“Algorithms Are the New Drugs”
We sat down with digital health expert Dr. Hugh Harvey to better understand the challenges that startups face when building and certifying AI technologies for medical devices. Hugh is the Managing Director at Hardian Health, where he helps companies bring digital solutions to healthcare. He has worked with countless startups and is familiar with everything from formulating the problem to building, deploying, and regulating AI technologies. We are very excited to be able to learn from his experience and share his answers to our questions with teams getting started on this journey.
From the start, companies building medical AI face two key challenges: finding the right data and finding a use case. I’ve seen companies successfully start with either – as this is, to a large extent, a chicken and egg problem. In my opinion, the best way forward is to first find a use case that creates clear clinical value and then look for the right data. It is particularly important to make sure that AI is the right tool for the job, as often a simple rule-based solution would suffice!
With today’s tools, building the model is the easy part. The challenges lie elsewhere and start with the data. Doctors do not have machine learning in mind when they annotate patient data, and obtaining high-quality labels is often much harder than startups expect. Yet, it is the key to success, and investing in relabeling datasets is often required. Data challenges do not stop there. Even with clean labels, having access to representative data, free of biases, is often impossible. Moreover, one is often unaware of some hidden sub-stratifications. You need not only more data but also better data. Actually, it is often necessary to do double the work since the gold standard for validation is to collect and annotate a dataset from a distinct source!
