Artificial intelligence (AI) is on the road to dramatically change the practice of medicine. And the effects will be profound for patients, clinicians and diagnostic device manufacturers.
The incidence of diagnostic errors in a non-AI world has been estimated at 10-20 percent, with laboratory results implicated in just 2-4% of the errors. Studies have shown that physicians make diagnostic errors, in part, because they rely on heuristics, or “rules of thumb” as a quick route to a diagnoses. Irrelevant information and external cues can also influence diagnoses.
In contrast, AI diagnostics rely on probabilities and iterative decision rules applied to large quantities of data. While humans may still specify underlying assumptions (such as the statistical measures used), computers can compare thousands of predictions and outcomes, learn, and modify the diagnostic algorithms over time to improve the model. Bias, while still possible, may also be reduced, depending on the source of the data set.
Researchers are already testing ways to deploy computer algorithms to test for medical conditions. Using a neural network model, researchers trained a computer to differentiate between x-rays testing positive or negative for tuberculosis (TB). The resulting model was then tested on a different set of x-rays, and yielded an accuracy rate of 96 percent. Another study found that AI could detect diabetic retinopathy with a greater than 90 percent accuracy rate.
Implications are far reaching. With AI-generated diagnoses, patients should see an improvement in diagnostic accuracy along with better predictive capabilities. Earlier interventions can improve and prolong patient life. And patients in underserved areas with limited access to specialists will benefit from improved diagnostic support.
But downsides include privacy issues related to data sharing, and an increased risk of security breaches. For example, computer scientist Sebastian Thrun imagines a world where your mobile phone analyzes your speech patterns over time to detect Alzheimer’s disease. However, in the absence of regulation, insurers may try to exclude predicted existing and “pre-existing” conditions.
It’s also mixed bag for clinicians. Some clinicians may see their fields of specialization change dramatically or disappear. Using our earlier example, in theory a computer will be able to diagnose conditions like TB or cancer much faster and more accurately than a radiologist or oncologist. As a result, clinicians may find themselves on autopilot (“automation bias”), and lose the ability to diagnose, instead relying on the AI-generated diagnoses. The AI algorithm and the dataset that the computer uses will not be transparent to the clinician. And who is responsible if the AI-generated diagnosis turns out to be wrong?
On the plus side, with the exception of delayed diagnosis errors (i.e. errors created waiting for the AI-generated diagnosis results), AI will likely reduce the incidence of clinical diagnosis errors. Clinicians will benefit from improved access to valuable diagnostics information that goes far beyond what is currently available.
Similarly, for a diagnostics device manufacturer, AI confers many advantages. With its emphasis on outcomes and accuracy, a robust iterative AI model will allow the manufacturer to design and build better products. In many cases, this should translate into a competitive advantage. For example, Medtronic has partnered with IBM to improve diabetes care by using Watson Health to analyze 125 million patient days of anonymous data from glucose monitors and insulin pumps. Manufacturers will also have the means to assess if devices are being used correctly. And over time, as older devices become obsolete, customers will be motivated to purchase new AI-enabled devices.
From a product development standpoint, diagnostic devices and systems will need to be designed to incorporate current and projected AI capabilities. Data needs to be validated and transmitted reliably and securely back to a centralized database or blockchain enabled system, in a format that is compatible with the AI system. The AI output will need to be monitored, and feedback loops established to improve the AI model.
The ability to measure and predict outcomes with a predefined level of accuracy will have implications across all functional areas of the manufacturer’s organization.
On the other hand, AI-generated data will invite greater scrutiny, so expect more attention from regulatory agencies, insurers, prospective purchasers, patients, and legal counsel.
AI is here, and it’s here to stay.