AI Doctor

CHAPTER 2Building Robust Medical Algorithms

IF THERE'S ONE ISSUE that needs to be front and center in artificial intelligence (AI), it's the issue of data: getting enough of it to train the algorithms, having a steady flow of it when you implement the algorithms in the real world, ensuring that it's representative of the patient population and protecting it effectively. MIT Computer Science and Artificial Intelligence Laboratory Clinical Decision‐Making Group head Peter Szolovits told me a few years ago, “A bad algorithm trained with lots of data will perform better than a good algorithm trained with little data.” As such, the process of using AI to transform healthcare will only reach completion if the myriad issues with healthcare data are addressed over time. Fortunately, the quantity of digitized data in healthcare is exploding. From electronic health records (EHRs) to wearables and apps, this trend is expected to continue to increase exponentially (Figure 2.1).

But there are some serious issues. We'll delve into many of them in this and the next chapter, but those that come straight to mind include fragmented data that resides in different information systems (for the same patient!), unstructured data, data gaps, and errors, exchanging healthcare data between different providers, changing data, and so on. In the early days of digital health, I started a population health management software company., Acupera. Our platform used data from different sources such as EHRs, ...