Reimagining the FDA’s Role in Digital Medicine

In addition to evaluating the safety of software as a medical device (SaMD), the agency needs to devote more resources to evaluating its efficacy and quality.

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

The FDA’s approach to software as a medical device (SaMD) has been evolving. Consider a few examples.

In 2018, IDx-DR, a software system used to improve screening for retinopathy, a common complication of diabetes that affects the eye, became the first AI-based medical device to receive US Food and Drug Administration clearance to “detect greater than a mild level of … diabetic retinopathy in adults who have diabetes.” To arrive at that decision, the agency not only reviewed data to establish its safety, it also took into account prospective studies, an essential form of evidence that clinicians look for when trying to decide if a device or product is worth using. The software was the first medical device approved by the FDA that does not require the services of a specialist to interpret the results, making it a useful tool for health care providers who may not normally be involved in eye care. The FDA clearance emphasized the fact that IDx-DR is a screening tool not a diagnostic tool, stating that patients with positive results should be referred to an eye care professional. The algorithm built into the IDx-DR system is intended to be used with the Topcon NW400 retinal camera and a cloud server that contains the software.

Similarly, FDA looked at a randomized prospective trial before approval of a machine learning-based algorithm that can help endoscopists improve their ability to detected smaller, easily missed colonic polyps. Its recent clearance of GI Genius by Medtronic was based on a clinical trial published in Gastroenterology, in which investigators in Italy evaluated data from 685 patients, comparing a group that underwent the procedure with the help of the computer-aided detection (CADe) system to a group who acted as controls. Repici et al found that the adenoma detection rate was significantly higher in the CADe group, as was the detection rate for polyps 5 mm or smaller, which led to the conclusion: “Including CADe in colonoscopy examinations increases detection of adenomas without affecting safety.”

Their findings raise several questions: is it reasonable to assume that a study of 600+ Italians would apply to a U.S. population, which has different demographic characteristics? More importantly, were the 685 patients representative of the general public, including adequate numbers of persons of color and those in lower socioeconomic groups? While the Gastroenterology study did report enough female patients, there is no mention of these other marginalized groups.  

An independent 2021 analysis of FDA approvals has likewise raised several concerns about the effectiveness and equity of several recently approved AI algorithms. Eric Wu from Stanford University and his colleagues examined the FDA’s clearance of 130 devices and found the vast majority were approved based on retrospective studies (126 of 130). And when they separated all 130 devices into low- and high-risk subgroups using FDA guidelines, they found none of the 54 high-risk devices had been evaluated by prospective trials. Other shortcomings documented in Wu’s analysis included the following:

• Of the 130 approved products, 93 did not report multi-site evaluation.
• Fifty-nine of the approved AI devices included no mention of the sample size of the test population. 
• Only 17 of the approved devices discussed a demographic subgroup. 

We would certainly like to see the FDA take a more thorough approach to AI-based algorithm clearance, but in lieu of that, several leading academic medical centers, including Mayo Clinic, are contemplating a more holistic and comprehensive approach to algorithmic evaluation. It would include establishing a standard labeling schema to document the characteristics, behavior, efficacy, and equity of AI systems, to reveal the properties of systems necessary for stakeholders to assess them and build the trust necessary for safe adoption. The schema will also support assessment of the portability of systems to disparate datasets. The labeling schema will serve as an organizational framework that specifies the elements of the label. Label content will be specified in sections that will likely include:

• model details such as name, developer, date of release, and version,
• the intended use of the system,
• performance measures,
• accuracy metrics, and
• training data and evaluation data characteristics

While it makes no sense to sacrifice the good in pursuit of the perfect, the current regulatory framework for evaluating SaMD is far from perfect. Combining a more robust FDA approval process with the expertise of the world’s leading medical centers will offer our patients the best of both worlds.

Secure Computing Enclaves Move Digital Medicine Forward

By providing a safe, secure environment, novel approaches enable health care innovators to share data without opening the door to snoopers and thieves.

