Predictive analytics has been in the population health management (PHM) space for a couple of decades. Predictive modeling vendors have come and gone, and risk scoring has been somewhat commoditized. What is next? What should PHM leaders expect to help advance program effectiveness and efficiency?
Beyond periodic vendor batch processing of administrative data
Most legacy predictive models originated for actuarial purposes. Because they focused on underwriting costs, it was important to have an accurate and complete representation of historical claims experience for a fixed period that was often limited to one year. These constraints have been built into most common predictive modeling processes and are the source of the poor timeliness and actionability that are the primary weaknesses associated with the use of claims data. Moving beyond this legacy thinking, and beyond administrative (i.e. health plan) data in general, is the beginning of an incremental move toward more advanced, higher-value predictive modeling.
The legacy approach of using health plan claims and member data has been extremely powerful. The mining of these data can be made more effective through the use of submitted (pre-adjudicated) claims and incremental, near real-time processing.
- Use more clinically meaningful and timely data — from EMR/EHRs, HIE, etc. — to change the models’ focus and multiply their ability to impact care quality
- Parse and identify the valuable features of its unstructured data as well as its structured content
- This will be made even more valuable by timely (real-time) model execution and appropriately constructed, prioritized (or automated) and delivered decision support. This ultimately suggests the need to tightly integrate, or even merge, the EHR data collection application with a powerful PHM workflow engine.
- Help you change patient and health system behavior to achieve desired outcomes
- Use consumer and psychographic segmentation data to understand the behaviors and motivations of targeted individuals and enable more effective engagement and behavior change strategies
- Since these data are typically external to clinical or health plan sources, segmentation may be applied independently and robust master data management will be required to assure that the correct characteristics are associated with each individual.
- These data will also be important in understanding risks related to non-medical determinants of health and associated risks. Since medical care is believed to only determine 10 to 20 percent of health outcomes, there are significant opportunities to mine and model data representing other determinants, such as health behaviors, social/economic factors, and physical environment.
- Capture relevant social media streams to identify person-level engagement opportunities and community-level opportunities for enhancing the brand, which may or may not be tied to clinical programs
Beyond predicting high future cost
- Use predictive models to predict the specific events that care management programs target to achieve cost savings
- Use predictive models to predict specific, avoidable events such as admissions, readmissions, disease progression, complications of conditions and treatments, ER visits, and high-cost imaging studies
- The more time-sensitive these predictions are, the more important it is to have a current stream of claims or EHR data to input.
- Predict current patient attributes that have not yet been observed
- Create profiles based on characteristics that we can infer from data that can then be segmented to assign outreach messages, approaches, and personnel most likely to be successful and to maximize their efficiency
- Create high levels of personalization by segmenting populations based on features such as predicted motivating factors, aspirations, and preferences, and matching members to programs, interventions, and personnel based on these features
- Identify segments of individuals who have significant risk but are amenable to simple, direct messages delivered by mail or to their inboxes, and other segments with people of moderate risk who are not cost-effective to intervene with through personal telephone outreach but who likewise are likely to respond to appropriate, targeted messages delivered via a low-cost channel
- Change from simple case identification (e.g., this person has diabetes and high future cost and is to be referred to ‘the diabetes program’) to case matching (e.g., this person has diabetes and, based on all of his or her characteristics and risks, will have the best chance of the desired outcome if he or she receives this specific bundle of interventions, in this way, and from these sources.)
To advanced, multi-dimensional predictive analytics embedded in highly automated decision support and PHM workflows
Real-time scoring and segmentation occurring across multiple dimensions not only greatly enhance the value of data, they also increase the complexity and time-sensitivity of the results; their value is best leveraged through extremely tight integration with PHM software. Because of the high level of personalization, it is neither practical nor efficient to generate lists in one application to be implemented in another. Instead, complex logic will assign programs and specific interventions within programs to individuals based on multiple criteria designed to create the best possible results with the fewest possible resources. That is, highly configurable automation will drive highly personalized interventions ranging from the highest intensity to the extremely low-cost, to achieve the best outcome in the most efficient way possible.
- Automate the process of continuous improvement by embedding study design methodologies, including randomization (A-B testing) in some program settings and machine learning algorithms in others
- Use feedback loops to validate segmentation and stratification methodologies
- Identify priority areas for improvement, and optimize intervention and program designs