In the early 2000s, researchers borrowed a technique from astronomy to make ultrasound images sharper, creating a device to screen women with dense breast tissue for cancer. They found that using a sound signal can better differentiate between tumors and dense tissue than is possible with mammography X-rays.
The problem? It took hours to produce an image with the prototype machine, which wouldn’t work for a screening test.
The Detroit-based team from Wayne State University and the Karmanos Cancer Institute headed back to the lab to refine their ultrasound tomography device. In part, they developed algorithms for better resolution and contrast in the images it produced, and optimized the instructions’ coding to make them as efficient as possible in terms of compute time and memory used.
Today, some combination of software algorithms, artificial intelligence (AI) and machine learning (ML) have become essential to all but the most basic medical breakthrough devices, including imaging systems such as the SoftVue 3D whole breast tomography system, diagnostic equipment, remote monitoring systems and wearables.
“AI has indeed become an integral part of the evolution of medical devices,” says Zen Koh, co-founder and global CEO of Fourier Intelligence International Ltd., a Singapore-based fitness and wellness products company. “While not every medical device incorporates AI at this point, we are witnessing a remarkable trend toward AI integration in healthcare technology.”
AI likely will take on an even larger role in the future, according to Alexandra Murdoch, a medical analyst at the intelligence company GlobalData, London.
“Most therapy areas within medical are looking at AI or may have a device that incorporates AI into it, but some common applications of AI and ML in medical currently are data management, remote patient monitoring and surgery, diagnostic and procedural AI assistants, and clinical trial design,” Murdoch says. “Overall, AI and ML in medical devices can enhance diagnostic accuracy, enable personalized medicine, support clinical decision making and contribute to efficient healthcare delivery.”
AI’s ability to analyze complex medical data, detect patterns and provide actionable insights has made it a valuable tool for medical professionals in improving patient care, Koh says. However, regulatory considerations, the need for rigorous testing and the complexity of integrating AI into devices can affect the pace of adoption.
Companies such as Fourier and SoftVue’s maker, Delphinus Medical Technologies Inc., Novi, Mich., keep adding to the growing stock of breakthrough medical devices equipped with digital prowess.
“The U.S. Food & Drug Administration has approved more than 600 AI/ML-enabled devices since 1995,” notes Catherine Pugh, senior manager of digital health at the Consumer Technology Association (CTA), Hopewell, Va., sponsor of the annual CES electronics show in Las Vegas.
CTA generally supports the FDA’s risk-based approach to regulation, including for AI, according to Pugh. Non-FDA regulated healthcare applications that use AI, such as health administrative software or consumer health apps that don’t make a medical claim, should be considered low risk. For both scenarios, consensus-based industry standards should play an important role in driving transparency and accountability in AI, Pugh continues.
Algorithms and AI also help enhance bottom-line performance. Medtronic plc’s cranial and spinal technology business grew 7% in the second quarter of fiscal year 2024, driven in part by its ML-assisted UNiD ASI spine surgery product. The company’s endoscopy business grew 13%, pushed upward by continued adoption of GI Genius, which uses the power of AI to detect polyps in real time during a colonoscopy, resulting in a 50% reduction in missed polyps versus a standard colonoscopy, CEO and Chairman Geoff Martha told analysts on an earnings call last November.
The company is depending on digital technologies to spur growth, Martha noted. “We’re decisively allocating capital into fast-growth med-tech markets, and fueling innovative technologies in areas like robotics, AI and closed-loop systems that will drive our growth over the next decade,” he said during the analyst call.
One of the devices cleared by the FDA is SoftVue, which earned marketing approval from the federal agency in 2021. After 20 years in development and testing, Delphinus plans to commercially launch the product later this year.
When paired with mammography, SoftVue has shown as much as a 20% improvement in cancer detection while also reducing false positives and decreasing unnecessary call-backs and biopsies for women with dense breasts, according to company literature. To help alleviate discomfort during mammography, the device replaces a regular ultrasound’s gel with warm water, which surrounds the breast during imaging, and its design is purposely ergonomic and soft.
