Predictive analytics in healthcare must be timely, role-specific, and actionable to be successful. There are also three common types of healthcare predictive analytics: Risk scores (risk stratification using CMS-HCC or other models), What-if scenarios (simulations of specific outcomes given a certain combination of events, and Geo-spatial analytics (mapping a geographical location’s patient disease burden). The common thread in all of these is the element of action, or specifically, the intervention that really matters in healthcare predictive analytics.
Machine Learning / Predictive Analytics
Under value-based healthcare and the 2012 Hospital Readmission Reduction Program, healthcare organizations are more motivated than ever to reduce their incidence of preventable readmissions.
Health systems can reduce risk of hospital readmissions by developing readmission risk scores tailored specifically to their populations. A risk model that meets the following five requirements will have significant predictive value and is most likely to achieve systemwide adoption:
Identifies at-risk patients early.
Separates patients relevant to the disease-specific identification method and intervention strategy from all other in-hospital patients.
Uses organization-specific data to train a disease-specific model.
Exceeds performance of existing models.
Is developed in collaboration with domain experts.
Machine Learning in Healthcare: How it Supports Clinician Decisions—and Why Clinicians are Still in Charge
Machine learning in healthcare is transforming healthcare with its ability to tackle data variability and complexity. Everyone in healthcare should embrace this new technology and its ability to deliver more precise, faster, data-driven insight to clinical teams.
But just as machine learning has benefits, it also has limitations; for example, it loses its impact when implemented without realistic expectations or without thorough integration with existing clinical processes.
As the FDA works to publish guidance on digital health services, including governance regarding the use of algorithms to support clinical decisions, it’s important for everyone in the industry to hold themselves accountable for the quality of the data and the processes that put this data in front of clinicians.
Machine learning is transforming the way health systems deliver care to patients by surfacing insights to clinicians at the point of care; but, ultimately, the clinician considers the entire clinical picture to determine the most appropriate plan for patients.
Machine learning, used in the context of healthcare, is not about computers replacing doctors or rolling robots dispensing bedside care to patients. Perhaps a better term would be data-driven healthcare because it is the process of using historical patient data in a predictive model to determine the likelihood of a healthcare-related outcome. The black box that is the machine is no more than an algorithm trained by data. Most importantly, the predictions can be used by doctors to optimize decision making in real time, thus reducing readmissions, infection rates, and other complications that drive up costs and lower the quality of care.
This article explains some machine learning basics, dispels some misconceptions, and outlines five steps to its implementation:
Define the use case.
Prepare the data.
Train the model.
Make predictions on new data.
Deliver the risk score for use in clinical decision support.
Machine learning is destined to be a digital partner for physicians, executives, and health systems focused on improving clinical, financial, and operational performance. To get there, it must first be understood.
The Adverse Childhood Experience (ACE) study conducted by the CDC and Kaiser Permanent showed a strong correlation between ACEs and negative health outcomes later in life (e.g., risky health behaviors, chronic health conditions, and early death). ACE scores help paint a more complete picture of a person’s health history—a more comprehensive data snapshot of the entire patient.
Given that ACE scores build better data sets and machine learning relies on high-quality data, health systems should incorporate these nutrient-rich data sources into their machine learning models to better predict negative health outcomes, allow for earlier interventions, and improve outcomes.
Healthcare machine learning is evolving to use ACE scores and lifestyle data (e.g., eating habits) to improve population health management.
Machine learning is a term that crops up often in healthcare lately, but it’s important to understand what really constitutes learning in this context. What some call machine learning is actually unintentional programming, but true learning is derived during the process of building a predictive model. This article delves into the nuts and bolts of a healthcare machine learning model and describes the training process a model undergoes to impact outcomes for patients.
The key ingredient is data and the key deliverable is to complete the feedback loop so those responsible for managing care have actionable information at their disposal.
Machine learning is beyond conceptual; it’s incorporated into a growing list of predictive models for various disease classifications.
There are limitless opportunities for machine learning in healthcare. Defined as, “a field of computer science that uses pattern recognition to identify historical relationships in large data sets using an algorithm to create a generalized model of behavior,” machine learning is one of the most important life-saving technologies ever introduced to healthcare for several reasons:
Prevents hospital acquired infections (HAIs).
Reduces hospital length of stay (LOS).
Predicts chronic disease.
Reduces one-year mortality.
Improves sepsis outcomes.
It’s no wonder health systems are eager to start leveraging machine learning to save lives, improve outcomes, and make systemwide enhancements.
