What is medical artificial intelligence (AI)?

One of the most challenging and valuable domains for AI is medicine. Both the opportunities and the dangers are great in applying the technology to healthcare overall.

The value of improved medical care is immediate, especially for people suffering from diseases that cannot presently be adequately treated. Artificial intelligence (AI) may have the potential to see what humans cannot and provide a level of care that is otherwise beyond our reach. And when AI algorithms work well, they can be shared widely in cost-lowering ways. 

Risks and rewards

There are, however, both risks and rewards to medical AI. In a 2020 survey of medical professionals, 79% of respondents reported believing that the technology could be useful or very useful. But 80% completely or partially agreed that the risks to privacy could be very high, while 40% completely or partially rated the potential risks “more dangerous than nuclear weapons.”

AI has enabled the development of technologies that go beyond natural human processes, among other risks. Nanotechnology, gene editing, in-vivo networking (INV), the Internet of Bodies and amalgams such as the Internet of Bio-Nano Things (IoBNT) are among the technologies that offer both promise and potential harm.

Also read: 10 top artificial intelligence (AI) applications in healthcare

What are the challenges of medical AI? 

Scientists approaching medical AI want to leverage the technology’s natural abilities while limiting the potential harm. All applications of AI come with challenges, but using this technology to improve health is particularly complicated. Here are some of the challenges:

• Imperfect sensors: Data gathered from medical sensors is often noisier and less precise than in other domains such as photographic classification. This is especially true when sensors penetrate a living, breathing human being. CT or MRI scanners return pixelated and blocky images with many artifacts that can cloud or obscure the details in question. X-rays may be better but they, like CT or MRI scanners, can detect some types of body parts better than others. 
• Chaotic systems: The parts of living things are often changing, sometimes dynamically. They’re not fixed targets for algorithms. In many cases, sick people have more complex and dysfunctional systems that are harder for algorithms to analyze because they are not behaving normally. 
• Privacy: Medical information is often protected by stringent laws and regulations because patients are sensitive about their personal data being shared with the larger world. While several clever and useful approaches can protect people’s identities, these require more work, can inject potential error and hold inherent risk. 
• Limited knowledge: While the medical profession has accumulated a reservoir of understanding about human bodies, there are still many areas that are a mystery. Although AI algorithms can sometimes be useful when we don’t know much about the topic, there are still limitations. Sometimes we don’t even know the right questions to ask.
• Cautious approach: Because doctors and nurses understand that there could be dangers, they are often quite careful and hesitant to try new techniques.
• Tight regulation: Governments tightly regulate medical devices and software. Levels of testing and development that are acceptable for other domains often do not meet the standards for medical technology set by the governing U.S. agencies, for example. 

What are the opportunities for medical AI? 

While there remain deep challenges to using AI in medicine, there are also many opportunities for improving care. The technology can offer solutions that humans aren’t able to duplicate. Here are a few ways it can help: 

• Chaotic systems: The human body is quite complex and human caregivers often have trouble seeing complex or chaotic events. Noise or random extraneous events can cloud their vision. AI algorithms focus on data and learn to extract valuable insights from hundreds of data readings. They may be able to do a better job of focusing than humans can. 
• Better sensors: AI applications may have access to information that humans can’t see. Some sensors pick up infrared or other wavelengths that the human eye can’t perceive. The technology can identify small changes with more precision than normal human perception. Better information can lead to better decisions.  
• Unbiased: AI processes only the data it is given. While there can be biases in the data itself, this blind focus still offers us the opportunity to eliminate confounding variables that can trigger human bias. 
• Indefatigable: As long as electrical power is available, AI apps can see patients and render opinions. This can be extremely valuable late at night or at times when human caregivers are tired or not available. 
• An assistant, not a replacement: In most cases AI is not in direct competition with human caregivers. The technology can offer advice to humans, who decide how much of the advice to accept. This hybrid approach can ultimately capture the best of both human and machine intelligence. 

What are some of the best roles for AI in medicine? 

VentureBeat has elsewhere covered the 10 top AI applications in the healthcare sector more broadly, and here in brief are the medical areas addressed there: 

• Research
• Training
• Professional support
• Patient engagement
• Remote medicine
• Diagnostics
• Surgery
• Hospital care

How are the major companies handling medical AI?

