Medical and technological advancement in the last century has led to the unprecedented increase of the populace's quality of life and lifespan. As a result, an ever-increasing number of people live with chronic health conditions that require long-term treatment, resulting in increased healthcare costs and managerial burden to the healthcare provider. This increase in complexity can lead to ineffective decision-making and reduce care quality for the individual while increasing costs. One promising direction to tackle these issues is the active involvement of the patient in managing their care. Particularly for chronic diseases, where ongoing support is often required, patients must understand their illness and be empowered to manage their care. With the advent of smart devices such as smartphones, it is easier than ever to provide personalised digital interventions to patients, help them manage their treatment in their daily lives, and raise awareness about their illness. If such new approaches are to succeed, scalability is necessary, and solutions are needed that can act autonomously without costly human intervention. Furthermore, solutions should exhibit adaptability to the changing circumstances of an individual patient's health, needs and goals. Through the ongoing digitisation of healthcare, we are presented with the unique opportunity to develop cost-effective and scalable solutions through Artificial Intelligence (AI).

This thesis presents work that we conducted as part of the project improving Medication Adherence through Person-Centered Care and Adaptive Interventions (iMedA) that aims to provide personalised adaptive interventions to hypertensive patients, supporting them in managing their medication regiment. The focus lies on inadequate medication adherence (MA), a pervasive issue where patients do not take their medication as instructed by their physician. The selection of individuals for intervention through secondary database analysis on Electronic Health Records (EHRs) was a key challenge and is addressed through in-depth analysis of common adherence measures, development of prediction models for MA and discussions on limitations of such approaches for analysing MA. Furthermore, providing personalised adaptive interventions is framed in the contextual bandit setting and addresses the challenge of delivering relevant interventions in environments where contextual information is significantly corrupted.       

The contributions of the thesis can be summarised as follows: (1) Highlighting the issues encountered in measuring MA through secondary database analysis and providing recommendations to address these issues, (2) Investigating machine learning models developed using EHRs for MA prediction and extraction of common refilling patterns through EHRs and (3) formal problem definition for a novel contextual bandit setting with context uncertainty commonly encountered in Mobile Health and development of an algorithm designed for such environments.  

Background and purpose: Patients’ adherence to medication is a complex, multidimensional phenomenon. Dispensation data and electronic health records are used to approximate medication-taking through refill adherence. In-depth discussions on the adverse effects of data quality and computational differences are rare. The purpose of this article is to evaluate the impact of common pitfalls when computing medication adherence using electronic health records.

Procedures: We point out common pitfalls associated with the data and operationalization of adherence measures. We provide operational definitions of refill adherence and conduct experiments to determine the effect of the pitfalls on adherence estimations. We performed statistical significance testing on the impact of common pitfalls using a baseline scenario as reference.

Findings: Slight changes in definition can significantly skew refill adherence estimates. Pickup patterns cause significant disagreement between measures and the commonly used proportion of days covered. Common data related issues had a small but statistically significant (p < 0.05) impact on population-level and significant effect on individual cases.

Conclusion: Data-related issues encountered in real-world administrative databases, which affect various operational definitions of refill adherence differently, can significantly skew refill adherence values, leading to false conclusions about adherence, particularly when estimating adherence for individuals. © 2020 The Authors. Published by Elsevier B.V. 

Background and purpose

Low adherence to medication in chronic disease patients leads to increased morbidity, mortality, and healthcare costs. The widespread adoption of electronic prescription and dispensation records allows a more comprehensive overview of medication utilization. In combination with electronic health records (EHR), such data provides new opportunities for identifying patients at risk of nonadherence and provide more targeted and effective interventions. The purpose of this article is to study the predictability of medication adherence for a cohort of hypertensive patients, focusing on healthcare utilization factors under various predictive scenarios. Furthermore, we discover common proportion of days covered patterns (PDC-patterns) for patients with index prescriptions and simulate medication-taking behaviours that might explain observed patterns.

Procedures

We predict refill adherence focusing on factors of healthcare utilization, such as visits, prescription information and demographics of patient and prescriber. We train models with machine learning algorithms, using four different data splits: stratified random, patient, temporal forward prediction with and without index patients. We extract frequent, two-year long PDC-patterns using K-means clustering and investigate five simple models of medication-taking that can generate such PDC-patterns.

Findings

Model performance varies between data splits (AUC test set: 0.77–0.89). Including historical information increases the performance slightly in most cases (approx. 1–2% absolute AUC uplift). Models show low predictive performance (AUC test set: 0.56–0.66) on index-prescriptions and patients with sudden drops in PDC (Recall: 0.58–0.63). We find 21 distinct two-year PDC-patterns, ranging from good adherence to intermittent gaps and early discontinuation in the first or second year. Simulations show that observed PDC-patterns can only be explained by specific medication consumption behaviours.

Conclusions

Prediction models developed using EHR exhibit bias towards patients with high healthcare utilization. Even though actual medication-taking is not observable, consumption patterns may not be as arbitrary, provided that medication refilling and consumption is linked.  © 2020 The Authors. Published by Elsevier Inc.

We consider a variant of the novel contextual bandit problem with corrupted context, which we call the contextual bandit problem with corrupted context and action correlation, where actions exhibit a relationship structure that can be exploited to guide the exploration of viable next decisions. Our setting is primarily motivated by adaptive mobile health interventions and related applications, where users might transitions through different stages requiring more targeted action selection approaches. In such settings, keeping user engagement is paramount for the success of interventions and therefore it is vital to provide relevant recommendations in a timely manner. The context provided by users might not always be informative at every decision point and standard contextual approaches to action selection will incur high regret. We propose a meta-algorithm using a referee that dynamically combines the policies of a contextual bandit and multi-armed bandit, similar to previous work, as wells as a simple correlation mechanism that captures action to action transition probabilities allowing for more efficient exploration of time-correlated actions. We evaluate empirically the performance of said algorithm on a simulation where the sequence of best actions is determined by a hidden state that evolves in a Markovian manner. We show that the proposed meta-algorithm improves upon regret in situations where the performance of both policies varies such that one is strictly superior to the other for a given time period. To demonstrate that our setting has relevant practical applicability, we evaluate our method on several real world data sets, clearly showing better empirical performance compared to a set of simple algorithms.