Abstracts

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

by Rafiq Muhammad

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…