An extensible and flexible process to speed time to value for our clients and improve health outcomes

Complex healthcare problems require detailed, impactful solutions. At Decode Health, we have shaped our data solutions platform to help decision makers proactively address myriad challenges faced in healthcare. Our process is flexible and fast: when presented with a new need by our customers, we join together the relevant parts of our platform to produce results that are meaningful and applicable to their needs. Whether modeling chronic disease risk or incorporating emerging public health data in real time to predict COVID-19 infection rates, we adapt, test, and deliver insights that help stakeholders understand and manage healthcare and disease-related risk.

Identify the problem

The Decode Health process begins with identifying the healthcare risk we intend to predict and understand. Before engaging our data team, we conceptualize the challenge at hand, understand the business use case and decide how we’ll measure success. We then collect and format the data and employ our transformation pipelines to lay out a framework for the solution.

Claims data, commonly analyzed in descriptive and predictive efforts throughout the healthcare space, is notoriously difficult to wrangle and can vary in completeness and length of the overall longitudinal record tied to a plan member. We developed our early platform using this widely available dataset and focused on chronic disease risk challenges using population-level data combined with advanced machine learning methods to define specific data patterns for patients with potentially undiagnosed, misdiagnosed and uncontrolled chronic disease. This effort honed our ability to solve modeling pain points, like repeatable data collection and allowed us to streamline and optimize our raw data structuring process prior to generating machine learning models and final results. Our early work addressing these challenges led us to architect flexible pipelines to handle future data sources.

These capabilities were put to the test as the COVID-19 pandemic began to unfold. Our partners asked Decode to provide insights to illuminate the evolving risk profiles associated with COVID-19 case growth and COVID-19 related health outcomes. Because of our unique ability to use repeatable, flexible sections of our existing data science pipeline, we rapidly began predicting COVID-19 case growth and community vulnerability to poor outcomes, including hospitalization or death, nationally. Use of our workflows, and community level predictions allowed us to identify and track evolving demographic risk factors contributing to COVID-19 spread in Tennessee and Georgia.

Collect, clean, and transform the data

Adapting use of our chronic disease risk prediction framework to deliver predictions of COVID-19 required that we amass new data sources to inform our modeling efforts. Our process quickly identified relevant datasets specific to the COVID use case that would produce highly accurate insights tied to a member, a population or a geography. These data sources ranged from dynamic, rapidly changing national and statewide public health datasets to information obtained directly from patients through real-time COVID-19 risk surveys. The Decode Health process considered all possible data elements and found the optimal data features to produce accurate predictions.

Decode’s process matures with every engagement. We continue to expand our use cases and the datasets necessary to predict and explain disease risk. As we deploy our process we continually refresh our models to maintain high levels of predictive accuracy and to spotlight the most important data features that contribute to Decode’s ability to accurately predict future risk. Our goal is ongoing examination and understanding of the interrelationship of all data elements to spotlight unique correlations tied to disease outcomes. For example, our recent work provides direct evidence that higher SARS-CoV-2 vaccination rates in communities plays a significant role in lowering predictions of future COVID-19 case growth and poor outcomes including hospitalizations. This work has also highlighted the association of COVID-19 case growth in Tennessee with younger demographics in the spring of 2021. As these correlations are identified, Decode uses this knowledge to consider inclusion of additional enrichment datasets, like information to model the social determinants of health (SDOH) within a community or real-time survey data that we collect through direct member-level engagement. This allows us to provide greater context and insights to supplement baseline data sources.

Beyond vetting and incorporating novel data sources, we constantly iterate through data transformations to ensure that each new data source is integrated seamlessly into Decode’s expanding data warehouse. Decode has highly scalable workflows to enable longitudinal inclusion of expanding data schemas over time. Our data architecture process transforms individual data fields to prepare each data input for machine learning analysis. Decode has built a proprietary repository of novel data transformation techniques to maximize prediction accuracy. As we continue to build on our platform, we have processes in places that rapidly test new data transformation logic. This flexible data structuring approach increases the likelihood that our machine learning methods will discover specific patterns in the data that produce the strongest predictions and context for each prediction during each partner engagement.

Apply and adapt advanced machine learning models to produce actionable, explainable results

Building upon this rich collection of data and data transformation processes, we train and test hundreds of machine learning models to identify the modeling approach that best captures a pattern in the data for the problem we set out to solve. This ensemble method leverages the power of an automated machine learning process, taking advantage of the latest, methods to build these models in a competitive environment until we attain the models that best fits the disease risk profile with highest overall accuracy.

While this approach identifies the best model from a data science perspective, we have developed additional metrics to compare model performance as we vary the unique data features that serve as inputs into our automated machine learning process. Varying the volume and structure of the input datasets prior to the modeling effort is critical as we search for the model that delivers high quality results. Regardless of the challenge we are trying to understand, Decode Health ensures that the process of training, testing, and deploying models to predict disease risk are flexible and scalable to deliver and maintain the best model for each data problem.

Decode implements both model-level metrics and real-world accuracy to transform results into actions. These results are interpreted with knowledge of the business use case and how Decode’s partners intend to take action. The context of these results is essential, from explanation of potential dollars to be saved to discovery and explanation of the specific clinical or social determinants of health risk factors that lead to the best opportunity for teams to mitigate healthcare risk. 

The power of our methods – speed and flexibility

Beginning with the unique challenge we are trying to solve and working toward actionable results, each aspect of the Decode Health pipeline is intentionally developed to quickly collect new data, test new ideas, and incorporate modeling refinements over time that generate immediate results as soon as each partner engages our team. We constantly refine Decode’s approach to understand our data problems, old or new, and to uncover the best solutions for our partner’s needs.

Healthcare companies including those who are delivering care directly, those who are paying for care, those who are developing therapies, or those who are conducting diagnostic testing, benefit from the actionable insights enhanced with predictive context uniquely developed by Decode Health. These insights allow for risk evaluation on an individual- or at a population- level for chronic or infectious diseases, including emergence or exacerbation of these diseases. Decode’s insights identify individuals or populations who will benefit from early actions. With the powerful combination of insights and context, organizations can help to prevent the onset, exacerbation, or spread of illnesses and can change outcomes. Additionally, our solutions help our clients engage directly with those who will benefit from specific interventions through simple, secure channels to engender information sharing or actions on the part of the targeted individuals.

Healthcare companies not directly involved in patient care use our risk identification and stratification methodologies and tools to increase their operational efficiency and improve financial results, as they may not have the capacity, time or appetite to build their own capabilities in this arena. Instead, they work with Decode Health to leverage the knowledge and experience we have gained as a direct result of our work with data tools, machine learning, and analytics platforms.

Organizations that engage with Decode Health benefit from our expertise and from the flexible, extensible machine learning platform developed by our expert team of data scientists.

Contact the team at Decode Health to learn more about our unique approach to truly predictive analytics. Put our adaptability, agility, and speed to work for your organization to help improve outcomes while better managing risk and costs.