In modern market theory, everything is about supply and demand, and healthcare informatics is no exception. As informatics technology and business competition rapidly grow in the healthcare industry, business leaders and users are expecting something more than retrospective reporting to gain insights.
The desire to answer forward-looking questions—such as “What’s next?” and “What should we do about it?” with all data points available—has become the key focus in healthcare business intelligence. Abundant electronic healthcare data, advancement in web analytic tools and the prevalence of mobile devices are driving analytics investments to bring healthcare operations data into the cloud and onto mobile devices for business leaders.
Providence Health & Services sought to take an innovative approach to implementing an analytics strategy. A team of its data analysts and engineers was charged with creating and delivering a working cloud-based, cross-platform app employing machine learning to solve a healthcare operational problem.
To do so within 90 days, the team broke with traditional organization culture and used a combination of open source and internally available resources, driven by an agile process methodology normally only seen in Silicon Valley startups. In doing so, the effort went in a direction that’s typically not seen in healthcare, where technical projects typically use large software implementations.
Providence has a healthcare intelligence team devoted to data expertise. The purpose of this team is to enable and empower all other caregivers within the Providence system with information, leveraging business intelligence technologies and solutions using or employing EDW’s, dashboards, automated reporting and advanced analytics.
The intelligence group has gained a high level of competency in building visually appealing dashboards that serve up complex retrospective data from disparate sources. However, the comprehensive retrospective reporting still falls short of what users expect from data reporting. Reports may give users insights into the past, but rarely inform and let users know what is happening now or what will happen in the near future.
A core team of developers, designers and data scientists created a proof-of-concept app that serves up prospective data by calling a cloud-based web service. The use case focused on hospital-based claim denials; it would predict the likelihood of hospital claim denial before a claim is submitted. The goal was to help revenue cycle management utilize predictive analytics to detect denial patterns and to lower denial rate on authorization and medical necessity.
Typically, the denial rate for primary claims is around 11 percent over a six-month period, and the top five denial categories include non-covered; additional information needed; information/contractual/payment; medical necessity; and missing authorization/referral. These five account for 71 percent of primary claim denials.
There are multiple instances of Epic within Providence and Swedish Health Services across five states. The data used in the denial predictive model was from Epic Clarity, a database used primarily to store our organization’s Epic data purely for reporting purposes. Key elements for a denial predictive model include patient demographics, visit details such as visit providers, CPT, diagnoses, and billing information like payers and plans. The team used six months of historical patient, clinic, billing data and the claim denial details posted to train and test our predictive model for primary claim denials. It consisted of around 12 million rows of data across systems.
In developing a model to predict claim denials, the team faced two main challenges. The first was to analyze the massive volume of data with limited memory and compute power available on local computers that analysts use, relative to the scale of the problem. It is difficult to take full advantage of programs like this to experiment and run many different models without running into memory bottlenecks.
Building models can be done on local desktop machines, but there’s a price to be paid in time and scale that is prohibitively expensive for the timeframes that were established. An alternative is to utilize the cloud computing platform to overcome computing and memory issues that tend to bottleneck data science projects.
However, solving one problem sometimes causes another, and moving to the cloud for machine learning produced a new challenge—protecting personal health information (PHI) on a cloud-based infrastructure. Anonymizing data to specifications set by the HIPAA Privacy Rule Data De-Identification Method requires care, but doing so while making a year's worth of data look real and consistent for R modeling is particularly difficult. To prevent exposure in a cloud context, an in-situ data masking architecture was applied that used substitution, shuffling, scrambling, synchronization and variance techniques on names, dates and numbers.
Selecting a machine learning algorithm depends on the size, quality and nature of the data. A deep understanding of how hospital clinical, billing and claim denials data work enables analysts to shape or subset the data in a meaningful way and select the best predictive model for the purpose.
For this pilot project, the team experimented with several algorithms, including two-class boosted decision tree, logistic regression and other binary classification predictive models. Several evaluation metrics were compared, including accuracy, F1 Score, and AUC (area under the ROC curve to inspect the true positive rate vs. the false positive rate), to assess the performance among selected models before deploying the trained model to web services.
To call the model in a web service, the team designed a simple app with input/output setup. When users type in required claim information, encoded parameters will be used as inputs to consume the web service and return the probability of denial on the mobile device. For the denial use case, the team used Microsoft’s Azure Machine Learning services (MAML). This particular service makes it very fast and easy to develop predictive models.
The web services were established, and the team created consumable RESTful service to connect R models to the mobile applications. Generally, trained models are available but have a tendency to be stuck in R space and are not able to be easily used by other services or clients. By turning the models into standard RESTful consumable web service, and copying the model to a server that hosted the model, it can then be called by clients.
The input/output functions as follows: MAML provides a URL that you send a POST request, and it encodes the parameters that will be used as inputs to the predictive model. In our case we have about 8 parameters. The output from MAML is in JSON format with a label and probability (ex: 70 percent denial).
The product team designed a minimally viable cross-platform product: a mobile app accessed through the Providence Health and Services domain only. Simplicity and ease of use were the main goals for UI design. The flow goes like this: A user fills out the given fields and clicks the single “Submit” button. The app ingests these fields as inputs into the web service, and outputs the predicted probability of a denial. The user interface will evolve into a more robust workflow as the predictive model matures and more variables are added.
The denial use case is an example of how healthcare data can be analyzed and accessed to achieve a useful purpose to an organization. By embracing the latest in cloud technology, the healthcare intelligence team was able to bring predictive analytics to business users in a functional and preferable way without being bounded to desktop or EMR login.
As more use cases are developed, issues such as safeguarding privacy and security, establishing analytics standards, and enhancing tools and platforms will gain more attention and improve the process of bringing healthcare operations data into the cloud.
Register or login for access to this item and much more
All Health Data Management content is archived after seven days.
Community members receive:
- All recent and archived articles
- Conference offers and updates
- A full menu of enewsletter options
- Web seminars, white papers, ebooks
Already have an account? Log In
Don't have an account? Register for Free Unlimited Access