694.

2Moher D et al. Statistical power, sample size, and their reporting in randomized controlled trials. JAMA. 1994; 272:122.

3. Look AHEAD Research Group. Cardiovascular effects of intensive lifestyle intervention in type 2 diabetes. N Engl J Med. 2013. 369:145-154.

4. Baum, A et al. Targeting Weight Loss Interventions to Reduce Cardiovascular Complications of Type 2 Diabetes: A Machine Learning-Based Post-Hoc Analysis of Heterogeneous Treatment Effects in the Look AHEAD Trial. Lancet Diabetes Endocrinology, 2017;  5: 808–815.

Thursday, February 4, 2021

The Therapeutic Potential of Voice Technology

John Halamka, M.D., president, Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform, wrote this article.

The human voice is capable of extraordinary feats of genius and everyday acts of kindness.  It can recite Shakespearean sonnets, teach our children moral values, stir audiences with a dramatic performance, and much more. But few of us ever imagined it capable of assisting in the diagnosis of disease. That’s about the change, as evidenced by several innovative projects in voice technology.

With the right digital tools, it is now possible to analyze a patient’s speech patterns to detect clues to underlying pathological issues. Elad Maor, MD, Ph.D., with the Mayo Clinic Department of Cardiovascular Medicine, and his colleagues have looked at voice samples from about 100 patients who underwent coronary angiograms, asking them to read text excerpts and respond to questions about positive and negative emotional experiences. Their recorded responses found subtle differences in vocal pitch and intensity between patients who were ultimately diagnosed with heart disease and normal controls. [1] Dr. Maor and his associates concluded: “One possible explanation for our interesting finding is the documented association between mental stress, the adrenergic system, and voice. . . . Emotional stress conditions change the human voice, including an increase in fundamental frequency. . . . [O]ne possible hypothesis to interpret our findings is that the association between voice and atherosclerosis is mediated by hypersensitivity of the adrenergic system to stress. The association between stress, the adrenergic system, and atherosclerosis is well established on the basis of robust data.”

There is also reason to believe that voice technology may help detect pulmonary hypertension. Jaskanwal Deep Singh Sara, M.B., Ch.B. (also with the Mayo Clinic), in collaboration with scientists from Vocalis Health (an Israeli vendor), has analyzed voice recordings among patients who had invasive cardiac hemodynamic testing, the standard approach to diagnosing pulmonary hypertension. The recordings were analyzed to measure pitch, loudness, jitter, and other metrics. Dr. Sara and his colleagues found a significant association between an invasively derived hemodynamic index used to measure pulmonary hypertension and the vocal biomarkers. Patients with pulmonary arterial pressure at or above 35 mmHg had higher mean vocal biomarker readings than those with pressure readings at or below 35 mmHg. Given that invasive testing occurred during cardiac catheterization, the non-invasive collection of a patient's voice patterns holds promise as a safer alternative. If controlled clinical trials confirm the findings, it will likely reduce the cost and risk associated with pulmonary hypertension (PH) diagnosis. [2]

Investigators in Belgium also have had success using vocal characteristics as part of a suite of tools to help detect Parkinson's Disease in its early stages. [3]  Some studies suggest that 60-90% of patients have subtle changes in their voice and speech patterns when initially diagnosed. In situations in which the diagnosis is questionable, however, a neurologist might administer a loading dose of levodopa, one of the standard drugs used to treat the disease and watch for improvements in a patient’s speech and voice. After administering the drug, the Flemish researchers monitored the strength of a patient's facial or mouth muscles and evaluated vocal quality, frequency, breathiness, phonation time — how long a person can sustain a vocal sound on one deep breath — and several other parameters, folding them into a metric called a Voice Handicap Index. They found that these markers helped distinguish patients with idiopathic Parkinson's disease from healthy individuals.

Mayo Clinic is also exploring the value of voice technology, coupled with artificial intelligence, to address patients' needs with neurological disorders and related motor speech disorders. The project's goals include creating digital tools for voice-based disease detection in the office, over the phone, and in a patient's home; earlier, more accurate and more holistic diagnoses; and the provision of individualized, in-home markers of disease progression and treatment response. To accomplish those goals, the Neurology AI Program and other Mayo researchers and clinicians are developing a fully automated digital speech diagnostics platform that can take a patient's speech sample and provide probabilistic diagnoses. Those taking the lead in this initiative believe that the primary goal should not be to map directly onto a clinical diagnostic label but instead use AI to extract clinically meaningful information from a speech sample. 

Mayo Clinic-based speech pathologists Darley, Aronson and Brown [4] proposed a similar idea in 1969 in their seminal work on the dysarthrias, speech disorders caused by muscle weakness.  While this theoretical construct was previously based on conventional speech labels, advances in AI have made it feasible to build such feature space using latent patterns in the data, which can then be labeled and linked to dense, multivariate medical data. This Deeply Annotated Speech Latent space can then be used to characterize new speech samples. When connected to non-speech medical and demographic data, a broader neurologic or systemic diagnosis can be rendered. The model will likely have the most considerable impact if deployed on smartphone devices, smart speakers or other wearable technology.

Voice technology has yet to surpass Shakespeare’s genius, replace a gifted actor's heart-wrenching performance,  or serve as a moral compass for the next generation. Still, the evidence suggests it is ushering in a new generation of intelligent digital tools that will likely transform patient care. 


1. Maor E, Sara JD, Orbelo DM, et al. Voice signal characteristics are independently associated with coronary artery disease. Mayo Clinic Proceedings. 2018;93:840–847.

2. Sara JDS, Maor E, Borlaug B, et al. Non-invasive vocal biomarker is associated with pulmonary hypertension. PLOS ONE. 15(4):e0231441.

3. Lechien JR, Delsaut B, Abderrakib A et al. Orofacial Strength and Voice Quality as Outcome of Levodopa Challenge Test in Parkinson Disease. Laryngoscope, 130:E896–E903, 2020.

4. Frederic L Darley, Arnold E Aronson, and Joe R Brown. Clusters of deviant speech dimensions in the dysarthrias. Journal of speech and hearing research, 12(3):462–496, 1969.