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

We know that bringing together AI algorithms and data in ways that preserve privacy and intellectual property is one of the keys to delivering the next generation of clinical decision support. But meeting that challenge requires health care innovators to look to other innovators who themselves have created unique cybersecurity solutions. Among these “Think outside the box” solutions are products and services from vendors like TripleBlind, Verily, Beekeeper.AI/Microsoft, Terra, and Nvidia.

The concept of secure computing enclaves has been around for many years. Apple created its secure enclave, a subsystem built into its systems on a chip (SoC), which in turn is “an integrated circuit that incorporates multiple components into a single chip,” including an application processor, secure enclave, and other coprocessors. Apple explains that “The Secure Enclave is isolated from the main processor to provide an extra layer of security and is designed to keep sensitive user data secure even when the Application Processor kernel becomes compromised. It follows the same design principles as the SoC does—a boot ROM to establish a hardware root of trust, an AES [advanced encryption standard] engine for efficient and secure cryptographic operations, and protected memory. Although the Secure Enclave doesn’t include storage, it has a mechanism to store information securely on attached storage separate from the NAND flash storage that’s used by the Application Processor and operating system.” The secure enclave is embedded into the latest versions of its iPhone, iPad, Mac computers, Apple TV, Apple Watch, and Home Pod.

While this security measure provides users when an extra layer of protection, because it’s a hardware-based solution, its uses are limited. With that in mind, several vendors have created software-based enclaves that are more readily adapted by customers. At Mayo Clinic Platform, we are deploying TripleBlind’s services to facilitate sharing data with our many external partners. It allows Mayo Clinic to test its algorithms using another organization’s data without either party losing control of its assets. Similarly, we can test an algorithm from one of our academic or commercial partners with Mayo Clinic data, or test an outside organization’s data with another outside organization’s data.

How is this “magic” performed? Of course, it’s always about the math. TripleBlind allows the use of distributed data that is accessed but never moved or revealed; it always remains one-way encrypted with no decryption possible. TripleBlind’s novel cryptographic approaches can operate on any type of data (structured or unstructured images, text, voice, video), and perform any operation, including training of and inferring from AI and ML algorithms. An organization’s data remains fully encrypted throughout the transaction, which means that a third party never sees the raw data because it is stored behind the data owner organization’s firewall. In fact, there is no decryption key available, ever. When two health care organizations partner to share data, for instance, TripleBlind software de-identifies their data via one-way encryption; then, both partners access each other’s one-way encrypted data through an Application Programming Interface (API). That means each partner can use the other’s data for training an algorithm, for example, which in turn allows them to generate a more generalizable, less biased algorithm. During a recent conversation with Riddhiman Das, CEO for TripleBlind, he explained: “To build robust algorithms, you want to be able to access diverse training data so that your model is accurate and can generalize to many types of data. Historically, health care organizations have had to send their data to one another to accomplish this goal, which creates unacceptable risks. TripleBlind performs one-way encryption from both interacting organizations, and because there is no decryption possible, you cannot reconstruct the data. In addition, the data can only be used by an algorithm for the specific purpose spelled out in the business agreement.”

Developing innovative technological services is exciting work, with the potential to reshape the health care ecosystem worldwide. But along with the excitement is the challenge of keeping data safe and secure. Taking advantage of the many secure computing enclaves available on the market allows us to do just that.

Breast Cancer Screening: We Can Do Better

The three risk assessment tools now in use fall far short. Using the latest deep learning techniques, investigators are developing more personalized ways to locate women at high risk.

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

The promise of personalized medicine will eventually allow clinicians to offer individual patients more precise advice on prevention, early detection and treatment. Of course, the operative word is eventually. A closer examination of the screening tools available to detect breast cancer demonstrates that we still have a way to go before we can fulfill that promise. But with the help of better technology, we are getting closer to that realization.