They can do so by understanding machine learning basics, the importance of customized machine learning (not one-size-fits-all models) and historical data (a requirement for answering basic machine learning questions), and how machine learning helps their patients, clinical teams, and bottom lines.
As data availability and open source tools make predictive analytics increasingly accessible for health systems, more organizations are adopting this advanced capability. Organizations won’t, however, use predictive analytics to its full potential—making it routine, pervasive, and actionable—without a deployment strategy that scales the technology.
Three recommendations can help health systems successfully deploy predictive analytics and leverage data experience to improve data-driven interventions and outcomes:
Fully leverage your analytics environment.
Standardize tools and methods using production quality code.
Deploy with a strategy for intervention.
Machine learning in healthcare is already proving its worth in clinical applications. From identifying tumors in mammograms, to diagnosing skin cancer and diabetic retinopathy from images, algorithms can perform certain duties more quickly and reliably than humans. While only used for specialized medicine now, the time will come where every practitioner and patient will benefit from cyber-assisted bedside care. This won’t develop without ethical implications, but the advantages that machine learning will bring to healthcare in terms of lower costs, improved quality of care, and greater provider and patient satisfaction, will easily outweigh those concerns.
In this article, Dr. Ed Corbett explores the intricacies of machine learning from two perspectives: as a physician and as a family caregiver with a personal story about how this data science could benefit patient lives today.
Healthcare machine learning, predictive analytics, and artificial intelligence (AI) are starting to play a much bigger role in care management.
As care managers continue to have a growing number of patients like Ruth, who use digital devices at home, machine learning offers a solution to the resulting exponential increase in healthcare data.
Defined as the practice of extracting information from existing data sets to determine patterns and predict future outcomes and trends, the advantages of using predictive analytics to improve care management are infinite, from chronic disease management to cost control.
Health systems must prioritize learning how to use healthcare machine learning to not only improve their care management programs, but also outcomes for patients like Ruth.
Despite its prevalence in many other industries (and its use by most Americans every single day), machine learning in healthcare is far behind. But not for long, because Health Catalyst® is bringing this life-saving technology to healthcare with catalyst.ai™—a new machine learning technology initiative that helps healthcare organizations of any size use predictive analytics to transform healthcare.
The clinical, operational, and financial opportunities catalyst.ai gives health systems are limitless:
Prevent hospital acquired infections.
Predict chronic disease.
Reduce hospital Length-of-Stay.
Catalyst.ai™ (machine learning models built into every Health Catalyst application) together with healthcare.ai™ (a collaborative, open source repository of standardized machine learning methodologies and production-quality code that makes it easy to deploy machine learning in any environment), represents a new era of powerful predictive analytics that will not only improve outcomes, but also save lives.
Before the introduction of healthcare.ai, an open source, healthcare-specific machine learning software, only a small subset of healthcare staff (primarily data scientists) had the ability to leverage predictive analytics to improve outcomes.
Healthcare.ai will democratize machine learning by empowering everyone in healthcare with the appropriate technical skills (BI developers, project managers, data architects, etc.) to download the healthcare.ai tools (packages for R and Python), request features, ask questions, and contribute code.
What sets healthcare.ai apart from other machine learning tools is its healthcare-specific functionality:
Pays attention to longitudinal questions.
Offers an easy way to do risk-adjusted comparisons.
Provides easy connections and deployment to databases.
Healthcare.ai will do more than just democratize machine learning—it will transform healthcare.
Retrospective and predictive analytics are familiar terms for practitioners of clinical outcomes improvement, but the new kid on the block is prospective analytics. This is the next level that uses findings from its predecessors to not only identify the best clinical routes, but also what the results might be of each choice. Prospective analytics gives bedside clinicians an expanded, branching view of operational and clinical options in a type of decision support that can lead to not only improving surgical and medical outcomes, but to making a positive financial contribution, as well. But, as expected with any new process or new way of thinking, prospective analytics requires careful introduction and stewardship to help drive its adoption within the organization.
We developed a predictive analytics framework for patient care based upon concepts from airline operations. Using the idea of an aircraft turnaround time where the airline wants to put the aircraft back into operation as soon as possible, we’ve created a way to help patients headed toward poor outcomes, along with their providers, “turnaround” and get the best possible, most cost-effective outcome. For example, in a diabetes patient, we might use variables such as: age, alcohol use, annual eye/foot exam, BMI, etc. to look for patterns that might influence two outcomes: 1) Diabetic control and 2) The absence of progression toward diabetic complications. The notion of our Patient Flight Path is useful at both the conceptual level, as well as the predictive algorithm implementation level.
