When IBM Watson won Jeopardy in 2011, the public was awakened to the exciting potential of Artificial Intelligence (AI). The platform demonstrated a state-of-the-art capacity for natural language processing and synthesizing answers from massive textual databases with near real-time response. This one-of-a-kind feat was the culmination of a four-year effort by a dedicated 20-person engineering team. Professionally, I applauded their remarkable achievement, having directed multiple interdisciplinary R&D organizations responsible for rolling out innovative products. Moreover, having spent years as a PhD researcher investigating AI and related technologies in industrial lab settings, I found it more than a little intriguing that AI was finally becoming mainstream.
IBM capitalized on the apparent success of Watson by launching AI corporate divisions in different industries, including Watson Health for healthcare in 2014. AI algorithms are known to be hungry consumers of data, and healthcare is in no short supply with the rise of the Internet of Things (IoT) where hospitals now have 10 to 15 connected devices per bed. The industry hype suggested that IBM was on the verge of creating an AI doctor called, Watson. However, after significant investment by IBM with high-profile technology acquisitions, the AI achievements realized were rather modest compared to the initial promise. It took an entrepreneur with both a computer science and medical background to realize an AI healthcare breakthrough. Michael Abramoff MD PhD founded a startup, IDx, that applied both a deep understanding of diabetic retinopathy and a realistic appraisal of the capabilities of the AI.
In a press release from April, 2018, the FDA announced permission for IDx to market the first medical device leveraging AI to detect a form of eye disease in adults who have diabetes. Diabetic retinopathy is the most common cause of vision impairment for working-age adults and the leading cause of vision loss (24,000 yearly) for more than 30 million Americans living with diabetes. It’s estimated that 7.7 million Americans have diabetic retinopathy and this number is expected to double by 2050. Early detection of retinopathy is key to preventing damage in blood vessels of the retina caused by high levels of blood sugar. Unfortunately, 50 percent of patients with diabetes do not see their eye doctors on a yearly basis, so they are not adequately screened for the disease. IDx-DR is a software program that uses AI to detect greater than mild diabetic retinopathy in images captured with a retinal camera. This medical device enables primary care providers to make the quantitative determination that a patient should see an eye care specialist for diagnostic evaluation and possible treatment.
Today, there is a large pipeline of AI healthcare startups trying to replicate IDx’s success for alternate indications. As Director of Technology and Innovation at Syncro Medical, I see no shortage of new AI opportunities emerging for the MedTech community. Given the FDA’s reliance on precedents, the IDx case study provides an excellent roadmap for moving AI products through the regulatory process. The challenge is to navigate around the commercialization pitfalls, such as the ones IBM faced. Medical device manufacturers need to adopt a systematic approach to AI, and the best starting point for this exploration is the data.
What story does the data tell?
Given the data-driven nature of AI, data needs to be at the heart of any viable AI strategy. MedTech firms may have access to vast amounts of information through their products and strategic partnerships, but the important question to consider is “What story does the data tell?” For example, in the digital health space, wearables such as Fitbit® are known to capture a virtual stream of information literally around the clock. However, if a given user has a sedentary lifestyle or poor sleeping habits, he or she may not find this feedback useful. Ezekiel J. Emanuel MD PhD, Professor of Medical Ethics and Health Policy at the University of Pennsylvania, published an article that medical conclusions drawn by algorithms from these novel data sources are often suspect. Consequently, the initial stage of any AI project typically involves systematic number-crunching to uncover trends and relationships between elements of the dataset. This analysis is called unsupervised learning and involves discovering the signal hidden in the noise.
Does the story from the data make sense?
When a consistent pattern is found in the data, it should be validated against healthcare subject matter expertise. It is vital to find out early whether the data is trending in the appropriate direction. There was a well-publicized study where researchers tried to predict the risk of death for a hospitalized individual with pneumonia, so that low-risk individuals could be sent home while high-risk individuals would be treated. The data-driven AI model promoted the counterintuitive conclusion that patients with asthma and pneumonia were less likely to die than those with asthma only. The researchers traced this result back to the policy that individuals with both pneumonia and asthma are typically admitted to an intensive care unit and receive aggressive treatment, which positively impacted their survival rate. The failure in the AI model design was not considering that the level of treatment for pneumonia depended on whether the patient had asthma. This is commonly cited in the literature as an example failure for AI, but better to be viewed as a cautionary tale about the need for rigorous validation of AI analysis.
Is the data fair?