Disease screening is about risk assessment. Researchers collect data on thousands of patients who develop breast cancer, for instance, and discover that the age range, family history and menstruation history of those who develop the disease differs significantly from those who remain free of it. That in turn allows policy makers to create a screening protocol that suggests women of a certain age who have experienced early menarche or late menopause are more likely to develop the malignancy. That risk assessment is consistent with the fact that more reproductive years means more exposure to the hormones that contribute to breast cancer. Similarly, there’s evidence to show that women with first degree relatives with the cancer and those with a history of ovarian cancer or HRT use are at greater risk.

Statistics like this are the basis for several breast cancer risk scoring systems, including the Gail score, the IBIS score, and BCSC tool.  The National Cancer Institute, which uses the Gail model, explains: “The Breast Cancer Risk Assessment Tool allows health professionals to estimate a woman's risk of developing invasive breast cancer over the next 5 years and up to age 90 (lifetime risk). The tool uses a woman’s personal medical and reproductive history and the history of breast cancer among her first-degree relatives (mother, sisters, daughters) to estimate absolute breast cancer risk—her chance or probability of developing invasive breast cancer in a defined age interval.” While the screening tool saves lives, it can also be misleading. If, for example, it finds that a woman has a 1% likelihood of developing breast cancer, what that really means is a large population of women with those specific risk factors has a one in 100 risk of developing the disease. There is no way of knowing what the threat is for any one patient in that group. Similar problems exist for the International Breast Cancer Intervention Study (IBIS) score, based on the Tyrer-Cuzick Model, and the Breast Cancer Surveillance Consortium (BCSC) Risk Calculator. These 3 assessment tools can give patients a false sense of security if they don’t dive into the details. BCSC, for instance, cannot be applied to women younger that 35 or older than 74, nor does it accurately measure risk for anyone who has previously had ductal carcinoma in situ (DCIS), or had breast augmentation. Similarly, the NCI tool doesn’t accurately estimate risk in women with BRCA1 or BRCA1 mutation, as well as certain other subgroups.

During a conversation with Tufia Haddad, M.D,, a Mayo Clinic medical oncologist with specialty interest in precision medicine in breast cancer and artificial intelligence, she discussed the research she and her colleagues are doing to improve the risk assessment process and identify more high-risk women. Dr. Haddad pointed out that there are numerous obstacles that prevent women from obtaining the best possible risk assessment. Too many women do not have a primary care practitioner who might use a risk tool. And those that do have a PCP are more likely to have an evaluation based on the Breast Cancer Risk Assessment tool (the Gail model). “We prefer the Tyrer-Cuzick model in part because it incorporates more personal information for each individual patient including a detailed family history, a woman’s breast density from her mammogram, as well as her history of atypia or other high risk benign breast disease,” says Dr. Haddad. Unfortunately, the Tyrer-Cuzick method requires many more data elements to assess breast cancer risk, which discourages busy clinicians from using it.

Another obstacle to using any of these risk assessment tools is the fact that they don’t readily fit into the average physician’s clinical workflow. Ideally these tools should seamlessly integrate into the EHR system. Even better would be the incorporation of AI-enhanced algorithms that automate the abstraction of the required data elements from the patient’s record into the assessment tool. For example, the algorithm would flag a family history of breast cancer, increased breast density as determined during a mammogram, as well as hormone replacement therapy and insert those risk factors into the Tyrer-Cuzick tool.

Even with this AI-enhanced approach, all of the available risk models fall short because they take a population-based approach, as we mentioned above. Dr. Haddad and her colleagues are looking to make the assessment process more individualized, as are others work in this specialty. That model could incorporate each patient’s previous mammography results, their genetics and benign breast biopsy findings, and much more. Adam Yala, and his colleagues at MIT recently developed a mammography-based deep learning model designed to take this more sophisticated approach. Called Mirai, it was trained on a large data set from Massachusetts General Hospital and from facilities in Sweden and Taiwan.  The new model generated significantly better results for breast cancer risk prediction than the TC model.

