Doing Predictive Analytics? Maybe You're Doing It Wrong

Sriram Vishwanath is a professor of Data Science and Informatics at the Cockrell School of Engineering, The University of Texas at Austin, offering unique and profound insights into the application of predictive analytics in healthcare.

Figure: Sriram Vishwanath

First, academic perspectives on this issue are divided into two major camps. One camp argues that any attempt to apply predictive analytics to healthcare is futile. Healthcare is an extremely complex field, and most people have an inadequate understanding of it. Given the sheer number of variables, achieving accurate predictions is unrealistic; at best, such efforts can only be considered descriptive studies.The other camp, namely the data science community, holds a highly optimistic view, believing that they can predict everything. No matter what challenges arise, they are confident in their ability to overcome them successfully.In Vishwanath’s view, both of these perspectives are incorrect.He considers the former too pessimistic. Due to a lack of proper understanding of data science, this group resists challenges and fails to recognize that data science, when applied appropriately, can generate positive benefits for the healthcare industry. The latter, meanwhile, is overly confident, failing to appreciate the magnitude of the task at hand.Striking a balance between these two attitudes may yield surprising results, namely:Do not attempt to predict everything; instead, focus your attention on what is within your control.For those looking to harness the power of big data, this is perhaps the most fundamental piece of advice. The “Big Data and Healthcare Analytics Forum” was held in Boston, United States, this past November. Drawing on his extensive experience accumulated over many years in the fields of data science and engineering, Vishwanath offered practical strategies and recommendations to healthcare organizations seeking to effectively leverage high-precision analytics.As value-based and outcomes-based reimbursement models become increasingly prevalent, the boundary between consumers and healthcare providers is becoming increasingly blurred. Although clinical risk and financial risk are two distinct concepts, they are often considered together in predictive modeling, much like two sides of the same coin.Under requirements set by the Centers for Medicare & Medicaid Services (CMS), healthcare providers are also beginning to assume financial risk. Therefore, it will not be long before distinguishing between clinical risk and financial risk becomes largely unnecessary. As for health insurance, it has always been a high-risk industry; however, since insurers began acquiring healthcare service providers, they too have started to bear both clinical and financial risks.Vishwanath has explored the field of data analysis for many years: he served as a professor at the University of Texas for over a decade and more recently assumed the role of CEO at the healthcare startup Accordion Health. These experiences have led him to understandWhat Are the Unique Challenges Facing the Healthcare Industry: Pain Points from a Data Science PerspectiveFrom clinical and financial perspectives, what understanding and operational approaches are required for predictive analytics to truly generate positive impact and efficacy in the healthcare sector?He has been engaged in predictive analytics for many years, spanning network environments, network traffic, consumer analysis, efforts to understand consumer behavioral characteristics, and research into how consumer behavior trends evolve over time. However, when he shifted his focus specifically to the healthcare industry, one point stood out to him in particular:Compared to the past, the industry is now changing at a much faster pace. The advent of the digital era has enabled people to access vast amounts of data that were simply nonexistent in the paper-based age.Finally, the healthcare industry is poised to achieve a dual leap forward in quality improvement and cost control.Nevertheless, Vishwanath has noted some unfavorable factors: much like “artificial intelligence” and “synergy,” the term “predictive analytics” has been overused and reduced to a buzzword. Those who claim to engage in predictive analytics are often far from practicing true predictive analytics. Their work might be better described as summarization, interpretation, or extrapolation, but not prediction.Vishwanath likes to use a cartoon to illustrate the pitfalls of conducting predictive analytics with unreliable methods. The cartoon roughly depicts the following scenario: The protagonists are a newlywed couple, with a chart in the background where the horizontal axis represents time and the vertical axis represents the number of husbands the female protagonist has had. Based on actual conditions, the male protagonist plots two coordinate points: (yesterday, 0) and (today, 1). He then solemnly derives a prediction curve from these two points and seriously predicts that, by next month, the female protagonist will have dozens of husbands.

Clearly, this approach is unworkable.Recently, the healthcare industry has made significant strides in predictive analytics. These advancements are particularly noteworthy given that the healthcare sector lags behind other industries in this regard.Vishwanath noted that data scientists prefer working with consumer data because it is far more accessible than healthcare-related data. From each consumer, approximately 100,000 data points can be collected, yielding an unimaginably vast amount of information. The situation in healthcare is quite different: data volume poses a major challenge. Typically, only three to four data points can be obtained from each individual—perhaps ten to twelve at best. However, such limited data quickly proves of little practical value, as it has been extensively utilized by others and thus lacks predictive power.For all those seeking to apply predictive analytics in healthcare, the most challenging aspect is generating accurate predictions despite limited data availability. While countless curves can be fitted to just two or three data points, determining which one is truly valid remains difficult.It is precisely for this reason that Vishwanath emphasizes focusing on what is feasible within existing constraints.In the healthcare industry, the right approach is to focus on your strengths and learn to say no to areas where you are less proficient.For instance, predicting cardiac arrest is extremely difficult. Do not assume that you can provide high-precision predictive data, as this is impossible. In contrast, predicting the outcomes of knee replacement surgery is far less challenging, because the underlying influencing factors and timelines are relatively easier to manage.Regarding population health management, Vishwanath highlighted two key issues: First, too many healthcare institutions treat the entire population as their study subject, whereas only approximately 3% of the population is truly critical. Focus should be placed on this crucial 3%, with efforts directed toward properly managing their data, since what we need are not average values but outliers. Second, most people overlook the impact of physicians and allied health professionals on predictive data. The consequence is that even if you target the minority of outliers, you are at best only half right, because you have failed to account for the influence of physicians and allied health support. In addition, factors such as zip code, ethnicity, and age are also highly significant.Therefore, for healthcare institutions seeking to effectively leverage predictive analytics for population health management, Vishwanath offers two recommendations:First, focus on the critical 3%, identify who these individuals are, and obtain as precise information about them as possible. Second, while taking into account patients’ individual conditions, place emphasis on physician-related and healthcare support factors. Only in this way can predictive analytics in the healthcare industry be effectively conducted.。Compiled by Lü XiaoyiEditor: Mo Renying