By Prashanth Kini
Health care is a data-gathering powerhouse.
But that means nothing if the industry can’t effectively harness this fire hose of information into discrete, actionable insights to improve outcomes, eliminate waste and uncover new revenue streams. Emerging value-based models of care, and upcoming regulatory mandates, demand that health care organizations quickly transform from data-gathering caterpillars into agile analytics butterflies. Millions of lives and billions of dollars are at stake.
People deliver most of the value of health care. Clinician expertise, intuition, diligence and warmth are invaluable to the patient experience and health care outcomes. But the big data problem in health care is beyond the scope of the work of the human mind. Many health care organizations employ a patchwork of data and analytics technology solutions to bridge the gap between what clinicians do today and what they should do tomorrow to achieve better results for patients and the bottom line. This patchwork is marked by breakpoints, handoffs, widespread administrative burden and limited visibility into the systemic health of the organization. This approach is also marked by its long lag times in identifying areas for improvement and operationalizing those insights.
By contrast, machine intelligence applications offer an alternative that can simultaneously take on data analytics challenges across the continuum of care. Machine learning applications can examine complex and continuously growing data elements far faster and capture insights more comprehensively than traditional or homegrown analytics tools. The ability to meaningfully track clinical practice adherence to clinical pathways, predictively manage population health and optimize revenue cycle management will prove to be a model for innovative providers and integrated health care systems to follow. Applications of machine intelligence for home health and telemedicine also are promising.
The answer inside
Health systems need to transition quickly to value-based care. A lever for accelerating change already exists within every health care organization — and that lever is the organization’s own data. However, analytics involves a different set of skills, a different organizational mind-set and a different suite of technologies. The meaningful use of data requires more than light-lift SQL queries, dashboard software or marginal enhancements to flow charts. Most health care organizations are overwhelmed by the complexities of their data. There are simply not enough data scientists or analysts to make sense of the exponentially growing data sets within each organization.
Machine intelligence distills complexity, finds patterns within billions of data points and gives organizations data-driven insight into the best opportunities for improving the quality and cost of health care. Here are four examples of the power of machine intelligence to transform health care data into better, more efficient care:
Next: Clinical/surgical variation
1: Clinical/surgical variation
One of the best opportunities for using machine intelligence to improve care is the development of clinical pathways. By guiding clinicians to follow best practices through each step of care delivery, clinical pathways ensure that all patients receive consistent high-quality care at the lowest possible cost.
Variation is a necessary and natural element in most health care delivery because every patient is unique. A certain amount of informed variation is critical to driving innovation in clinical practices. However, variation caused by gaps in knowledge or lack of data-based evidence makes it difficult — if not impossible — for health systems to reduce costs, improve patient outcomes and quality of care, and minimize medical errors. Eliminating unnecessary variation is fundamental to achieving value-based care.
Health reform also has upped the ante. The Department of Health & Human Services recently announced an accelerated timeline for tying payments to quality and value. In addition, the Centers for Medicare & Medicaid Services (CMS) published its final rule for a mandatory bundled payment model for total knee and hip replacements. It can be expected that CMS will expand the list of procedures for which it mandates alternative payment models.
Since the early 1990s, hospitals and health systems have used clinical pathways as a way to reduce unwarranted variation. Sometimes branded as integrated care process models or collaborative care pathways, they all share the same basic structure: a multidisciplinary team of providers use peer-reviewed literature and patient population data to develop and validate best-practice protocols and guidance for specific conditions, treatments and outcomes.
The case of Mercy Health in St. Louis is a perfect example of machine learning in action. In 2014, Mercy tested a machine learning application to recreate and improve upon a clinical pathway for total knee replacement surgery. Drawing from Mercy’s integrated electronic medical record (EMR), the application grouped data from a highly complex series of events related to the procedure, discovered natural variations in clinical practice of the procedure across the health system and identified those flavors associated with the best outcomes.
It was then possible to adapt other methods from biology and signals processing, and incorporate extant evidence-based guidelines into the problem of determining an optimal way to perform the procedure — which drugs, tests, implants and other processes contribute to that optimal outcome. It also was possible to continue to monitor and measure adherence against these standardized pathways. Moreover, predictive machine learning methods can be used to examine the impact of potential pathway modifications to outcomes, enabling a variety of what-if scenarios.