Breast cancer risk assessment continues to evolve. And with better utilization of existing assessment tools and the assistance of deep learning, we can look forward to better patient outcomes.

Can Social Determinants of Health Predict Your Patient’s Future?

The evidence is mixed but suggests that these overlooked variables have a profound impact on each patient’s journey. 

This article was written by Tim Suther, Nicole Hobbs, Jeff McGinn, Matt Turner with Change Healthcare, John Halamka, MD, MS, president of Mayo Clinic Platform, and Paul Cerrato, senior research analyst and communications specialist, Mayo Clinic Platform.

By one estimate, social determinants of health (SDoH) influence up to 80% of health outcomes. Although reports like this suggest that these social factors have a major impact, thought leaders continue to debate whether they can also enhance the accuracy in predictive models. Resolving that debate is far from simple because the answer depends on the type, source and quality of the data, and the design of the model under consideration.

In general, we derive SDoH from subjective and objective sources. Subjective data includes self-reported or clinician-collected data such as patient reported outcomes, Z codes from ICD-10-CM that report factors that influence health status and interactions with health service providers, and other unstructured EHR data. Objective data includes individual-level and community-level data from government, public and private (and consumer behavior) sources; it’s usually more structured and often derived from national-level datasets.

Unfortunately, the research on the value of SDoH in predictive models varies widely. Some studies report no appreciable differences when SDoH are injected into models, while others report significant enhancements to predictive power. Unsurprisingly, these varying study results depend in part on levels of reliance on traditional clinical models and, most importantly, on the types and sources of SDoH data employed in the studies.

For example, a group from Johns Hopkins Bloomberg School of Public Health demonstrated SDoH predictive models can fail in part due to predictive model design as well as to EHR-level data that is unstructured and collected inconsistently.  They also demonstrated that dependence on data from EHR-derived population health databases for SDoH can be problematic because the data tends to be used as a proxy for individual-level social factors.  The problem lies in the fact that these proxies are often based on assumptions, not evidence. Other research supports the above and showcases the challenges of using SDoH data from sources that traditionally struggle with the comprehensive collection and standardization of these data types.

On a more positive note, several studies and healthcare articles have reported success by relying on objectively collected and/or highly structured and consistent data. For example, one study that used EHR-derived SDoH data sources found that the addition of structured data on median income, unemployment rate,  and education from trustworthy non-EHR sources  enhanced their model’s health prediction granularity for some of the most vulnerable subgroups of patients. In another study, collaboration between Stanford, Harvard, and the Imperial College London found that adding structured SDoH data from the US Census, along with using machine learning techniques, improved risk prediction model accuracy for hospitalization, death, and costs. They also showed that their models based on SDoH alone, as well as those based on clinical comorbidities alone, could predict health outcomes and costs. Similarly, researchers at The Ohio State University College of Medicine added community-level and consumer behavior data not available in standard EHR data and found it enhanced the study of and impact on obesity prevention.  Juhn et. al. at Mayo Clinic tapped telephone survey data and appended housing and neighborhood characteristic data from local government sources to create a socioeconomic status index (HOUSES). They first showed that HOUSES correlated well with outcome measures and later showed that HOUSES could even serve as a predictive tool for graft failure in patients.

Patient Level SDoH + Clinical Data = Predictive Power

Incorporating social factors into the healthcare equation can fill gaps needed at the point of care, but it also generates better healthcare predictions, but only when these determinants are patient level and linked to robust clinical data. Change Healthcare, for example, has curated such an integrated national-level dataset, linking billions of historical de-identified distinct medical claims with patient-level social, physical and behavioral determinants of health. One of this dataset’s most important uses is to understand the relative weight of specific patient SDOH factors, in comparison to clinical factors alone, for various therapeutic conditions, including COVID-19. For example, across Change Healthcare’s research, economic stability is repeatedly ranked as the highest or among the highest predictors of the healthcare experience. Despite this realization, most end users, including providers and payers, lack such visibility (or rely on geographic averages that are unhelpful in making accurate predictive models).

