What is necessary to transform analytics tools into enterprise-level solutions?
With the steady movement towards shared risk and shared savings models in healthcare, the pressure on risk-bearing organizations to reduce the costs of care has only heightened the need for next-generation healthcare analytics tools. While industries such as financial services, retail and law enforcement have pioneered the latest in healthcare analytics tools, healthcare has been (predictably) slow to fully embrace these new technologies. But what exactly do we mean by next-generation analytics tools? More precisely, what is necessary to transform sophisticated, yet often narrowly focused tools, into full enterprise-level analytics solutions?
Coming on the heels of major investments in big data coupled with recent advances in computer science, large healthcare organizations have directed their attention toward AI in the form of machine learning to leverage their available data to increase efficiency and competitiveness. These advancements are making it possible to develop predictive models that increase the speed and accuracy of complex diagnoses, reduce readmission rates and focus expensive resources more appropriately. According to Accenture, 69% of healthcare executives report increasing investments in artificial intelligence technologies and machine learning applications compared to last year.
However, these investments have yet to be accompanied by the expected (or promised) ROI, arguably because they overlook an important point: at an enterprise level, next-generation analytics tools must be complemented by a comprehensive suite of additional technical capabilities and services, not just one or two good models. In order to generate the kind of transformative changes that organizations are looking to achieve, predictive models must be effectively surrounded by robust data integration, clinical breadth, a meaningful time horizon, and ultimately, transparency that delivers confidence and maximizes the likelihood of end user action.
Machine learning is most effective when all available and relevant data can be utilized to train predictive models. Unfortunately, many organizations still face a gap in the capabilities required to unify healthcare’s highly complex data – claims, EHR/EMR, labs, diagnostic imaging, behavioral patterns – into a comprehensive longitudinal record for each health plan member or IDN patient. Moreover, data integration must also address the fundamental issue of scale and efficiency. Data analysts can spend up to 90 percent of their time managing data rather than analyzing it. First and foremost, next-generation solutions must automate the integration of structured and unstructured data from a variety of sources into a readily consumable, standardized format for machine learning algorithms to consume. And, because healthcare’s data suffers from the curse of high dimensionality, once the data is cleaned and integrated, advanced feature extraction, selection and engineering becomes paramount to making models more accurate.
A host of data scientists from academia and private enterprise have spent years developing predictive models for diabetes. Given the prevalence and associated costs of diabetes in society, it’s not surprising that so much attention has been focused on this single condition. While having a ‘good diabetes model’ may be useful for some questions, it falls short of addressing the needs of the enterprise. At a minimum, a next-generation analytics solution must address the top 8-10 conditions that drive 80% or more of healthcare costs. The co-morbid nature of these conditions create a level of complexity for the predictive models that quickly overwhelm the capabilities of the singe model approach.
Meaningful Predictive Time Horizon
Keep in mind, the whole point of predicting clinical conditions or events is to be able to intervene early enough to avoid the incident, blunt the severity or the incident or defer the onset to a significantly later time. Predicting that someone is likely to acquire diabetes or CHF next week is interesting, but not terribly useful from a preventative clinical standpoint. Predictive models need to perform reasonably well in the 12- to 24-month horizon to allow providers and case managers the opportunity for meaningful intervention. It’s worth noting, that achieving this feat is largely dependent on the first item – data integration. Absent the ability to fully integrate different data types, it is highly unlikely that a single data set such as claims or labs will provide the necessary information to drive early predictions.
Perhaps one of the major challenges with predictive analytics solutions in healthcare today is that most algorithms are black boxes. While accuracy and interpretability of models can often be a tradeoff, delivering transparent analytics in healthcare is crucial to driving greater collaboration between payers and providers. There are a number of techniques which can be used to improve accuracy while retaining interpretability of models.
In today’s rapidly evolving environment, organizations need a robust solution that addresses the complex needs of enterprise-level businesses, and provides the agility needed to not just succeed, but thrive by constantly staying ahead of the curve. This requires more than building or acquiring one good model. While I have listed some of the characteristics above, there are others that need to be considered as well: robust infrastructure that can handle both volume and security needs; model monitoring and management after deployment; agility in retraining or rebuilding models, and; ability to integrate results into end user workflows. Health organizations, particularly risk-bearing organizations under increasing pressure to transition to new business models, should challenge themselves to think beyond the single-purpose model and instead think more holistically when it comes to enterprise analytics.