Flight Path

Between the years 2000 and 2010 – if you exclude September 11, 2001 – there were 250 airline fatalities in the US (1). And between 2010 and 2020 – there were only two (1). Evidently, the airline industry realized airline fatalities were unacceptable – and so damaging to the industry that relevant technology was needed to curb mortality. Interestingly, airline industries don’t compete against each other. Safety is taken as a given.

Why is it considered acceptable that there are misdiagnoses in our industry? Everyone knows that it happens – and tolerates it. Consider this: the airline industry didn’t set themselves a goal to reduce fatalities; they said zero fatalities. And I think the same kind of goal should be set for cancer diagnosis. With AI, we can reach zero diagnostic error. We have the technology, but that is not enough – we also need adoption within clinical practice. Payers need to reimburse. Regulators need to work with the industry, clinicians, and pathologists. In short, it really does require an industry wide effort to achieve this goal.

We have direct access to misdiagnosis data because we deploy our algorithms in pathology labs where they’re used on a day-to-day basis. The normal range of misdiagnosis ranges between two to three percent at excellent labs – but it can go as high as 10 percent (especially in labs that lack subspecialists). If your cancer is misdiagnosed, it’s very likely to progress further and become more severe. The treatment is going to be more aggressive, mortality goes up, and the cost of the healthcare system also rises. It’s a lose-lose situation for everyone involved. We also need to understand that this is happening because of technological and resource limitations. It’s not possible for pathology labs to have three pathologists investigating every case. We need something more…

Our AI algorithms are trained by pathologists, for pathologists – and there are two ways you can deploy them in the lab. First, as a pathologists’ assistant. The algorithm can act as a virtual fellow, where it can review the case in advance and fill out a draft of the report. This way, the chances of a pathologist missing anything is reduced. The algorithms are very objective and can conduct day-to-day tasks, such as measurements and quantifications without getting tired. Over time, pathologists will start to develop confidence in algorithms’ performance and they will be able to work faster – by up to 60 percent. Second, the algorithms can be used in places where it wouldn’t be economical to have an actual human pathologist review the case. So, you can do triage of cases earlier on and make sure that the more urgent cases reach the top of the pile. You can preorder IHC staining to reduce turnaround time. In general, you can plug in the technology at different decision points in the slide preparation phase to make the process more efficient. 

Our algorithms help reduce diagnostic error because they’re extremely accurate – trained on huge datasets. Not only can they accurately detect cancer, but also help differentiate between different cancer subtypes. For example, the technology can tell the difference between invasive lobular carcinoma and invasive ductal carcinoma – or between high and low grade ductal carcinoma in situ. The algorithm can also help detect non-cancerous features, such as microcalcifications, breast biopsies, and perineural invasion. But perhaps the most exciting thing about these algorithms is the way they contribute to health equity; I can see the technology being deployed as part of standard practice in every pathology lab worldwide. 

For this technology to actually become prevalent, it’s critical that physicians and pathologists recognize the value and actively want to use it. Whenever we let pathologists play around with a new technology, they are almost always eager to deploy it in their labs. Regardless, there are still barriers. First of all, we need to acknowledge the importance of regulatory frameworks – and recognize that AI introduces new challenges for regulatory science. Agencies, the industry, and physicians need to work together to accelerate the deployment of this technology; we owe it to patients. The other piece of the puzzle is the economic model. How do we pay for this technology? Again, we need to have a big conversation with all stakeholders.

The use of biopsy revolutionized the field of pathology in terms of diagnostics, and also in terms of predictive power of pathology. This advance has become completely standard. I hope we will soon be able to say the same about AI.

Coming full circle, I’ll finish with a final call to action for zero diagnostic error. 

We need to see misdiagnosis as something that went terribly wrong – because we can prevent it from happening. Clinicians, pathologists, healthcare providers, payers, and regulators need to work together to achieve this vision. Though I don’t think it’s something that is achievable overnight, within a decade I believe we can transform cancer diagnostics to achieve much higher levels of accuracy.

^{Credit: Images sourced from Unsplash.com}