By Dr Mark Clements, MD PhD
An ongoing project aims to improve the quality of diabetes care for young patients at Children’s Mercy Hospital and adults from the Joslin Diabetes Center.
In 2017, The Leona M. and Harry B. Helmsley Char- itable Trust (Helmsley) funded a project, ‘Improving Type One Diabetes (T1D) Care with Machine Learning’. The objective was to combine expertise gathered from new technologies such as advanced machine learning and natural language processing into improving the quality of diabetic care among pediatric patients from Children’s Mercy Hospital (CMH) and adult patients from Joslin Diabetes Center (JDC).
In the U.S. the current healthcare system is designed to treat symptoms and treat people after an adverse event. Cyft’s predictive technology allows us to get ahead of this and start intervening and preventing adverse events for high-risk patients to improve health outcomes as well as making cost savings.
With the help of digital solution and analytics company, Cyft, the project aimed to gain a deep understanding of the operational opportunities and challenges in learning and managing T1D at both institutions. Specifically, the project aimed to, 1) develop and refine predictive models for clinical and financial outcomes of interest in T1D; 2) develop a change package to pair predictions with specific in-clinic interventions using a quality improvement framework, and 3) Implement Plan-Do-Study-Act cycles in order to rapidly perform small tests of change using promising and novel interventions i.e. developing ‘a learning health system’.
Within weeks of the funding, the project received more data than had ever been shared by either institution, including electronic medical records (EMRs) and data generated by diabetes devices. Even though the two participating institutions had a diverse group of patients and focused on very different problems, several interventions were successfully deployed with many more in the planning.
Overall, this project demonstrated learning, deployment, and evaluation capabilities at a speed that was never before possible. But most importantly, everything that has been designed, from solutions to infrastructures, can be scaled to deploy or appraise for further interventions.
Together with Cyft, the JDC was able to predict clinic no-shows and short-term cancellations, which reduce cost efficiency and limit revenue and fiscal health of the diabetes center. The JDC imple- mented multiple structural changes to its sched- uling system to support the capture of Process, Balancing and Outcome metrics, and also imple- mented multiple interventions, including the use of text message outreach to patients to reduce short-term cancellations across its multiple adult clinic types.
Cyft assisted CMH to develop predictive models for clinically significant rise in glycemia (hemoglo- bin A1c) in an effort to identify young people with T1D who are at risk from deteriorating glycemic control, decreased engagement with, or new barri- ers to, self-management, and other outcomes asso- ciated with rising HbA1c, such as hospital readmis- sion for diabetic ketoacidosis.
The model to predict rising A1c has a 50-55% positive predictive value when it includes EMR data alone and the positive predictive value increases to 70-75% with the inclusion of diabetes device data i.e. glucose, insulin, and carbohydrate information. The models have been validated with live clinical data since October 2018.
Nevertheless, the current model does not include data from fitness trackers or other daily personal measurement data dubbed the “Chro- nome”; more recently implemented patient report- ed outcomes collected in clinics, or geo-contextual information like neighborhood characteristics, proximity to ‘food deserts’ (a geographic area with few grocery stores
a glucose monitoring system company and KLUE, which produces competitive intelligence software, about their diabetes monitoring and meal prediction tools respectively. By October, CMH hopes to have a just-in-time text messaging intervention in place for youths who are predicted to experience deterioration in glycemic control. that contain healthy food options), or frequency of visiting certain geolocations dubbed ‘Envirome’. All these represent potential areas of expansion.
Notably, CMH already has a designated place in its EMR for clinically relevant ‘Envirome’ data. It is also important to note that home address- es can be used along with census data i.e., American Community Survey to estimate over 400 different parameters, that can serve as proxies for household income level, top educational achievement, and so on. Collectively, these data are rich sources of information to improve the present and future predictive models.
These predictive models have been paired with implementation science or quality improvement methods using the Model for Change as a guide. The CMH clinical team has developed a formal change package for reducing rise in hemoglobin A1c among youths, and has convened an expert group to perform an environmental scan of the literature and marketplace for prospective interventions to deploy.
In their first Plan-Do-Study-Act (PDSA) cycle, the clinical team has increased the contact between the diabetes care team and families of youths with T1D who are predicted to have a rise of hemo- globin A1c of 0.3% or greater in the next 90 days, via Direct-to-Home Telehealth. This intervention leverages on the American Well mobile app system, so families can engage with the diabetes team even when they are on the go.
In July this year, CMH launched its next Plan- Do-Study-Act cycle. In this cycle, the team is implementing a short-term peer support model for youths who are predicted to experience deterioration in glycemic control in the next 90 days. The CMH team has also been communicating with Dexcom,Cyft and CMH have continued their collaboration to work on a new model which predicts an outcome that is both financial- and health-related. In particular, the duo has developed a recurrent neural net model to predict hospital readmission for youths with T1D. The performance and feasibility of the initial model is currently under evaluation, but early results indicate that the model has a nearly 100% positive predictive value among youths at highest risk for hospital readmission in the near future.
CMH is planning to implement Direct-to- Home Telehealth in order to intensify care among youths at risk for hospital readmission. CMH is also exploring the Novel Interventions in Children’s Health (NICH) program to address adolescents with predicted high risk of hospital readmission. The NICH program has been developed based on developmental and systems theory, to deal with various life challenges and psychosocial stressors which make it difficult for young patients to meet the demands of their health conditions and day-to-day lives. CMH is presently engaged in discussions with the Director of NICH program, Dr. Michael Harris to pilot-test the intervention for future deployment in the present project.
The CMH team will continue to monitor the effectiveness of each clinical intervention. But the outcome of the first phase of funding has allowed CMH to form a learning health system. For the first time, the hospital is capable of deploying and testing innovative approaches to care, preventing negative outcomes faster than any clinical center.
The larger vision is to pave the way for other diabetes centers and healthcare systems to pair predictive analytics with the right interventions to improve outcomes for people with type 1 diabetes and beyond