Project Lead: Richard Lewanczuk, University of Alberta’s Division of Endocrinology & Metabolism
There is a wide range of health conditions that can increase human frailty, which in elderly complex-needs people require intricate and costly care from the province’s health system that could be prevented.
It has been shown that, if frailty can be identified in an individual, there will be less need for utilization of the health-care system by applying proactive measures. Currently, there is not a comprehensive platform in place to identify frailty that includes regular patient contact, a robust patient-centred medical home model and resources to provide that screening.
Accurate frailty models for efficient patient care
Alberta's Health Link service provides 24/7 access to health professionals who respond to the public’s inquiries around health concerns. The interactions are recorded and transcribed for follow-up analysis. This includes detailed analysis of individual records and strategic analysis for further development of epidemiological models.
One of the challenges in the interpretation of Health Link’s raw transcripts is the variation in language use, including informal use of medical terms—like “temperature” or “fever”—and lack of clarity in personal reference such as the pronouns used, or temporal reference like identifying when a fever was observed.
A study conducted with approximately 5,000 transcripts produced a prototype system that will help create consolidated models of callers, identifying these factors and addressing the normalized use of medical terms.
As a second stage for the initial study, the team will include more data to support the application of machine learning, seeking to build more sophisticated and accurate models. This will include Health Link voice data and a sampling of primary care data collected via by the Canadian Primary Care Sentinel Surveillance Network (CPCSSN), in collaboration with the University of Calgary.
By combining all these data sources, the research team will extract appropriate information to construct calibrated models of patient frailty. Voice and vocal attributes can be extracted to build models of age and speaker anxiety. Coupling those models with the results from the previous Health Link study and adding selected CPCSSN, there is the opportunity to build the desired frailty model that can be widely and easily applied, even with an automated telephone system.
The project will allow for potentially more valuable health and medical interventions, determining the difference between regular health maintenance and imminent emergency by measures of caller anxiety.