Earlier this week, the Journal of the American College of Radiology (JACR) published a paper entitled “Projected growth in FDA-approved artificial intelligence products given venture capital funding.” The authors are affiliated with the Data Science Institute (DSI) which is part of the American College of Radiology. Since medical imaging constitutes 85% of the venture capital (VC) funding for digital health, the authors assumed that as VC funding grows, artificial intelligence (AI) products will increase in tandem. Their objective was to “project the number of new AI products given the statistical association between historical funding and FDA-approved products.”

The study drew on data from two sources: the DSI’s own database on the number of AI products approved by the FDA between 2008 and 2022; and data on VC funding for AI projects between 2013 and 2022 was supplied by Rock Health, a seed fund that supports start-ups working in digital health. They then employed a linear regression model, based on a 6-year lag between funding and FDA approval, to forecast the number of new approvals.

The figure seen left (© American College of Radiology) illustrates the projected FDA-approved products based on actual and projected AI funding, modelled as a 6-year lag between funding and project approval. For example, “2028 product-year funding reflects actual funding in 2022 lagged 6 years to correspond with the average from funding to FDA approval.” The light grey area on either side of the blue line represents the 50% prediction interval, while the dark grey area encompasses the 95% prediction interval.

The results showed there are 11.3 new AI products for every $1 billion in VC investment, assuming the 6-year lag. The were 69 new AI products approved by the FDA in 2022 and this was associated with almost $5 billion in VC funding. In the year 2035, when the funding for new AI products is projected to reach $31 billion, 350 new products will be approved by the FDA. The authors acknowledged that their study had some limitations.

First, they had limited historical data – just 4 years of FDA-approved products. Second, they relied on a linear model, and third, the model was based on association rather than a cause-and-effect relationship between VC funding and FDA-approvals. Fourth, they were unable to distinguish between products based on quality and impact, and finally, their models were based on assumptions that might not come to fruition, such as market fluctuations or mergers and acquisitions. That said, they concluded that AI is likely to change the practice of diagnostic radiology in the future and this in turn may lead to increased investment in AI projects.