How machine learning can drive clinical and operational improvements
As the healthcare industry pivots toward enhancing the delivery of care, improving the efficiency of their business operations and advancing the quality of scientific discovery, many organizations are increasingly embracing machine learning as part of their overall clinical and financial strategy.
Enthusiasm for advanced cognitive computing in healthcare is on the rise as providers recognize the need for analytical tools to gain predictive and preventative insights. By using machine learning algorithms, healthcare organizations can glean patterns in patient data and diagnostic imaging that will help them treat and diagnose patients with greater accuracy, map care pathways and processes, reduce costs in care, and improve outcomes. Yet, the task of implementing machine learning projects comes with challenges.
Machine learning projects can be expensive. They require highly skilled data scientists and IT capabilities that hospitals with thin budgets often can’t afford. These initiatives also require significant technology infrastructure and computing power.
Then there are data concerns. IT managers wonder how much data they’ll need to perform queries and whether too much data will bring them diminishing returns. They also must ask which data they should select and how the data should be collected. These questions and more need to be answered before embarking on a machine learning project.
Good analysis won’t result from bad data, but because clinical documents and medical images are far too large for the human mind to compute, machine learning projects are becoming more frequent, especially as population health and value-based care initiatives become increasingly critical.
Recent examples show that large hospitals, academic medical centers and technology companies are applying machine learning algorithms to a variety of use cases. Some of these include mapping cancerous immune-cell patterns that will help guide new cancer therapies; identifying heart blockages; and applying the technology to the data of discharged patients in an effort to identify which have the highest risk for readmission.
As the momentum to adopt machine learning grows, healthcare organizations will need to design a plan that takes advantage of the insights they gain as they look to customize treatments, improve diagnosis decision making and forecast the spread of infections. These goals will be key drivers of their operations, especially against the backdrop of the need to produce quality care metrics that enables them to be paid under value-based care payment programs.
To conduct a successful machine learning project many healthcare organizations are turning to third-party managed service providers that already have experience handling large volumes of patient data sets and can provide valuable offerings, such as patient-matching tools and cybersecurity technology. Engaging a managed service provider (MSP) that has a partnership with a reputable HIPAA-compliant cloud vendor can help the healthcare organization build, train and host their machine learning models at scale. Additionally, customized solutions with on-board computing power capable of running real time deep learning inference on sophisticated models helps deliver the performance, efficiency and responsiveness healthcare organizations want to see.
At the heart of machine learning—a subset of –is the ability to perform pattern recognition, probability theory, optimization and statistics. Machine learning algorithms can be trained to learn from the data, build a model to recognize common patterns, devise data-driven predictions and uncover insights that contribute to informed decisions.
One example of how machine learning can be applied in healthcare is the case of performing demographic matching of data for an Enterprise Master Patient Index (EMPI)—a centralized database containing patient medical records across various departments and geographic locations. Patients are assigned a unique identifier in the EMPI, but data that comes from multiple sources can have data input errors, name variances, duplications and other precarious inaccuracies.
Unlike traditional algorithms, machine learning algorithms are able to adjust themselves based on the feedback provided by human intervention. In the case of the EMPI and its primary goal of demographic matching, the training process for machine learning hinges on manual remediation typically performed by health information management (HIM) professionals responsible for reviewing and linking duplicate records together under a single identifier.
This manual intervention tends to occur in cases where there is ambiguity between two or more records, and the action performed represents an enormous amount of information that traditional algorithms simply discard. The challenge in using this kind of information is in the sheer number of human interactions required for an algorithm of this type to truly outperform human remediation. This is because the system must be able to detect broad patterns where users consistently take an action of marking a pair of records unique or as a match.
Training, however, is greatly simplified in a cloud environment where usage statistics across many implementations can be gathered to produce a highly intelligent record resolution algorithm, thereby reducing manual duplicate resolution tasks and diminishing false-positive/false-negative errors. Data centralization in the cloud is also cost effective because resources can be dynamically allocated to multiple customers on demand.
While healthcare organizations identify, harvest and normalize the data before starting a machine learning project, they’ll have to keep in mind their efforts will be used toward the greater goal of meeting specific performance metrics under their health insurers’ payment programs. For example, under the Medicare Access and CHIP Reauthorization Act of 2015 (MACRA), there are quality payment programs such as the Advanced Alternative Payment models (APMs) and the Merit-based Incentive Payment System (MIPS). These initiatives pay providers based on their performance and improved patient outcomes.
Other trends are impacting health data analytics, too. Population health management programs, which involve treating and monitoring groups of patients with specific medical conditions such as diabetes, hypertension or cancer, are increasingly being implemented.
Additionally, data that incorporates social determinants of health—such as biology and genetics, behavior, economic status or social environment (housing, education, transportation, income, and food insecurity) as well as other factors—are important health-related data that need to be included when analyzing the health and wellness of an individual, a group or a wider population.
As healthcare organizations embark on a machine learning project, they’ll have to ask themselves the following questions:
• What are the best use cases for a machine learning project, and where can the organization reap the best return on investment from the project?
• Does the organization have the right talent, data and technology to execute machine learning opportunities?
• How can the organization build trust and transparency into machine learning platforms and applications?
As HIT executives plan machine learning projects, remember that the health organization’s growth will hinge, in part, on how well it turns actionable insights into measures and initiatives that improve outcomes, reduce costs and raise the quality of care.