We describe fundamental challenges to estimating heterogeneous treatment effects in the context of the statistical causal inference literature, proposed algorithms for addressing those challenges, and methods to evaluate how well heterogeneous treatment effects have been estimated. We illustrate the proposed algorithms using data from two large randomized trials of blood pressure treatments. We describe directions for future research in medical statistics and machine learning in this domain. The focus will be on how flexible machine learning methods can improve causal estimators, especially in the RCT setting.

To address the challenge of limited training data for machine learning models in the healthcare domain, we advocate for human-in-the-loop machine learning, which involves domain experts in an inter- active process of developing predictive models. Interpretability offers a promising way to facilitate this interaction. We describe an approach that offers a simple decision tree interpretation for any complex blackbox machine learning model. In a case study with physicians, we find that they were able to use the interpretation to discover an unexpected causal issue in a personalized patient risk score trained on electronic medical record data. To account for dynamics in disease progression, we advocate for building decision models that integrate predictions of the disease progression at the individual patient level with system models capturing the dynamic operational environments. We describe a case study on hospital inpatient management, showing how to build a Markov decision framework that leverages predictive analytics on patient readmission risk and prescribes the optimal set of patients to be discharged each day.

In this chapter, we explore how data-driven modeling can improve the understanding of OHCA risk, help identify the limitations of current AED placement strategies, and guide the development of optimal AED networks to increase the chance of AED use and OHCA survival. More specifically, we frame AED network design and related response efforts as a facility location problem, focusing on the maximum coverage location and p-median problems. We also highlight how novel tools that combine techniques from areas including information theory and machine learning with optimization models can shape the future of OHCA response efforts and AED placement strategies.

This chapter provides an overview of how researchers form interdisciplinary partnerships that leverage analytics-driven methods in AI, IE, and OR to tackle the most difficult societal and operational problems in healthcare.

This chapter provides an introduction to analytics-driven hospital capacity management through three projects that employed mathematical programming and discrete event simulation to address common challenges. The first project used mathematical programming to identify the mix of patients at Stanford Hospital that would maximize revenue given the capacity of hospital resources after a planned hospital expansion. The second project used discrete event simulation to plan the physical capacity and operational profile of a new procedural space at a hospital in New England. The third project combined mathematical programming and discrete event simulation to create a tool to schedule surgical procedures at Lucile Packard Children's Hospital Stanford.

This chapter offers lessons from engineering and other industries that promise developments in healthcare, and practical guidance for clinician-engineer partnerships. Section 1 provides guidance on how to establish a shared vocabulary and common understanding between engineers and clinicians of what terms such as AI and ML do and don’t mean. Section 2 identifies challenges clinician-engineering partnerships must overcome to deliver sustained value and ways to avoid common causes of failure. Section 3 provides specific advice on how to design projects to produce value at a series of stages rather than rely on the success of one, ambitious final model. Section 4 concludes by drawing on cautionary lessons from healthcare and other industries.

Designing AI-assisted technology to better understand major depressive disorder and further develop appropriate strategies for monitoring and treatment of major depression under resource constraints is an important and challenging task. In this chapter, we present seven studies that developed methods for AI-assisted, data-driven decision support systems to aid healthcare professionals. These methods focus on modeling chronic depression’s complex disease trajectories, identifying patients at high risk of progression, and recommending adaptive and cost-effective follow-up care. Long-term goals of this research include improving patient health outcomes and facilitating efficient allocation of healthcare providers’ limited resource through the use of novel technology.

Mathematical models may be used to optimize the decision of when to screen for cancer and how invasive a test to use, for example a biopsy or a biomarker. Partially observable Markov decision process (POMDP) models may be used to optimize screening decisions based on a patient's belief state, which is calculated using Bayesian updating and comprises a patient's complete history of biomarker test results. POMDPs can be used to determine how, if at all, biomarkers should be used for cancer screening in order to maximize quality-adjusted life years, a population health measure of disease burden that incorporates both the quality and quantity of life.

In this chapter, we: (1) detail the problem scope and gatekeeper training as a solution to it, (2) delineate key issues that can impact GT as well as the consequences of ignoring them, (3) explore the value of social network information for GT recruitment, (4) explain the mathematical model of GT and an algorithmic solution approach that can account for gatekeeper, population, and institutional issues, and (5) describe experimental results that compare a novel algorithmic solution for GT recruitment with recruitment approaches that approximate common GT practice.

Introduction

In the previous chapter, we examined the extent to which flexible working is widespread across the world. This chapter continues on from the previous chapter by examining why companies provide flexible working arrangements and who – which individuals – gets access to and uses flexible working. Through these analyses, this chapter aims to show that despite popular belief, provision of/access to flexible working may be still driven by performance-enhancing goals, rather than work-life balance or well-being goals. When examining who has access to flexible working arrangements, family and care demands of workers have limited explanatory power. Rather, it is better explained by the type of work carried out, the relative value the worker has – that is, their skill level, and position of seniority/power they carry, and in general how much performance outcomes employers can expect from these workers. This explains why disadvantaged workers, possibly with the most demand for such flexibility, are the least likely to gain access to such arrangements. This results in a rather polarised access to flexible working arrangements across the labour market. This chapter will look into these issues further. First, I explore the dual nature of flexible working – namely the different purposes it meets. I will then examine the theories explaining why employers provide flexible working arrangements. Finally, the chapter presents empirical evidence testing these theories, and reviews other already published work. This is done to argue that performance outcome goals may trump work-family goals when examining workers’ access to flexible working policies – depending on the flexible working arrangement in question.

The dual nature of flexible working

Flexible working as a family-friendly arrangement

Karasek (1979) in his study on the determinants of job strain explains how job decision latitude or job control helps workers reduce the negative impact high-job demands brings. Here job demands relate to one's workload or responsibility, while control is autonomy over various aspect of one's work including decisions about the way in which it is done and, more important for this book, when and where it is done. Demerouti, Bakker and colleagues (Demerouti et al, 2001; Bakker et al, 2003; Bakker and Demerouti, 2007) refines this model to a job demands and resources model.