While not incorporating AI, SoftVue’s development reveals the complexity of integrating advanced algorithms and processing data inside of devices versus sending it to the cloud for analysis. For example, the problem of hours-long imaging was resolved with software and hardware fixes, explains Neb Duric, Delphinus’ co-founder and chief technology officer.
In addition to optimizing an algorithm’s code, paring down imaging time required hardware to be assembled in parallel for the large volume of data that needed to be processed in the device.
“I believe the last system we built contains eight GPUs (graphic processing units), and many, many CPUs,” Duric says. “That way, you can reconstruct eight slices at a time out of a typical scan. And it’s that kind of parallelization that allowed us to get to a point where we could finally image on clinically relevant time scales. That took two or three years to fine tune.”
Borrowing from his background in astrophysics and astronomy, Duric applied sound signals to refine imaging for dense tissue.
“One of the things that we learn in astronomy is how to make images clearer,” he says. “When I switched over to medical imaging, we tried some of those techniques for ultrasound imaging, and we found that we can, in fact, improve imaging quite a bit by using a technique called ultrasound tomography.”
In addition to commercializing SoftVue, Delphinus plans to bring production in-house from a contract manufacturer this year.
“We want to be able to have control of the manufacturing process,” says CEO and President Scott White. “What we mean by that is, we want the ability to scale up, scale down, do second shift, do weekends, whatever it takes to get the product to the customer. We want to have the flexibility in the process to do that. And then we also want to manage the cost more effectively.”
White says in-house production and refurbished warehouse space at the company is expected to cut labor costs in half, with a target run of two of the hand-built devices per month.
Nanshu Lu, a professor in aerospace engineering and engineering mechanics at the University of Texas at Austin (UTA), knows very well the incremental process required to develop a breakthrough medical device such as an electronic “tattoo” to monitor the heart. In the latter’s case, the flexible, stretchable e-tattoo keeps watch over the heart’s electrical activity with electrocardiogram (ECG) and the vital organ’s valves with seismocardiogram (SCG) technology.
“The ECG sensor interfaces with the human body via bio-compatible graphite film electrodes, and the SCG is measured using a high-resolution, low-noise accelerometer that can detect subtle vibrations caused by the heart,” she and her co-authors wrote in the Advanced Electronic Materials journal. “Wireless connectivity incorporating Bluetooth Low-Energy (BLE) is utilized to stream the data in real time to a host device.”
The SCG picks up mechanical signals from the valve, similar to listening with a stethoscope that produces a phonocardiogram (PCG) wave. The PCG, however, isn’t suitable for a wearable, which led UTA to develop the SCG sensor.
“The only difference is that the PCG has higher frequencies in the audible range,” she notes. “But the SCG is lower frequency.” The next issue was power for the device. Lu and her team developed an energy-saving strategy by letting the SCG sensor’s chip sleep until it’s triggered by the ECG sensor’s chip. “Currently, our power consumption is only 3 mW, which is the lowest power-consuming, dual-mode wearable cardiovascular sensor that has ever been reported,” Lu asserts. In order to meet their electrical functionality and mechanical compliance goals, Lu’s group manufactures an electrode layer for the e-tattoo. Because conventional flexible circuit makers can’t make a stretchable circuit, a second, stretchable layer is outsourced to JLCPCB in Taiwan.
“In the future, as we acquire more and more data, which is both multi-modal—meaning we will have electrical, mechanical, thermal, optical or even biochemical data—and either for a few minutes or even a few weeks, in those cases we have to process huge amounts of data, which are very complex,” she says. “Therefore, we would have to use artificial intelligence to help us process the data, and also to help us interpret and make sense of the data so that doctors or athletic trainers will be able to make use of that.”
The heart e-tattoo can replace and/or enhance a Holter (portable ECG monitor), according to Lu. The technology is suitable for post-surgery patients, people with atrial fibrillation or hypertension, she says, as well as for elderly people who are at high risk for cardiovascular disease.
Innovators such as Lu may have been on CEO Chris Rogers’ mind when he started SensiML Corp., Beaverton, Ore., to market software that creates customized ML sensor algorithms. “Our tool is taking advantage of the ongoing advancements in cloud compute, on being able to take sensor data from edge devices and apply an automated approach to generating algorithms,” Rogers says.
