5 key learnings about conversational AI for healthcare analytics users
Healthcare payers can apply conversational AI to improve experiences – not just for members and physicians – but for their informatics teams and business users, too.
A few weeks ago, I helped my 10-year-old nephew make a special family dessert at his house. As we popped it into the oven, I explained how we needed to set a timer and surveyed the controls, looking for the right button. Then I heard him confidently command, “Set a timer for 45 minutes.” Done. It struck me how much technology has adapted to improve our user experiences at home, and how that approach has the power to improve experiences in my professional world of healthcare analytics.
Of course, analytics is often more complex than simple transactional requests, like setting a timer or getting a new member ID card. One analytic question often requires clarification, which may lead to another question. In our work developing conversational AI that works well for healthcare analytics users, we share five key practices and learnings:
1. Identify target use cases and personas.
It is important to be clear up front about the target use cases and personas you are designing for in incorporating conversational AI. There is a big difference in needs and usage patterns between a power user and a business user who uses the system infrequently.
One valuable use case we’ve identified relates to enabling business users to get answers to simple analytics questions. Many payers have implemented “self-service” analytic models. Conversational AI has the potential to be helpful in answering questions for business users, who are less likely to be proficient in the system. They are often trying to assess the big picture – understanding orders of magnitude – before requesting a more comprehensive analysis.
2. Anticipate questions and implement patterns.
Designers must involve users in testing and training AI to better anticipate what they will need most. Our experience with healthcare analytics and user research has enabled us to identify the top types of questions they ask. We’ve found that a large percentage of questions on healthcare analytics can be reduced to core set of queries. For example: What are the costs for X? What is the utilization trend for Y? How has enrollment changed over time?
The system should use a pattern-based approach, coupled with a rich healthcare ontology, so it can help users answer questions in other areas. For example, answering questions about cost trends for chiropractic services should also work for emergency department services, COVID-19 testing and specialty drugs.
3. Navigate to relevant content.
Anticipating the user’s true intent helps train conversational AI to deliver more accurate support. One study1 evaluated feedback from those using the technology in IBM Health Insights to improve program performance. User feedback drove changes to the system’s vocabulary, adding synonyms and variations of phrases that people commonly used, to improve accuracy of results.
A robust analytic interface can sometimes make it harder for the user to find valuable content. Applying conversational AI can help users quickly navigate to the answers they need. Users respond positively when technology guides them to useful content and anticipates additional insights that would be valuable.
4. Simplify access to meta data.
Robust analytic systems include tomes of supporting information on data lineage, enrichment and grouping methods, and adjustments embedded in the system. This is often presented in different system of documents.
A conversational overlay can help power users and casual users alike to quickly traverse this detailed analytic meta data. Being able to ask a question like What is the threshold for high-cost claimants? or How are avoidable admissions defined? gives easy answers within the workflow of defining or analyzing healthcare data.
5. Adopt closed-loop learning for long-term value.
To be effective over time, conversational AI must be a learning system that adapts to evolving user needs. For this, we need analytics on the conversational analytic experience.
Healthcare payers should be able to understand how their conversational AI is performing. For example, they should be able to see how analytic virtual assistants are performing and continuously understand patterns in unanswered queries to deliver a useful experience. Are there new patterns emerging we need to address? Where are conversations falling off?
We know from experience in other sectors that conversational AI requires an ongoing investment to ensure the system learns and adapts over time. Having the right analytic insights on the user experience is key to knowing what is working and what needs attention.
These key practices and learnings will help shape how well healthcare analytics users adopt conversational AI technologies. By the time my nephew’s generation enters the workforce, they will have high expectations for technologies that they will use to conduct business. Even in increasingly complex fields such as healthcare analytics, users will demand great experiences.
- Conversational BI – An Ontology-Driven Conversation System for Business Intelligence Applications. Very Large Databases (VLDB), Tokyo August 2020. Miller, et al