Scott Brimble, MD, FRCP is an Associate Professor of Medicine and Divisional Director of Nephrology at McMaster University, Hamilton, Canada. Ironically, he is also the Provincial Lead for Early Detection and Prevention of Progression (of CKD) at the Ontario Renal Network. He reviewed chapter 13 for the NephJC bookclub.
“Eventually, we won’t need the doctor. Machine learning makes a better Dr. House than Dr. House.”
As one who never thought much of Dr. House, I am not sure that is really saying much.
I should also point out that it is a quote from Vinod Khosla, the venture capitalist who happens to want to charge a community $30 million to use the road on his property to access the beach.
In this chapter, the author discusses the importance that data science will have on prevention of chronic illness. I would encourage you to read the timeless editorial by the brilliant and, sadly late, David Sackett. Rather unfashionably, he argued that preventative medicine displays all three elements of arrogance; it is:
- aggressively assertive
- presumptuous
- overbearing.
As he argues, if we are to go after the “unsuspecting healthy”, it must be based on the highest level of evidence. At the time he was referring to the practice of prescribing hormonal replacement therapy in post-menopausal women to prevent cardiovascular disease, but there are many such examples in medicine. This line of reasoning reminds me why I tend to have little regard for much of what those who practice alternative medicine have to say.
Topol describes the prediction of disease both at a population level and on an individual level. A relatively current-day example of the former is Google Flu Trends. Started in 2008 – an algorithm tracks the use of flu-related search terms and was reported to better predict flu outbreaks than standard methodology. Subsequent work has shown that this approach consistently over-estimates flu prevalence – something that isn’t really surprising (everyone thinks they have the flu...). He then goes on to describe the example of the HealthMap algorithm that “predicted” the 2014 Ebola outbreak in West Africa, 9 days before the WHO made their announcement. Of course this algorithm did not “predict” an outbreak – it detected evidence that there might be an outbreak of a disease associated with hemorrhagic fever in West Africa, based on the blogging of some workers in the area who were concerned it may indeed be Ebola. Subsequent work confirmed it to be the case. It should be noted that the same algorithm also “predicted” an outbreak of Ebola in New York City, which, of course did not occur.
I mostly agree with Topol that the point of predicting a disease is so that we can then prevent it. He cites the example of Alzheimer’s disease, and efforts to better predict who will be afflicted with the debilitating condition. In the absence of preventative strategies in those at high risk of developing Alzheimer’s, there is questionable benefit to predicting the disease at an earlier stage. As Sackett reminds us, it is arrogant to presume that in the absence of high-quality evidence, employing a strategy to prevent a disease is best for a patient; indeed it may in fact turn out to be harmful.
As a nephrologist, when reading this section of the chapter, I was reminded of Nav Tangri’s (@NavTangri) work on developing the Kidney Failure Risk equation (KFRE). This equation uses only four variables: age, sex, eGFR, and urinary albumin-creatinine ratio (ACR) to predict the 2-yr and 5-yr risk of end-stage renal disease (ESRD). Validated in over 700,000 subjects (big data?!), this immediately improved our ability to predict the likelihood of what we all consider a bad outcome. Can we prevent ESRD, armed with such information? Probably not, most of us are already well aware of what treatment options are available to treat patients with chronic kidney disease (and there are few), so such information is unlikely to influence treatment strategies. Nevertheless, the information can be useful to better allocate more intensive resources (eg. ESRD preparation, ongoing nephrology follow-up) and to counsel patients about their risk. More often than not, this tool helps to reassure patients that in fact their risk is low.
In Topol’s ongoing fascination with wearable sensor technology, it is not surprising that the chapter moves on to their potential usefulness in predicting disease. He discusses a number of conditions, including asthma, post-traumatic stress disorder, and specifically in congestive heart failure, where sophisticated wearable sensors and smartphones could be used to measure vital signs, apneic episodes, weight, as well as endogenous biomarkers such as brain natriuretic peptide and creatinine. In aggregate, such information might be used to predict an episode of heart failure prior to onset of symptoms, leading to changes in therapy to prevent adverse outcomes including hospitalization. Trials using implantable technology with telemonitoring have indeed already been conducted and more are likely to follow. Whether such technology ultimately is better than structured telephone monitoring without the costly technology is unclear. He then goes on to discuss more futuristic technologies, including a tiny sensor implanted in the bloodstream to monitor the immune system to, for example, predict a flare-up of an autoimmune disease. Such approaches would ultimately require machine learning, it is argued, to collect such complex information from thousands of afflicted patients to better understand patterns that predict onset of disease.
Topol argues that in the current era, for many conditions including cancer, the goal of physicians is to convert the condition into a manageable, chronic disease, rather than finding a cure. This is indeed true for the most part. However, the examples of wearable technology coupled with machine learning cited throughout the chapter focus on predicting flare-ups of a chronic disease (eg. heart failure, asthma exacerbation, autoimmune disease flare-up), and are not about predicting and preventing the chronic disease in the first place (eg. a cure). It seems unlikely that for many of the chronic illnesses (eg. diabetes mellitus, hypertension, coronary artery disease, chronic kidney disease etc.), predicting future risk will lead to preempting the disease. When a primary care physician is reviewing a patient in the office who is overweight, sedentary, with a strong family history of diabetes mellitus and hypertension, he or she does not need to consult IBM Watson or download information from an implantable sensor to predict that the patient will likely develop diabetes mellitus, hypertension, with their associated complications. At the end of the day, prevention will require active patient participation with a focus on lifestyle changes. If wearable technology is a motivator of sorts to achieve such changes, and is cost-effective, then we should be all for it. Indeed, at least one randomized controlled trial has evaluated such an approach, but was limited by a 40% attrition rate, which is telling in itself.
My own biased view is that we have more immediate work to do as a community. We live in a day and age when, despite advances in technologies and therapies and a wealth of information on risk factors for the development of a variety of diseases, significant proportions of the population in many countries, including Topol’s own, do not have affordable access to healthy food choices, basic health care and the proven therapies to treat common chronic conditions. I think we need to get our house in order to improve the social determinants of health as a strategy to prevent disease before we get too carried away with more sophisticated, costly and largely unproven strategies, as fascinating as they may be.