Eliminating guesswork : an exploration of the role of predictive modelling in care management for patients with multimorbidities

Abstract

Introduction: Patients with multiple chronic conditions of diabetes, cardiovascular and kidney diseases are one of the most complex group of patients and high consumers of care. Multidisciplinary integrated care delivery models such as Integrated Practice Units have been introduced to improve patient care and reduce health care utilization through offering comprehensive and coordinated care. In addition to the traditional approaches of improving care around patients with multiple chronic conditions, innovative approaches such as developing predictive technologies using machine learning and artificial intelligence are needed to reduce costs and improve care delivery processes of patients with multiple chronic conditions. Multidisciplinary integrated care units are an ideal setting for development and application of predictive technologies using artificial intelligence and machine learning.

Aim: The aim of this thesis is to develop and explore how a predictive decision support model for physicians can be used to improve the management of clinical processes applied to individual patients with multiple chronic conditions of diabetes, cardiovascular and kidney diseases (HND patients).

Method: This thesis consists of four studies. Study I used descriptive statistics from a randomized controlled trial CareHND (NCT03362983) to describe and compare HND patients’ care utilization patterns between traditional care and multidisciplinary integrated care. Study II implemented two different types of Recurrent Neural Networks to learn about vectors representations of HND patients to demonstrate how ICD codes and clinical procedures contribute towards predicting 30-day hospital readmission using electronic health records data. Study III was a mixed-methods study employing an experience-based co-design model to explore the needs and preferences of physician managers regarding the role of predictive analytics in decision support for patients with multiple chronic conditions, and inputs in the design and development of hospital readmission prediction model. Study IV employed supervised machine learning models to improve and validate a hospital readmission prediction model using electronic health records data and compared their performance.

Findings: Study I found that HND patients consumed large amounts of healthcare resources including high hospitalization rates, emergency department visits and frequent encounters with the healthcare professionals. This finding implies that innovative methods like machine learning models should be used to explore the impact of integrated care interventions on care utilization. Study II found that three distinct sub-types of HND patients could be identified using patients’ vectors representation and clustering approach, and deep learning models were able to identify and quantify key contributors to hospital readmission. Study III found that healthcare professionals’ involvement in the design of predictive technologies right from the outset can facilitate a smoother implementation and adoption and enhance their predictive performance. Study IV found that hospital readmission prediction models perform better at the patient sub-group level, and target patients should be clustered based on most similar characteristics before development of predictive modeling.

Discussion: This thesis demonstrates how predictive analytics can be applied to cluster patients with multiple chronic conditions into sub-groups having clinically distinct characteristics and develop hospital readmission prediction models. More broadly, this thesis demonstrates how to conceptualize, design, and develop predictive technologies in complex patients with multiple chronic conditions using electronic health records data. This thesis establishes a groundwork for improving management of clinical processes of patients with multiple chronic conditions using machine learning models, and has implications for the wider development, implementation, and adoption of predictive technologies in healthcare. The healthcare management implications of the thesis are centered around the potential improvement of healthcare management practices through patient segmentation and hospital readmission predictions. The thesis also has implications for how opportunities can be created around the design and development of predictive technologies through co-design approaches by actively involving healthcare professionals.

Conclusion: This thesis demonstrates that involving healthcare professionals in the design and development of predictive technologies for patients with multiple chronic conditions in a multidisciplinary care setting can produce better results in predicting hospital readmissions and identify clinically distinct patient sub-types. For a wider implementation and adoption of predictive technologies in healthcare, knowledge and competence development of clinicians and managers on the use of such technologies is important. Efforts should be made to actively involve healthcare professionals in the conceptualization, design, development, implementation, and adoption of predictive technologies that may also help reduce healthcare professionals’ unfounded anxieties and concerns about the role of predictive technologies in their daily practices. The difficulty to operationalize the large quantity of available healthcare data should be overcome by ensuring seamless access to healthcare data that can facilitate a smoother design, development, adoption, and implementation of predictive technologies in healthcare. Building proper data structures around the EHRs such that the healthcare data can be collected and easily utilized in the development of predictive technologies is important.