Incorporating SDoH data into predictive models holds much promise. Given the relative newness of SDoH data in predictive analytics, along with a lack of data standardization and scale, it’s not surprising to find varying degrees of success in using it to improve predictive health models. But as researchers learn more about the best types and sources of SDoH data to use, along with developing better-suited models for these types of data, we’re likely to see significant advances in healthcare predictive models. By combining the right data with the right models, SDoH are a powerful asset in predictive models of health, outcomes, and potential health disparities.

If you're still with us . . .

Please consider supporting Dr. Steve Parodi, Reed Abelson and I by "voting up" on our panel at the upcoming South by Southwest conference in March of 2022. Our proposed panel, "Extending the Stethoscope Into the Home," will dive into a discussion about acute health care for patients in their home and the infrastructures needed to support it. If you are so inclined to vote, please do so here.

We Need to Open Up the AI Black Box

To convince physicians and nurses that deep learning algorithms are worth using in everyday practice, developers need to explain how they work in plain clinical English.

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

AI’s so-called black box refers to the fact that much of the underlying technology behind machine learning-enhanced algorithms is probability/statistics without a human readable explanation. Oftentimes that’s the case because the advanced math or the data science behind the algorithms is too complex for the average user to understand without additional training. Several stakeholders in digital health maintain, however, that this lack of understanding isn’t that important. They argue that as long as an algorithm generates actionable insights, most clinicians don’t really care about what’s “under the hood.” Is that reasoning sound?

Some thought leaders point to the fact that there are many advanced, computer-enhanced diagnostic and therapeutic tools currently in use that physicians don’t fully understand, but nonetheless accept.  The CHA2DSA-VASc score, for instance, is used to estimate the likelihood of a patient with non-valvular atrial fibrillation having a stroke. Few clinicians are familiar with the original research or detailed reasoning upon which the calculator is based, but they nonetheless use the tool.  Similarly, many physicians use the FRAX score to estimate a patient’s 10-year risk of developing a bone fracture, despite the fact that they have not investigated the underlying math. 

It’s important to point out, however, that the stroke risk tool and the FRAX tool both have major endorsements from organizations that physicians respect. The American Heart Association and the American College of Cardiology both recommend the CHA2DSA-VASc score while the National Osteoporosis Foundation supports the use of FRAX score. That gives physicians confidence in these tools even if they don’t grasp the underlying details. To date, there are no major professional associations recommending specific AI-enabled algorithms to supplement the diagnosis or treatment of disease. The American Diabetes Association did include a passing mention of an AI-based screening tool in its 2020 Standards of Medical Care in Diabetes, stating: “Artificial intelligence systems that detect more than mild diabetic retinopathy and diabetic macular edema authorized for use by the FDA represent an alternative to traditional screening approaches. However, the benefits and optimal utilization of this type of screening have yet to be fully determined.” That can hardly be considered a recommendation.

Given this scenario, most physicians have reason to be skeptical, and surveys bear out that skepticism.   A survey of 91 primary care physicians found that understandability of AI is one of the important attributes they want before trusting its recommendations during breast cancer screening. Similarly, a survey of senior specialists in UK found that understandability was one of their primary concerns about AI.  Among New Zealand physicians, 88% were more likely to trust an AI algorithm that produced an understandable explanation of its decisions.

Of course, it may not be possible to fully explain the advanced mathematics used to create machine learning based algorithms. But there are other ways to describe the logic behind these tools that would satisfy clinicians.  As we have mentioned in previous publications and oral presentations, there are tutorials available to simplify machine learning-related systems like neural networks, random forest modeling, clustering, and gradient boosting. Our most recent book contains an entire chapter on this digital toolbox. Similarly, JAMA  has created clinician friendly video tutorials designed to graphically illustrate how deep learning is used in medical image analysis and how such algorithms can be used to help detect lymph node metastases in breast cancer patients. 