SensiML got its start in wearable computing as a companion software tool for Intel’s chips intended for small devices such as fitness bands and other types of personal, health/fitness or medical applications and derive insight or meaningful information from raw sensor data.
The process can be incredibly complicated to develop, confides Rogers, who worked at Intel nearly 20 years before starting SensiML. It usually requires a domain expert, a firmware engineer and data scientists, he notes, adding that the various experts don’t necessarily speak the same technical language. And, in most cases, an existing team doesn’t have all of those capabilities.
“So what we were trying to do was really put together a tool that could help automate the data science aspects of this process, take time series data and build models just by a search process to come up with valid, accurate models that meet the parameters the user defined.”
SensiML’s software is broader than medical, and can be used for applications such as predictive maintenance for machine tools, and activity recognition or keyword detection for smart home devices.
“So we’re really trying to take that process and democratize it to where embedded developers who might have an idea for new innovation projects can collect some data on their own, and then devise models that can function at the edge,” Rogers says.
DornerWorks Ltd., an embedded software engineering company in Grand Rapids, Mich., used SensiML software to create a solution that could be in the next breakthrough medical device. The technology classifies the type and location of movement in an environment based on specific patterns of change—learning and adapting to its unique environment and subject—to detect dangerous situations such as a person falling.
“The solution relies on Wi-Fi capabilities to overcome the typically high cost and application memory constraints of using cameras, microphones, accelerometers and other popular alternatives,” says David Norwood, technical strategist at DornerWorks. “Wi-Fi is a valid option to consider for new product features that rely on motion detection, whether that motion is fall detection in a bathroom or virtually any other business or home environment.”
If a person is injured in a fall, Fourier Intelligence’s ArmMotus EMU may be needed. The ArmMotus is a robotic trainer that’s part of Fourier Intelligence’s portfolio of advanced robotic exoskeletons and virtual reality-based therapy platforms to aid those with upper- and lower-body balance and movement impairments.
“Coupled with gamified therapy, these innovations enable healthcare providers to deliver personalized and effective rehabilitation programs, improving the patient’s overall recovery outcomes and quality of life,” Koh says. “While most of our users are adults, our devices can assist children with health conditions requiring neuro-rehabilitation, such as cerebral palsy.”
The device’s AI-supported, assist-as-needed technology can detect a patient’s intentions during therapy, Koh adds. This is especially useful for patients who have little muscle contraction. In these cases, the ArmMotus can help a patient complete a movement or action. It works by driving a cable-driven mechanism, helping the patient complete the movement and ultimately achieving neuroplasticity through sufficient repetitions.
The technology was transferred from the University of Melbourne after undergoing a user-centered design approach in its development with clinicians, engineers and users. The ArmMotus underwent extensive internal prototyping and 17 product design iterations before the final product was commercialized.
Koh underscores the need for automated rehab by pointing to the more than 1 billion people worldwide with disabilities that affect their independence and mobility, according to World Health Organization data. The problem is compounded by the aging global population, along with a shortage of clinical therapists.
As a solution, he continued to make the case for rehab robots. “Robot-assisted therapy has freed up the labor-intensive, repetitive work from the therapist, allowing them to focus on work that requires their professional knowledge, like therapy planning, prescription and dexterous manual therapy that a robot cannot replicate. By doing that, a therapist’s work efficiency has been maximized, offering more treatment sessions to the continually increasing patient population. Robot-assisted therapy reduces the therapist’s involvement in labor-intensive training requiring high neuroplasticity repetitions.”
The emergence of robotics and AI has significantly transformed the rehabilitation industry. When coupled with other advanced technologies, the future looks even more promising to Koh.
“As we look ahead, the possibilities of 5G (cellular) technology in remote rehabilitation or telerehabilitation are immense,” he says. “Patients who cannot travel to a hospital or clinic can now receive rehabilitation therapy from the comfort of their homes, enabling them to access rehabilitation services from anywhere in the world. This approach not only increases accessibility, but also makes rehabilitation services more convenient and efficient for patients.”
Connect With Us