What this analysis revealed was an unforeseen and groundbreaking care pathway for high-quality total knee replacement. The common denominator between all patients with the shortest length of stay (LOS) and best outcomes was administration of pregabalin — a drug generally prescribed for shingles. A group of four physicians had seen something in the medical literature that led them to believe that administering the drug prior to surgery would inhibit postoperative pain, reduce opiate usage and produce faster ambulation. It did. Mercy never would have discovered this best practice using traditional approaches. This single procedure was worth over $1 million per year for Mercy in direct costs.
Next: Revenue cycle management
2: Revenue cycle management
Machine learning can easily be applied outside the realm of patient care, for example, in claims denial management. Denials are one of the most persistent problems in the revenue cycle, with the health care system devoting tremendous time, resources and money to recoding and resubmitting hundreds of billions of dollars in denied or rejected claims.
The denial of a claim can be due to multiple and complex variables, including patient, procedure, location, doctor, sequencing or payer. This means that uncovering solutions can be scattershot and infrequent. As with clinical variation, query-based approaches are time- and resource-intensive, and often fail to target the root cause of denied claims.
Machine intelligence applications are designed to find the answers by detecting all of the relationships associated with the data. Whereas human investigations into denied claims are slow and inaccurate, machine learning is able to drill down and identify the characteristics of denied or rejected claims holistically and identify changes that can be proactively driven upstream into the claims preparation workflow and potentially into point-of-care guidance.
From a revenue cycle management perspective, the ability to understand, monitor and manage clinical variation for a variety of episodes of care across the care continuum enables health systems to have a clear line of sight to their performance against bundled payments and other value-based arrangements. They will now be able to make the necessary course corrections to minimize an end-of-year shock when payers reconcile performance against contracts.
Next: Population health
3: Population health
With the health care industry moving toward outcomes-based reimbursement, population health management initiatives are no longer optional. Machine learning applications enable hospitals to support data-driven population health management initiatives by enabling predictive patient population risk assessment, care gaps identification and prescriptive patient intervention guidance.
Hospitals rely on patient risk scores to help physicians achieve the best possible outcomes while optimizing the use of scarce resources. With machine intelligence, clinicians can assess patient risk more effectively than they could with a patient questionnaire assessing readmission risks, which can be highly subject to interpretation. Machine intelligence approaches can incorporate current risk scores alongside a wide variety of patient clinical, financial and socioeconomic data to assess patient subpopulation risk for future disease states, cost, utilization and other types of risk from the patients’ and health system’s perspective.
By using machine learning to support population health management initiatives, hospitals and health systems benefit from being able to better understand quality indicators and identify ways to improve care outcomes while managing financial risk against emerging value-based reimbursement schemes. Mt. Sinai recently used machine intelligence to analyze clinical and genomic data from a population of type 2 diabetes (T2D) patients. The unbiased machine intelligence analysis identified heretofore unknown three distinct subtypes of T2D patients with distinct clinical and genomic characteristics. This advanced population stratification capability will inform the design of precision treatment regimens.
Next: Patient monitoring and telehealth
4: Patient monitoring and telehealth
Health care providers are increasingly looking for ways to monitor the health of their patients in outpatient settings to determine points of intervention to help reduce readmissions. The increase in adoption of wearable devices provides more data for closely monitoring chronic conditions and planning timely interventions. Using machine intelligence, health care providers can monitor patients’ medication adherence and disease outcomes over time to determine the types of intervention that would be most beneficial.
Machine intelligence is also promising for applications in preventive care and telemedicine. For example, patients may use a nurse call service and have data uploaded from their smart-phones or smart sensors to analyze the best and most cost-effective next step in treatment. Patients managing chronic conditions can use smart platforms at home to interact and feed data to their care providers. A prominent nonprofit organization has successfully utilized machine intelligence to distinguish a control set of healthy patients from Parkinson’s patients based purely on smart-phone accelerometer and gyroscope sensor data. The organization was also able to identify two distinct sub-cohorts of Parkinson’s patients, again based on the gait patterns revealed in the sensor data.
The ability to detect such differences in disease states holds large promise for the ability to monitor and proactively manage patients at home, particularly with debilitating diseases like Parkinson’s where the patient may not always be able to provide the necessary updates on their health. The best use of artificial intelligence is to augment, amplify and guide human intelligence. In health care, that means better best practices, sooner. Machine learning tools can deliver faster insights into which processes are working well for which patients and which ones need to be optimized. These distinctions are essential to deliver on the promise of value-based health care.
About the author: Prashanth Kini is vice president and head of product, Healthcare, for Ayasdi, a developer of machine intelligent applications for health systems and payer organizations.