Uses of Artificial Intelligence in Measuring the Impact of Research

To get to an answer, we can assess how AI has impacted medical research, what strides have been made in converting research & development into commercialized AI-based technology. We can then begin to truly evaluate whether in another 20 years, AI might displace physicians or more benignly, serve as their personal assistants.

AI, the basics

AI is essentially a branch of computer science whereby data are collected and algorithms created based on patterns found across the data using deep machine learning or neural networks. The output may be a diagnostic, prognostic or disease prediction that appears as if a human had analyzed the data and determined the output, all at a fraction of the time it would take a human to complete. For AI to be reliable, it is absolutely critical that the data numbers be high and of sufficient breath and quality to avoid skewed, biased results that are not generalizable (Sinz, Pitkow, Reimer, Bethge, & Tolias, 2019). Otherwise, we’re left with garbage in, garbage out. The AI field has matured wherein the scope and quality of training data, augmentation of data and enhanced computational power have resulted in ever more precise output, enabling processes that are particularly repetitive, ripe for AI (Gardezi, Elazab, Lei, & Wang, 2019).

AI in medical research

Several areas of medicine have been particularly amenable to AI based on the sheer volume of data readily available: radiology, ophthalmology and pathology (Ahuja, 2019; Gardezi et al., 2019). The data are derived from the vast numbers of patient-derived images and recordings that these medical segments collect to make diagnoses: from X-rays to CT scans, MRI imaging, retinal imaging and tissue histology images.

The field of opthalmology has pioneered the way with the first-ever FDA approved AI-medical device that was granted to IDx based on the autonomous analysis of 900 patient retinal images. Their diagnostic, the IDx-DR retinal imaging device, can detect higher than mild levels of diabetic retinopathy (DR) in diabetic patients with an accuracy of 87.4% (Abràmoff, Lavin, Birch, Shah, & Folk, 2018). It is the first authorized medical device screening tool that does not require a clinician to interpret retinal images for (DR) (FDA, 2018).

Another compelling example is AI applications in radiology, specifically breast mammography diagnostics. Based on 100,000 breast mammogram images, Google’s health research arm very recently announced that their AI-trained software resulted in 5.7% fewer false positive and 9.4% fewer false negative rates than trained radiologists (Collins, 2020). While their AI-software has not been approved yet by the FDA for diagnostic purposes, the results are proving out the revolutionary impact of AI in medical research.

AI in pathology has also made strides on the research and development front, particularly in cancer diagnostics given its extensive dependence on (digitized) tissue morphology. The push for some form of automatized assistance comes partially from interobserver (aka pathologist) variability in the analysis of H&E stained tissue and the sheer volume of images (Harbias, Salmo, & Crump, 2017)Advances in AI-research by Philips led the FDA to grant approval of their IntelliSite Pathology Solution, the first ever whole slide review imaging system to be marketed (FDA, 2017). While pathologists are still required to review and interpret the images, they can do so from digitized images rather than tissue samples.