When conducting clinical trials, medical device manufacturers understand the need to assemble a reasonably sized, diverse pool of subjects. For example, the distribution of race for candidates in studies funded through National Institutes of Health grants needs to match the general population in the geographic area where the research is conducted. However, researchers analyzing real-world data also should be cognizant that there may be systematic bias across the healthcare system. For example, a common practice within organizations is to use healthcare dollars spent as a proxy for healthcare need. It has been shown that healthcare expenses for those in minority groups tend to be more modest for equivalent medical conditions. Consequently, patients with similar symptoms and health history are getting divergent prognoses with race being the main differentiator. If unchecked, the inherent disparities and inequities in the healthcare system can skew AI risk assessments to favor white patients. On top of the important ethical concerns, AI platforms bias can lead to a public relations nightmare. Consider Microsoft’s release in 2016 of a chatbot that learns patterns of speech through Twitter conversations. Within 16 hours, the platform tweeted 95,000 times with a surprising number of messages that were both abusive and offensive.
Is the data representative of the real world?
Another valid concern for AI is this: a system developed in one hospital may not work nearly as well when deployed at another facility. Subtle differences between facilities, such as the brand of MRI device or operational procedures, can produce significantly different outcomes. Consequently, it is important to have access to relevant data that accurately reflects the actual scenario of operation. For example, the retinal images captured for both training and testing of IDx-DR covered the range of diabetic retinopathy in adults. Moreover, given that the product is labeled to be operated by an individual with a high school diploma and 4 hours of training, non-expert technicians naturally were used in the product validation. In order to achieve strong results, the algorithm was designed not to make a retinopathy determination in about 4% of the cases for the target population.
Is the story from the data reliable?
A common approach in AI, called supervised learning, is to “teach” algorithms the correct solution that should be derived from a given dataset. For example, orthopedic doctors routinely assess cartilage deterioration of knee joints in x-rays as stages. An AI algorithm designed to mimic this assessment can use stage labels, generated by subject matter experts, as ground truth data during training and testing. The key to robust algorithm performance is to adopt a design process that is invariant to nuances in the data. The labeled dataset is often randomly separated into three sets: training, testing, and validation. Test and validation sets are used to confirm algorithm performance by checking the output against the label for a given input image. An algorithm that performs robustly, no matter how the dataset is randomly separated, is expected to perform as robustly in the field of operation. The IDx-DR platform achieved nearly 90% accuracy rate in determining greater than mild diabetic retinopathy.
Do you want to tell a compelling story from your data?
Embarking on an AI journey can be a daunting challenge for companies new to the field, especially when even established players, such as IBM, had difficulties. However, the medical device manufacturer is often uniquely positioned with both the passion for the patient population and the subject matter expertise necessary for a successful AI project. Real-world domain knowledge guides the systematic exploration of various analytic approaches to reach a commercially viable solution. The natural starting point for an AI investigation is a deep dive of the relevant data:
- What compelling conclusion can be drawn from the data?
- Does the conclusion make sense?
- Will it generalize across medical facilities?
- Is the supporting data analysis fair?
- Does the analysis suggest the conclusion is robust?
At the most basic level, AI is a tool kit of algorithms that leverage historical data analysis to perform a series of complex operations on input data. If you do not have the appropriate resources in-house, it is important to find a partner with the right balance of AI and MedTech experience to realize your vision. Syncro Medical is ISO 13485 and ISO 9001 certified and has been building medical devices that incorporate sophisticated data analytics for over 30 years. Let Syncro Medical help you navigate the AI hurdles in your path, so your data can tell a compelling story.
About Gregory House, PhD
Gregory House spearheads innovation and new technologies at Syncro Medical. Greg earned his PhD in Electrical and Computer Engineering from Carnegie Mellon University and has a 20-year track record of commercializing digital technologies. He was Principal Investigator for millions in biomedical funding, awarded 11 US patents, and published 8 peer reviewed journal articles. Greg is credited with launching several innovative healthcare products in demanding real-world environments. His hands-on experience includes driving intellectual property strategy, building multi-disciplinary engineering and operation teams, and collaborating with research and medical institutions. In addition to applying his expertise in Artificial Intelligence, Greg is continually researching new technologies to advance client products.
About Syncro Medical
Syncro Medical is a leading provider of software development services for medical device and life science companies. With 30 years of proven expertise, Syncro Medical maintains a culture that values innovation, quality, and craftsmanship. In its ISO 13485/9001 certified development center, Syncro Medical teams accelerate product development efforts and provide critical resources and fresh insights. Across 400+ projects, Syncro Medical has maintained a best-in-class reputation in the industry with its 100% client satisfaction rate.