These resources require clinicians to take the initiative and learn a few basic AI concepts, but developers and vendors also have an obligation to make their products more transparent.  One way to accomplish that goal is through saliency maps and generative adversarial networks. Using such techniques, it’s possible to highlight the specific pixel grouping that a neural network has identified as a trouble spot, which the clinician can then view on a radiograph, for example.  Alex DeGrave, with the University of Washington, and his colleagues, used this approach to help explain why an algorithm designed to detect COVID-19-related changes in chest X-rays made its recommendations. Amirata Ghrobani and associates from Stanford University have taken a similar approach to help clinicians comprehend the echocardiography recommendations coming from a deep learning system. The researchers trained a convolutional neural network (CNN) on over 2.6 million echocardiogram images from more than 2,800 patients and demonstrated it was capable of identifying enlarged left atria, left ventricular hypertrophy, and several other abnormalities. To open up the black box, Ghorbani et al presented readers with “biologically plausible regions of interest” in the echocardiograms they analyzed so they could see for themselves the reason for the interpretation that the model has arrived at. For instance, if the CNN said it had identified a structure such as a pacemaker lead, it highlighted the pixels it identifies as the lead. Similar clinician-friendly images are presented for a severely dilated left atrium and for left ventricular hypertrophy.

Deep learning systems are slowly ushering in a new way to manage diagnosis and treatment, but to bring skeptical clinicians on board, we need to pull the curtain back. In addition to providing evidence that these tools are equitable and clinically effectively, practitioners want reasonable explanations to demonstrate that they will do what they claim to do.

The Future Belongs to Digital Pathology

Advances in artificial intelligence are slowly transforming the specialty, much the way radiology is being transformed by similar advances in digital technology.

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

Any patient who faces a potential cancer diagnosis knows how important an accurate, timely pathology report is.  Similarly, surgeons often require fast pathology results when they are performing a delicate procedure to determine their course of action during an operation. New technological developments are poised to meet the needs of patients and clinicians alike.

AI can improve pathology practice in numerous ways. The right digital tools can automate several repetitive tasks, including the detection of small foci. It can also help improve the staging of many malignancies, make the workflow process more efficient, and help classify images, which in turn gives pathologists a “second set of eyes”. And those “eyes” do not grow tired at the end of a long day or feel stressed out from too much work.

Such capabilities have far-reaching implications. With the right scanning hardware and the proper viewer software, pathologists and technicians can easily view and store whole slide images (WSIs). That view is markedly different from what they see through a microscope, which only allows a narrow field of view. In addition, digitization allows pathologists to mark up WSIs with non-destructive annotations, use the slides as teaching tools, search a laboratory’s archives to make comparisons with images that depict similar cases, give colleagues and patients access to the images, and create predictive models. And if the facility has cloud storage capabilities, it allows clinicians, patients, and pathologists around the world to access the data.

A 2020 prospective trial conducted by University of Michigan and Columbia University investigators illustrates just how profound the impact of AI and ML can be when applied to pathology.  Todd Hollon and colleagues point out that interoperative diagnosis of cancer relies on a “contracting, unevenly distributed pathology workforce.”¹ The process can be quite inefficient, requiring a tissue specimen travel from the OR to a lab, followed by specimen processing, slide preparation by a technician, and a pathologist’s review. At University of Michigan, they are now using Stimulated Raman histology, an advanced optical imaging method, along with a convolutional neural network (CNN) to help interpret the images. The machine learning tools were trained to detect 13 histologic categories and includes an inference algorithm to help make a diagnosis of brain cancer. Hollon et al conducted a 2-arm, prospective multicenter, non-inferiority trial to compare the CNN results to those of human pathologists. The trial, which evaluated 278 specimens, demonstrated that the machine learning system was as accurate as pathologists’ interpretation (94.6% vs 93.9%). Equally important was the fact that it took under 15 seconds for surgeons to get their results with the AI system, compared to 20-30 minutes with conventional techniques. And that latter estimate does not represent the national average. In some community settings, slides have to be shipped by special courier to labs that are hours away.

