No abstract

Background Successful integrations of machine learning into routine clinical care are exceedingly rare, and barriers to its adoption are poorly characterized in the literature. Objective This study aims to report a quality improvement effort to integrate a deep learning sepsis detection and management platform, Sepsis Watch, into routine clinical care. Methods In 2016, a multidisciplinary team consisting of statisticians, data scientists, data engineers, and clinicians was assembled by the leadership of an academic health system to radically improve the detection and treatment of sepsis. This report of the quality improvement effort follows the learning health system framework to describe the problem assessment, design, development, implementation, and evaluation plan of Sepsis Watch. Results Sepsis Watch was successfully integrated into routine clinical care and reshaped how local machine learning projects are executed. Frontline clinical staff were highly engaged in the design and development of the workflow, machine learning model, and application. Novel machine learning methods were developed to detect sepsis early, and implementation of the model required robust infrastructure. Significant investment was required to align stakeholders, develop trusting relationships, define roles and responsibilities, and to train frontline staff, leading to the establishment of 3 partnerships with internal and external research groups to evaluate Sepsis Watch. Conclusions Machine learning models are commonly developed to enhance clinical decision making, but successful integrations of machine learning into routine clinical care are rare. Although there is no playbook for integrating deep learning into clinical care, learnings from the Sepsis Watch integration can inform efforts to develop machine learning technologies at other health care delivery systems.

As debates about the policy and ethical implications of AI systems grow, it will be increasingly important to accurately locate who is responsible when agency is distributed in a system and control over an action is mediated through time and space. Analyzing several high-profile accidents involving complex and automated socio-technical systems and the media coverage that surrounded them, I introduce the concept of a moral crumple zone to describe how responsibility for an action may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system. Just as the crumple zone in a car is designed to absorb the force of impact in a crash, the human in a highly complex and automated system may become simply a component—accidentally or intentionally—that bears the brunt of the moral and legal responsibilities when the overall system malfunctions. While the crumple zone in a car is meant to protect the human driver, the moral crumple zone protects the integrity of the technological system, at the expense of the nearest human operator. The concept is both a challenge to and an opportunity for the design and regulation of human-robot systems. At stake in articulating moral crumple zones is not only the misattribution of responsibility but also the ways in which new forms of consumer and worker harm may develop in new complex, automated, or purported autonomous technologies.

Algorithmic impact assessments (AIAs) are an emergent form of accountability for organizations that build and deploy automated decision-support systems. They are modeled after impact assessments in other domains. Our study of the history of impact assessments shows that "impacts" are an evaluative construct that enable actors to identify and ameliorate harms experienced because of a policy decision or system. Every domain has different expectations and norms around what constitutes impacts and harms, how potential harms are rendered as impacts of a particular undertaking, who is responsible for conducting such assessments, and who has the authority to act on them to demand changes to that undertaking. By examining proposals for AIAs in relation to other domains, we find that there is a distinct risk of constructing algorithmic impacts as organizationally understandable metrics that are nonetheless inappropriately distant from the harms experienced by people, and which fall short of building the relationships required for effective accountability. As impact assessments become a commonplace process for evaluating harms, the FAccT community, in its efforts to address this challenge, should A) understand impacts as objects that are co-constructed accountability relationships, B) attempt to construct impacts as close as possible to actual harms, and C) recognize that accountability governance requires the input of various types of expertise and affected communities. We conclude with lessons for assembling cross-expertise consensus for the co-construction of impacts and building robust accountability relationships.

The wide‐spread deployment of machine learning tools within healthcare is on the horizon. However, the hype around “AI” tends to divert attention toward the spectacular, and away from the more mundane and ground‐level aspects of new technologies that shape technological adoption and integration. This paper examines the development of a machine learning‐driven sepsis risk detection tool in a hospital Emergency Department in order to interrogate the contingent and deeply contextual ways in which AI technologies are likely be adopted in healthcare. In particular, the paper bring into focus the epistemological implications of introducing a machine learning‐driven tool into a clinical setting by analyzing shifting categories of trust, evidence, and authority. The paper further explores the conditions of certainty in the disciplinary contexts of data science and ethnography, and offers a potential reframing of the work of doing data science and machine learning as “computational ethnography” in order to surface potential pathways for developing effective, human‐centered AI.