Analytics in 2016: Big advances coming to healthcare
Healthcare’s awash in data, but new opportunities emerge to gain value, says Tom Hill
Information technologies are again at an inflection point, and 2016 will see many new ideas and technologies convert into useful innovation. My colleague and Dell Statistica's Chief Research Officer, Shawn Rogers, made several predictions recently regarding the use of and application of analytics in 2016. These predictions are very applicable to the healthcare domain.
Technologies enabling the efficient collection and storage of data are driving increasingly faster (high-velocity) and bigger (high-dimensional) data volumes. It is now possible to monitor patients around the clock, inside or outside of the hospital, to verify compliance with treatment protocols, risk, progress towards better health, and so on. This year may well be the year of the Internet of Things (IoT), where such applications enabled through real-time data collection become commonly used, commercially successful and possibly indispensable for modern healthcare.
IoT technologies will open up tremendous opportunities for more efficient and complete health, activity, equipment and other types of monitoring. However, it is important to remember that data volumes will continue to grow exponentially, but information will not. For example, honest self-reported information of the amount of daily exercise will likely correlate highly with minute-by-minute recordings of physical activity using a wearable device. Also, both can be deliberately manipulated.
The availability of new types of data promises to deliver new insights into best practices for healthcare. Analyzing infinitely large or growing amounts of streaming IoT data is not easy, and sometimes may not deliver new information and insights. For any given analytic or prediction problem, it is important to think through where the most valuable and relevant information can most likely be found. Understanding the specific problem domain should always inform data collection scope and strategies, as well as final implementation details.
The availability of new types of data promises to deliver new insights into best practices for healthcare, calibrated through personalized medicine to manage the specific risks for specific patients. There are really two ways to derive actionable insights from data: traditional hypothesis testing and predictive modeling.
In the traditional hypothesis testing approach, where actionable predictors—those that can be tweaked to drive better health outcomes—that are suspected to be important for some outcomes are used for building transparent and interpretable models. For example, logistic linear regression models are very popular for risk modeling.
The other approach is to include all available data into a predictive modeling project and rely on general deep-learning algorithms to identify repeated patterns in historical data. Those recurrent patterns then can be used to drive action, even if the nature and causes of the relationships are not understood.
A couple of things are worth noting here: While deep learning methods are becoming increasingly popular and are showing enormous promise to transform and disrupt various industries (such as self-driving cars, which are on track to change personal transportation), our current model of health and healthcare is still based on an understanding of causes and effects.
The old adage that “correlation does not prove causation” applies, and even though, for example, adults with tattoos appear to be more likely to take risks of various kinds, it is not reasonable to expect that that forcibly removing tattoos will reduce risky behaviors, nor will forcing a tattoo onto a non-tatoo’ed person increase subsequent risky behavior.
The point is that black-box models can generate interesting, useful, and sometimes transformative insights. But the standard for prediction models that potentially affect patient well-being will probably always include careful review, repeated testing and understanding of relationships found in historical data.
Almost certainly, regulatory recommendations for best practices will soon emerge to ensure the quality, fairness, and transparency of predictive models.
We are undoubtedly at an inflection point where technologies based on or enabled by advanced predictive analytics methods will transform healthcare. There are a now a large number of examples how this technology can significantly improve patient health outcomes, reduce cost and more.
I had the opportunity to participate in many such projects, and my colleagues and I will present two of these projects in detail at the upcoming HIMSS16 conference in Las Vegas. We are looking forward to discussing the exciting opportunities unfolding before us.
Technologies enabling the efficient collection and storage of data are driving increasingly faster (high-velocity) and bigger (high-dimensional) data volumes. It is now possible to monitor patients around the clock, inside or outside of the hospital, to verify compliance with treatment protocols, risk, progress towards better health, and so on. This year may well be the year of the Internet of Things (IoT), where such applications enabled through real-time data collection become commonly used, commercially successful and possibly indispensable for modern healthcare.
IoT technologies will open up tremendous opportunities for more efficient and complete health, activity, equipment and other types of monitoring. However, it is important to remember that data volumes will continue to grow exponentially, but information will not. For example, honest self-reported information of the amount of daily exercise will likely correlate highly with minute-by-minute recordings of physical activity using a wearable device. Also, both can be deliberately manipulated.
The availability of new types of data promises to deliver new insights into best practices for healthcare, calibrated through personalized medicine to manage the specific risks for specific patients. There are really two ways to derive actionable insights from data: traditional hypothesis testing and predictive modeling.
In the traditional hypothesis testing approach, where actionable predictors—those that can be tweaked to drive better health outcomes—that are suspected to be important for some outcomes are used for building transparent and interpretable models. For example, logistic linear regression models are very popular for risk modeling.
The other approach is to include all available data into a predictive modeling project and rely on general deep-learning algorithms to identify repeated patterns in historical data. Those recurrent patterns then can be used to drive action, even if the nature and causes of the relationships are not understood.
A couple of things are worth noting here: While deep learning methods are becoming increasingly popular and are showing enormous promise to transform and disrupt various industries (such as self-driving cars, which are on track to change personal transportation), our current model of health and healthcare is still based on an understanding of causes and effects.
The old adage that “correlation does not prove causation” applies, and even though, for example, adults with tattoos appear to be more likely to take risks of various kinds, it is not reasonable to expect that that forcibly removing tattoos will reduce risky behaviors, nor will forcing a tattoo onto a non-tatoo’ed person increase subsequent risky behavior.
The point is that black-box models can generate interesting, useful, and sometimes transformative insights. But the standard for prediction models that potentially affect patient well-being will probably always include careful review, repeated testing and understanding of relationships found in historical data.
Almost certainly, regulatory recommendations for best practices will soon emerge to ensure the quality, fairness, and transparency of predictive models.
We are undoubtedly at an inflection point where technologies based on or enabled by advanced predictive analytics methods will transform healthcare. There are a now a large number of examples how this technology can significantly improve patient health outcomes, reduce cost and more.
I had the opportunity to participate in many such projects, and my colleagues and I will present two of these projects in detail at the upcoming HIMSS16 conference in Las Vegas. We are looking forward to discussing the exciting opportunities unfolding before us.
More for you
Loading data for hdm_tax_topic #better-outcomes...