As COVID19 swept the US, we realized that analytical warehouses were not supporting our need for real-time analytics. Administrative and clinical leadership at health systems were asking: Where are beds available now? How many ventilators, and other highly sought-after supplies, are available and where are they located in each hospital? Standard registries in transactional stores were focused on their use cases and are out-of-date in the analytical stores that primarily refreshed overnight. Thus emerged the need for more time-sensitive data and analyses.
KenSci’s engineers built its Real-Time Messaging Pipeline to listen to the stream of HL7 and/or FHIR messages that move within a health system and its partners. We were able to use this information to build tools to provide operational leadership with a 10,000-ft view of their facilities and resources. (see our work on COVID19 Response Planning here)
As we have moved into 2021, needs for this pipeline are shifting beyond operational settings (what is happening?). We are applying our advanced analytics to real-time streaming data. For many predictive models (what will happen?), analytical warehouses are sufficient sources of information: if we are predicting the likelihood that a health plan member has an undiagnosed chronic condition, this does not require real-time insights and considers the similarity to other patients. However, many models are more sensitive to time due to their dependency on events that occur during an active encounter. Examples of time-sensitive models include:
Identifying patients entering the emergency department (ED) who are likely to be sent home with services or require a higher level of care (inpatient)
Identifying patients who are likely to discharge to a skilled nursing facility based on past organizational patterns, creating a potential moment of intervention to prevent an unwanted SNF discharge.
Determining the likely discharge date for those patients still in active hospital care to facilitate discharge orders and inform future capacity or bed utilization
These models are looking for changes in the trajectory of an encounter. Using the real-time pipeline for message consumption and as a model scoring trigger allows models to score/re-score whenever a change occurs and enables the delivery of an updated prediction as soon as its supporting information available.
Models fed and predicting against real-time data have access to the same information as a user him or herself: this risk score is not just informed by last night’s analytical warehouse refresh, but includes latest lab orders or results alongside historic utilization patterns. KenSci’s real-time models reflect a comprehensive view of risk.