The medical industry is evolving rapidly due to greater deployment of information technology. While the industry has always been slow to adopt more IT due to privacy and security concerns, the market is poised for a dramatic shift as artificial intelligence (AI)—both algorithmic and machine learning—makes an entrance and new AI technologies are approved by the FDA. The key in reducing the AI fear factor and increasing AI acceptance will be using the right tool for the right job.
Medical Imaging: Poised for Change
In particular, medical imaging and radiology are set to change. Market drivers are creating major opportunities. First, the volume of imaging is growing rapidly, while the number of radiologists is dwindling. Infinium Global Research estimates the global market will grow from $30 billion in 2015 to $40 billion by 2021. The growth is predicated on the establishment of developments such as new 5G phone service networks that speed communication of large imaging files, reduced costs for cloud storage, and an aging population requiring more care in developed countries.
Yet, in the United States, the American College of Radiology reports that radiologists over 55 have been retiring at an accelerated rate. At the same time remaining radiologists are moving into subspecialties, draining the general pool even further.
And that’s just in the United States. In China, where the doctor-patient ratio is far higher, radiologists typically must interpret more than 100 scans per day, per doctor—and each scan may have thousands of images to choose from.
The result? More mistakes. A 2015 study published by the Journal of the American College of Radiology showed that faster reading and reporting times fostered by a volume-based reimbursement system can result in a 27% increase in error rates compared to a 10% error rate at normal speeds. This doesn’t include the fatigue experienced because of repetitive, tedious work.
AI is Already Helping
As one of the first sectors to deploy AI, radiology has already made demonstrable progress, becoming a showpiece of sorts for machine intelligence in the medical sector. In April, the FDA approved an AI-based device for screening for diabetic retinopathy in ophthalmic patients and in May gave the go-ahead to an algorithm that identifies wrist fractures. The agency hinted it has many more such applications in the pipeline.
Even AI products still restricted to investigational use are yielding positive results. My own company’s products have been deployed and tested across more than 150 hospitals where a large number of scans have been screened for indications of lung cancer and strokes. The ability to test the system on tens of thousands of scans is critical in documenting the effectiveness of AI tools and is also key in improving the system’s intelligence through continuous machine learning.
AI deployment in the industry has been highly encouraging. Side-by-side comparisons are demonstrating the value of AI as an assistant to radiologists. A recent competition in China between AI-assisted and unassisted radiologists set up by Infervision indicated that image reading speed and accuracy rates increased with AI assistance with only five minutes of AI instruction for participating radiological professionals.
Why AI is Here to Stay: A Useful Tool
The arrival of AI onto the medical scene is a long-awaited event. The stage was set for its arrival with the emergence of digital health platforms, such as EMRs, that collect enormous amounts of data. The logical follow-on to digital health is that some form of machine intelligence is needed to analyze both the large sets of individual diagnostic data (such as thousands of pixelated images) as well as aggregated, anonymized meta data.
Additionally, the advent of personalized medicine, targeted directly to the individual, also makes AI a must-have technology for improved therapy management. In this model, the continuous data collected on an individual patient’s multiple physiological markers from both wearable and in-vivo sensors requires strong computational power that can identify problematic conditions into which doctors previously had no visibility. Here, AI has great potential to extrapolate unexpected connections between a variety of data streams through the use of continuous machine learning as it learns about different phenotypes.
The American Medical Association is fully on board. The organization recently came out with “Augmented Intelligence” recommendations and guidelines for deployment, which include “promotion of development of thoughtfully designed, high-quality, clinically validated health care AI that is designed and evaluated in keeping with best practices in user-centered design, particularly for physicians and other members of the health care team.”
Consumers are gradually accepting the benefits of AI. As much as we like to rail against automated voice systems, those banking and airline systems are pretty good at understanding what we want. And devices like Amazon’s Alexa has furthered that confidence. While the medical AI technology may be transparent to many, the average patient can expect to see an increase in the standard of care, based on staff ability to more quickly process diagnostic screening and provide immediate therapy. Similarly, AI may facilitate equal access to care by offering effective screening while reducing physician workloads in disadvantaged areas with many patients. In the case of more remote and rural hospitals, many of which are in danger of closing in the United States, sending scans to off-site AI services for almost real-time evaluation can ensure the preservation of radiology services even when there are limited or no staff resources.
Looking to the Future
Based on what I’ve seen in the past few years, we are at the beginning of a new age for medicine. With ever-increasing bandwidth, storage and computational power, I expect to see nothing less than a radical transformation of our visual and analytical capabilities in medicine.
Already we are seeing companies that have achieved significant milestones in newer forms of AI-based imaging: 3-D rendering; a technology called AUTOMAP that enhances resolution for MRI scans; tools that can reconstruct images where visual data might be missing; and the emergence of what’s called cinematic volume rendering techniques (CVRT), which can display physiological structures in a photorealistic way. We are clearly on the cusp of an amazing new era.
AI still faces a number of challenges, the largest of which is a cultural resistance to changing workflow habits among busy physicians and medical staff. These professionals are already under pressure in a volume-based reimbursement world. The key for AI developers is to work with doctors during product design to ensure minimal workflow disruption. Developers will need to show how a short ramp-up to deploying new intelligent tools—such as the contest previously mentioned—can help improve outcomes, save time, reduce workload and even help grow business. Working far in advance with well-respected institutions and medical thought leaders on critical requirements can also help pave the way to industry acceptance.
Regulation is another potential hurdle. Since the FDA has only recently begun to evaluate and approve AI tools, there are of course few precedents as a basis for assessment, and the necessary sets of criteria are only now being developed. In radiology, AI has an advantage in the approval process in that it has a proven track record of rapidly facilitating accurate screening. This will be particularly important as the industry transitions to MIPS (Merit-based Incentive Payment System) and a value-based reimbursement system built on improved outcomes. An AI system’s history of analyzing large data sets—in Infervision’s own case about 18,000 scans per day across China and other countries—provides significant depth for the agency to validate effectiveness. In contrast to approval of other technologies such as static MRI equipment where “Day One” of regulatory authorization represents the highest bar, for AI imaging analysis Day One of approval is actually the lowest bar—since the system will continue to learn immense amounts after commercial deployment.
Using AI for the Right Applications
AI should not be seen as a threat to medical professionals, or to patients. AI is most useful for pattern recognition, whether it’s making sense of pixels in an image or looking for trends in many streams of numeric data. AI is designed to take the tedium out of computationally intensive tasks. Those of us in AI are not proposing that medical professionals rely on it for diagnosis—that’s where the experience and intuition of doctors is most beneficial, based on evaluation of patients’ specific case histories. But AI can help us in pointing out problematic areas for closer inspection and can bring considerable machine power to the task so that doctors can spend more time doing what they do best: Results interpretation, meeting with patients and providing care.
As is the case for all new products, medical AI technology must be developed and deployed in partnership with its key stakeholders: Users and patients. This includes the needs to make it reliable, trustworthy, easy to use, transparent, with minimal disruption to workflow used by medical staff, and to patient lifestyle. AI has the potential to add value for doctors, patients and payers alike by improving outcomes, increasing equal access to reliable medical care, and reducing healthcare costs. On top of that, the promise of AI—like digital health—is to assist doctors in making better judgments.
Lets not forget the opportunities for Artificial Mediocrity and Artificial Stupidity.