Mayo Clinic is among several forward-thinking health systems that are in the process of implementing a variety of digital pathology services. Mayo Clinic has partnered with Google and is leveraging their technology in two ways. The program will extend Mayo Clinic’s comprehensive Longitudinal Patient Record profile with digitized pathology images to better serve and care for patients. And we are exploring new search capabilities to improve digital pathology analytics and AI. The Mayo/Google project is being conducted with the help of Sectra, a digital slide review and image storage and management system. Once proof of concept, system testing, and configuration activities are complete, the digital pathology solution will be introduced gradually to Mayo Clinic departments throughout Rochester, Florida, and Arizona, as well as the Mayo Clinic Health System.

The new digital capabilities taking hold in several pathology labs around the globe are likely to solve several vexing problems facing the specialty. Currently there is a shortage of pathologists worldwide, and in some countries, that shortage is severe. One estimate found there is one pathologist per 1.5 million people in parts of Africa. And China has one fourth the number of pathologists practicing in the U.S., on a per capita basis. Studies predict that the steady decline of the number of pathologists in the U.S. will continue over the next two decades. A lack of subspecialists is likewise a problem. Similarly, there are reports of poor accuracy and reproducibility, with many practitioners making subjective judgements based on a manual estimate of the percentage of positive cells for a biomarker. Finally, there is reason to believe that implementing digital pathology systems will likely improve a health system’s financial return on investment. One study has suggested that it can “improve the efficiency of pathology workloads by 13%.” ²

As we have said several times in these columns, AI and ML are certainly not a panacea, and they will never replace an experienced clinician or pathologist. But taking advantage of the tools generated by AI/ML will have a profound impact of diagnosis and treatment for the next several decades.

 

References

1. Hollon T, Pandian B, Adapa A et al. Near real-time intraoperative brain tumor diagnosis using stimulated Raman histology and deep neural networks. Nat. Med. 2020. 26:52-58.

2. Ho J, Ahlers SM, Stratman C, et al. Can digital pathology result in cost savings? a financial projection for digital pathology implementation at a large integrated health care organization. J Pathol Inform. 2014;5(1):33; doi: 10.4103/2153-3539.139714.

Shift Happens

Dataset shift can thwart the best intentions of algorithm developers and tech-savvy clinicians, but there are solutions.

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

Generalizability has always been a concern in health care, whether we’re discussing the application of clinical trials or machine-learning based algorithms. A large randomized controlled trial that finds an intensive lifestyle program doesn’t reduce the risk of cardiovascular complications in Type 2 diabetics, for instance, suggests the diet/exercise regimen is not worth recommending to patients. But the question immediately comes to mind: Can that finding be generalized to the entire population of Type 2 patients? As we have pointed out in other publications, subgroup analysis has demonstrated that many patients do, in fact, benefit from such a program.

The same problem exists in health care IT. Several algorithms have been developed to help classify diagnostic images, predict disease complications, and more. A closer look at the datasets upon which these digital tools are based indicates many suffer from dataset shift. In simple English, dataset shift is what happens when the data collected during the development of an algorithm changes over time and is different from the data when the algorithm is eventually implemented. For example, the patient demographics used to create a model may no longer represent the patient population when the algorithm is put into clinical use. This happened when COVID 19 changed the demographic characteristics of patients, making the Epic sepsis prediction tool ineffective.

