Artificial Intelligence: What it means for health care
For the past several years, experts have predicted that artificial intelligence (AI) would make a splash in health care. I decided it’s time to write about it for my blog, Trends in Health IT. Why now? Because it seems like a good time to tackle this topic, given the recent, high-profile debate between prominent tech leaders about whether or not AI poses a threat to humanity.
AI in health care will be a major shift for the industry, but one that happens slowly and gradually reaches a tipping point. It won’t follow the usual hype cycle we’ve seen recently.
Health care customarily embraces technology in big chunks, especially when financially motivated. The recent wave of EHR adoption, largely from Meaningful Use, is an example of a big chunk of technology. All of the record keeping – plus a great deal of workflow, decision support, and clinical best practices – became encapsulated in monolithic systems. EHRs became a recognizable product category as the industry began to agree on a definition of what constitutes an EHR. So too with billing systems, lab information systems, analytics packages, and dozens of other examples. Health care is accustomed to packing up business functions as software products and then categorizing those products.
AI, however, will not be a standalone product with a strict definition and clean boundaries. “AI solutions” won’t be a product category. Rather, AI will become a built-in feature of each and every product on the market. There will be AI in EHR software, analytics software, patient monitors, and everything else. Users might not even notice that it’s there; they’ll just notice that features work better. Computerized physician order entry systems will recommend more appropriate order sets. Analytics packages will calculate better patient risk scores. Radiology software will provide assistance with interpreting imaging studies.
There are at least two reasons why AI will infiltrate health care as features, rather than separate products. First, AI is not a business function in the same way that record-keeping or order entry is a business function. AI is a collection of techniques for doing various computational tasks. It’s a novel, subtler way of predicting, recognizing, recommending and categorizing. Second, AI is highly specialized. The same AI software that is good at recognizing potential tumors on an x-ray image will not be able to group patients into cohorts based on similar risk factors. As such, it’s a great add-on to existing software, but it’s not great on its own.
Vendors will need to find ways, over time, to improve their software by incorporating AI techniques. It’s a predictable consequence of the competitive drive to deliver better software. By the same token, it’s unlikely that we’ll see software procurement requirements that specifically require AI.
AI will change health care, but it’ll happen without anyone specifically attributing the change to AI.
Will AI be an existential threat to human existence, as Elon and Zuck debated? Not in the five-year timeframe that Trends in Health IT spans, but there will be a number of serious implications to deal with. For example:
- The first several generations of AI-based features will probably be a nuisance as often as they are a benefit. This can be a big problem when it comes to patient care. “Alarm fatigue” is already a problem in clinical settings. Will “AI suggestion fatigue” become an issue?
- A key part of many AI techniques is that they are self-training or self-optimizing, and that means there isn’t always a human-comprehensible way to explain how the AI arrived at an answer, recommendation or suggestion. When a patient’s life depends on a critical decision, it’s not satisfactory for a clinician to choose a therapy “because the computer said so.”
- How will therapeutic devices and systems with advanced AI features be regulated?
- Many AI processes rely on huge data sets, which might mean pooling data across several institutions. This will need to be done in a way that complies with HIPAA, various state laws and institutional policies.
These are questions that will take a lot of time and careful thought to work through, and that’s why AI adoption in health care has not happened yet. Luckily for prognosticators, this means that they can always copy it from one year’s list of Health IT predictions to the next!