Abstract
Medication non-adherence poses a significant healthcare challenge globally, impacting treatment effectiveness and patient outcomes. Nearly half of patients with chronic illnesses do not follow their prescribed medication regimens. Non-adherence is a complex and multifactorial issue influenced by multiple factors, including physiological, affective, cognitive, behavioral, social, and healthcare system related factors. Despite the life-saving benefits of long-term therapies, non-adherence rates persist at concerning levels, warranting a deeper understanding of the underlying mechanisms driving medication adherence behavior.
Prior research in the field has predominantly emphasized cumulative adherence behaviors rather than exploring longitudinal adherence patterns. However, with technological advancements, smart computing devices like smart pill bottles and pill boxes have been introduced to remotely monitor and track patients’ adherence continuously. Current intervention strategies to improve adherence often overlook the context in which adherence or non-adherence occurs, leading to potential limitations in their applicability to individual patients given the timing and context of implementation. Additionally, traditional methods in the medication adherence domain follow a one-size-fits-all approach and overlook individual variations that personalized models could address. However, the limited availability of data hampers the practicality of developing personalized models.
To address these challenges, we propose a theory-guided approach to develop computational models for longitudinal adherence assessment. Our framework is rooted in Social Cognitive Theory (SCT), which delves into the interplay of environmental, personal, and behavioral factors. We demonstrate the efficacy of our proposed approach using real-world data from two different studies involving distinct populations. Furthermore, we introduce Multi-task Learning (MTL) for personalized modeling, enabling us to tailor models to each individual while still being able to leverage sample-wide data for automated medication adherence prediction. Our approach entails building personalized Multi-task Learning Neural Network (MTL-NN) by treating each person's adherence prediction as a separate task.
The contributions of our work encompass three main aspects: 1) A theory-guided computational framework to predict medication adherence integrating factors at different temporal scales; 2) Enhancing our understanding of the significance of potential reasons for non-adherence; and 3) Hierarchical MTL-based personalized models of medication adherence. Our research sets the stage for future work personalizing interventions targeted at specific non-adherence behaviors.