Samuel Finlayson, PhD, with Harvard Medical School, and his colleagues described a long list of data set shift scenarios that can compromise the accuracy and equity of AI-based algorithms, which in turn can compromise patient outcomes and patient safety. Finlayson et al list 14 scenarios, which fall into 3 broad categories: changes in technology; changes in population and setting; and changes in behavior. Examples of ways in which dataset shift can create misleading outputs that send clinicians down the wrong road include:

• Changes in the X-ray scanner models used
• Changes in the way diagnostic codes are collected (e.g. using ICD9 and then switching to ICD10)
• Changes in patient population resulting from hospital mergers

Other potential problems to be cognizant of include changes in your facility’s EHR system. Sometimes updates to the system may result in changes in how terms are defined, which in turn can impact predictive algorithms that rely on those definitions. If a term like elevated temperature or fever is changed to pyrexia in one of the EHR drop down menus, for example, it may no longer map to the algorithm that uses elevated temperature as one of the variable definitions to predict sepsis, or any number of common infections. Similarly, if the ML-based model has been trained on a patient dataset for a medical specialty practice or hospital cohort, it’s likely that data will generate misleading outputs when applied to a primary care setting.

Finlayson et al mention another example to be aware of: changes in the way physicians practice can influence data collection: “Adoption of new order sets, or changes in their timing, can heavily affect predictive model output.” Clearly, problems like this necessitate strong interdisciplinary ties, including an ongoing dialogue between the chief medical officer, clinical department heads, and chief information officer and his or her team. Equally important is the need for clinicians in the trenches to look for subtle changes in practice patterns that can impact the predictive analytics tools currently in place. Many dataset mismatches can be solved by updating variable mapping, retraining or redesigning the algorithm, and multidisciplinary root cause analysis.

While addressing dataset shift issues will improve the effectiveness of your AI-based algorithms, they are only one of many stumbling blocks to contend with. One classic example that demonstrates that computers are still incapable to matching human intelligence is the study that concluded that patients with asthma are less likely to die from pneumonia that those who don’t have asthma. The machine learning tool used to come to that unwarranted conclusion had failed to take into account the fact that many asthmatics often get faster, earlier, more intensive treatment when their condition flares up, which results in a lower mortality rate. Had clinicians acted on the misleading correlation between asthma and fewer deaths from pneumonia, they might have decided asthma patients don’t necessarily need to hospitalized when they develop pneumonia.

This kind of misdirection is relatively common and emphasizes the fact that ML-enhanced tools sometimes have trouble separating useless “noise” from meaningful signal. Another example worth noting: Some algorithms designed to help detect COVID 19 by analyzing X-rays suffer from this shortcoming. Several of these deep learning algorithms rely on confounding variables instead of focusing on medical pathology, giving clinicians the impression that they are accurately identifying the infection or ruling out its presence. Unbeknownst to their users, the algorithms have been shown to rely on text markers or patient positioning instead of pathology findings.

At Mayo Clinic, we have had to address similar problems. A palliative care model that was trained on data from the Rochester, Minnesota, community, for instance, did not work well in our health system because the severity of patient disease in a tertiary care facility is very different than what’s seen in a local community hospital. Similarly, one of our algorithms broke when a vendor did a point release in its software and changed the format of the results. We also had a vendor with a CT stroke detection algorithm run 10 of our known stroke patients through its system and was only able to identify one patient. The root cause: Mayo Clinic medical physicists have optimized our radiation exposure to 25% of industry standards to reduce radiation exposure to patients, but that changed the signal to noise ratio of the images and the vendor’s system wasn’t trained on that ratio and couldn’t find the images.

Valentina Bellini, with University of Parma, Parma, Italy, and her colleagues sum up the AI shortcut dilemma in a graphic that illustrates 3 broad problem areas: Poor quality data, ethical and legal issues, and lack of educational programs for clinicians who may be skeptical or uninformed about the value and limitations of AI enhanced algorithms in intensive care settings.

As we have pointed out in other blogs, ML-based algorithms rely on math, not magic. But when reliance on that math overshadows clinicians’ diagnostic experience and common sense, they need to partner with their IT colleagues to find ways to reconcile artificial and human intelligence.