Leading tech providers that are investing heavily in AI are targeting the medical market as well.

• Oracle has invested heavily in medical informatics in part by acquiring Cerner, a leading medical records company. Its product line will leverage Oracle’s investment in data science and AI to optimally treat patients at medical centers that use Cerner medical records. Its Integrated Behavioral Health, for instance, watches patient data to help prevent patient suicide. 
• Microsoft’s Azure cloud supports various medical applications. Its internet of things (IoT) software support can absorb data from hospital medical devices, then connect the data with the various AI packages. Its software for analyzing images can unlock details in radiological data. It is also investing in specialized tools for analyzing the vast sets of data captured by genomics research. 
• Amazon is creating specialized versions of its various AWS products to support medical practice and research. Its SageMaker AI platform can work with HIPAA-protected patient records stored in its HealthLake service. The algorithms support forward-looking research by searching for connections and patterns and can decode some of the unstructured text data by using natural language models. 
• Google’s Healthcare Data Engine allows researchers and care providers to track and query the information gathered from patients and research subjects. This HIPAA-compliant space offers direct connections to all of Google’s data analytics and artificial intelligence options such as VertexAI.
• IBM offers a cautionary tale about the challenges facing the application of AI to medicine. After significant investment, the company recently sold its IBM Watson Health assets to Francisco Partners, which has launched it as Merative. The software links algorithms under one brand for helping researchers, regulators, doctors, hospitals, insurance companies and patients. Its Clinical Development product, for instance, manages the coding and data storage needed for tracking patients across studies and visits. MarketScan Treatment Pathways searches large patient databases to identify optimal options for providing care. 

How are some startups delivering medical AI? 

Thousands of startups want to use the power of AI algorithms to change medicine. Summarizing them in a short article like this can’t be done. Still, it’s possible to offer a brief list with some examples for illustration. 

• Some startups are working on the front lines of care. Sensely, for example, is creating a bot that can offer patients automated advice. This can save nurses time while bringing faster answers to patients. 
• Others are working deeper in the labs. Atomwise, for instance, wants to improve drug discovery by helping chemists and pharmacologists evaluate the many different potential drugs for efficacy.
• A common use case is the analysis of medical images produced and interpreted by radiologists. Medical Harbor, AetherAI, ButterflyNetwork, Enlitic and RadLogics are just a few of the startups creating platforms that help radiologists capture and interpret images. They are focused on improving productivity, limiting mistakes and in some cases enabling earlier and more detailed findings.
• Molecular Devices and PathAI are examples of companies bringing similar algorithms to the work of pathologists, who often use imagery to analyze blood and tissue samples. The algorithms can speed up and automate repetitive tasks like counting cells that match criteria that indicate malignancy, for example. 
• Companies that specialize in capturing and storing medical records are also working to integrate artificial intelligence algorithms. Roam Analytics and Sopris are integrating AI techniques to automate medical records by improving accuracy, automating classification and increasing the accuracy of any internal data science studies of this information. 

Is there anything that medical AI can’t do? 

Some obvious limitations of medical AI are similar to those that confound all AI algorithms: If the training data is spotty, biased, noisy or limited, the resulting model will echo all of these problems. 

Gathering data is often more challenging in healthcare than in other domains. Between regulations, sensitivity of information, and difficulty in gathering information in a clinical setting, the datasets will naturally be less comprehensive and more prone to error. There is also not the same opportunity to redo the data-gathering that is possible in some other fields. 

In many cases, medical datasets are much too small for training AI. While some AI models are based upon millions or billions of data elements, some medical studies include only a handful of patients. The scale is markedly different and there’s not the same opportunity to rely upon large datasets to squeeze out error.

Medical AI is also limited by the power of medicine itself. If human intelligence doesn’t have a workable solution, then AI can’t provide one either. If the medical science is unclear or imperfect, the AI will be too. As reflected in the survey of medical professionals referenced above, both the potential and the risks of medical AI are great. The challenge is to maximize the former, while keeping the latter within an acceptable range.