It’s Like Uber, but for Neurologists
Automation is one of the engines of modernity, and what it should or could be is one of our society’s central discussions. However, when we discuss automation, it is never as a change that affects everyone in our community, but instead as one targeted at certain groups. Manufacturing workers on the assembly line have been replaced to an ever-greater extent over the past few decades, and a fairly broad consensus agrees that long-haul truckers are among the next in line. Few think of jobs considered elite, specialized, or knowledge-based, like doctors and financiers (and, yes, academics too), to be automation’s next potential targets, though. Many would find the possibility surprising, or even dismiss the prospect out of hand as unrealistic.
What if I told you one of neurologists’ central tasks could soon be automated? McGill University’s Translational Neuroimaging Laboratory has developed a novel algorithm to predict which patients with mild cognitive impairment (MCI) will progress to Alzheimer’s disease (AD) within the next two years . MCI is a condition characterized primarily by a decline in memory, but also potentially attention and/or language as well, far beyond those normally associated with aging. The algorithm requires only a series of PET scans, which use safe injected radioactive particles to measure relative activity within various regions of the brain.
This algorithm is neither the first to predict progression from MCI to AD, nor the first on this timescale or that uses PET scans, but it is the most powerful so far, predicting progression correctly 84% of the time. Around 7 million Americans over 65 have MCI, and just under a third of them will go on to develop AD (Roberts and Knopman, 2013). In a disease as widespread and costly as AD, any real increase in predictive power could result in more accurate diagnosis for tens of thousands of people while allocating millions of dollars in health care resources each year more efficiently (e.g. reducing amount of AD drugs wrongly prescribed for truly non-progressive MCI).
So what makes this algorithm more accurate than the others before it? The accuracy of a predictive algorithm depends on a balance of two opposing aspects: the sensitivity and specificity. Sensitivity is the proportion of true positive results marked as positive, i.e. correctly identifying those with MCI that will progress to AD. Specificity is the opposite, the proportion of true negative results marked as negative, correctly identifying those with MCI that will not progress to AD. This new algorithm is not remarkable for its sensitivity, but for its specificity, which is the highest yet measured; in other words, while it does not necessarily find cases of future AD that other tests cannot, it gains its high accuracy by avoiding wrongly predicting progression to AD.
Could we soon see this algorithm replace neurologists in local hospitals screening patients diagnosed with MCI to identify AD early? Not likely, as this technology’s applications are a bit more limited than some would expect. A prediction of whether or not a person with MCI will progress to AD cannot take the place of a formal diagnosis of dementia, which requires the person actually develop the corresponding symptoms first. Its greatest utility would likely be in clinical trials for future AD treatments, not in the clinic itself.
Unfortunately, AD-related brain changes precede the onset of MCI by years . By the time MCI is first noticed, a lot of irreversible damage has already been done, which severely limits the algorithm’s potential clinical utility. Current treatments simply slow the progression of symptoms, which limits their (and by extension, the algorithm’s) usefulness to palliation alone. However, palliation could potentially begin earlier, extending the patient’s window of time with before their symptoms noticeably progress.
The best current use for the algorithm is in clinical trials, not treatment. The algorithm could be of great use in clinical trials pushing treatment earlier. Using the algorithm, researchers could more reliably identify candidates for treatment. Researchers could use the algorithm as a screen, more reliably identifying progressive- and non-progressive cases of MCI. With greater accuracy, these researchers could be more confident that their treatment is itself responsible for observed differences, not that they accidentally chose a high proportion of MCI cases that would never have progressed to AD regardless.
Using algorithms to predict future AD in patients with MCI is certainly a promising prospect, and its continually increasing accuracy is certainly heartening, especially for future drug development. However, those who claim this specific algorithm will revolutionize clinical AD diagnosis may be getting a bit ahead of themselves. Brain scans could someday screen patients for AD treatment, but we will continue to need neurologists. For now.
- Mathotaarachchi, S., Pascoal, T.A., Shin, M., Benedet, A.L., Kang, M.S., Beaudry, T., Fonov, V.S., Gauthier, S., and Rosa-Neto, P. (2017). Identifying incipient dementia individuals using machine learning and amyloid imaging. Neurobiol Aging. doi:http://dx.doi.org/10.1016/j.neurobiolaging.2017.06.027
- Roberts, R., and Knopman, D.S. (2013). Classification and epidemiology of MCI. Clinics in geriatric medicine 29, 753-772. doi:10.1016/j.cger.2013.07.003.
- Burggren, A., and Brown, J. (2014). Imaging markers of structural and functional brain changes that precede cognitive symptoms in risk for Alzheimer’s disease. Brain imaging and behavior 8, 251-261. doi:10.1007/s11682-013-9278-4.