As far back as the 1840s, clinicians have been using maps to inform them about population health trends. Today, the geo-analytics industry is well-developed in almost every application, with the exception of healthcare and medicine. There is potential to use mapping technologies to show patient disease burden in geographic form, map locations of health care facilities, and a plethora of accountable care population health initiatives would benefit from geo-analysis. Health Catalyst is working to integrate inputs into analysis like maps that can show geographic care boundaries, population health demographics, and more.
Dale Sanders, SVP at Health Catalyst, gave a presentation covering healthcare analytics strategy at the recent Plante Moran executive Healthcare Summit. He covered similarities between his former career in assessing nuclear threats for the US Air Force and NSA, and his role today in advising effective use of healthcare analytics. His session covered the Healthcare Analytics Adoption Model and how organizations need to take a systematic, strategic approach to implementing new software solutions. He also covered the importance of establishing a healthcare data acquisition plan, in light of the coming patient-driven sources of data, such as wearable devices.
This is the complete 4-part series demonstrating real-world examples of the power of data mining in healthcare. Effective data mining requires a three-system approach: the analytics system (including an EDW), the best practice system (and systematically applying evidence-based best practices to care delivery), and the adoption system (driving change management throughout the organization and implementing a dedicated team structure). Here, we also show organizations with successful data-mining-application in critical areas such as: tracking fee-for-service and value-based payer contracts, population health management initiatives involving primary care reporting, and reducing hospital readmissions. Having the data and tools to use data mining and predict trends is giving these health systems a big advantage.
Interest in predictive modeling is part of a larger trend to employ business and clinical intelligence applications in healthcare. Until recently, organizations that had the ability to mine and analyze data were mostly conducting retrospective analyses. Using tools available today, organizations with the right technical infrastructure, including a data warehouse, can link predictions to specific clinical priorities, set up new workflows, apply analytics to emergency departments and to slowly changing clinical situations and more.
Predictive analytics alone cannot offer meaning without context, especially in health care. In order to be successful, prediction tools should be content-drive and clinical-driven. Prescriptive analytics can improve health care better than simple predictions can. Analytics should be used with clinical leaders that have the willingness to act on appropriate intervention measure. This “in context” prediction should include not only the evidence, but also the interpretation and recommended actions for each predicted category or outcome. An underlying data warehouse platform is key to gathering rich data sets necessary for training and implementing predictors.
Those in Big Data and Healthcare Analytics circles will seldom hear the phrase “less is more.” In a clinical setting however, there is an important lesson to learn in regards to the effective execution of predictive analytics. We should not confuse more data with more insight. More data is simply more—as in more tables, more lists, more replicates, more clinics, more controls, more rows, tables of tables and lists of lists, etc. You get the idea. In short, for predictive analytics to be effective in a clinical venue, a specific focus will always trump global utility.
Predictive analytics is quite a popular current topic. Unfortunately, there are many potential side tracks or pit falls for those that do not approach this carefully. Fortunately for healthcare, there are numerous existing models from other industries that are very efficient at risk stratification in the realm of population management. David Crocket, PhD shares 4 key pitfalls to avoid for those beginning predictive analytics. These include 1) confusing data with insight, 2) confusing insight with value, 3) overestimating the ability to interpret the data, and 4) underestimating the challenge of implementation.
In healthcare, popular buzzwords and hot topics always come and go. Technically sexy topics such as big data, bioinformatics, predictive analytics or genomic medicine are tossed in and about sales calls, funding proposals, journal articles and blogs for a few years and then folks move on to the next big thing. The buzzword fever around predictive analytics will likely continue to rise and fall. Unfortunately, lacking the proper infrastructure, staffing and resource to act when something is predicted with high certainty to happen, we fall short of the full potential of harnessing historic trends and patterns in patient data. In other words, without the willpower for clinical intervention, any predictor – no matter how good – is not fully utilized.
3 Reasons Why Comparative Analytics, Predictive Analytics, and NLP Won’t Solve Healthcare’s Problems
I had a recent opportunity to engage in an online discussion with a well-known healthcare analytics vendor about the value of comparative analytics, predictive analytics, and natural language processing (NLP) in healthcare. This vendor was describing a beautiful new world of the future, in which comparative data, in particular, would be the cornerstone of our industry’s turnaround. The executive summary of my response: “Beware the smoke and mirrors” because 1) comparative data doesn’t drive improvement, 2) predictive analytics fails to include outcomes, 3) gaps in industry healthcare data limits the effectiveness of NLP.