Clinical Artificial Intelligence -Technology Application or Change Management?

Introduction

The practice of medical care is changing and changing rapidly (Coiera, 2018, 2019a; Topol, 2019). Health is not immune to the wide societal changes brought on by our increasingly digital existence (.The Economist, 2017). It has, however, been late to realize these transformative changes, unlike banking and other industries (Brynjolfsson & McAfee, 2011). This is because health has always been heavily reliant on human interaction, which has delayed many aspects of adoption. Artificial intelligence (Al), in its various guises, is increasingly having an impact on the clinical world (Coiera, 2019b; Topol, 2019)- With an initial emphasis on pattern recognition (radiology, dermatology), the usefulness of Al is broadening to other areas. Facing this clinical medicine is having to adjust how it creates and conceptualizes care. Indeed, as we will argue later, the changes wrought by digital health care are as fundamental to changing care delivery as the advent of the scientific method was in the 18th century.

To frame the conversation, we will examine our initial experiences with the development of an Al tool designed to inform decision-making at the consultation level in a primary care context, at this time best placed to have an impact on patient care. Essentially, we used machine learning to design a means of informing general practitioners (GPs - also known as family practitioners) and patients of the risk of attending an emergency department within the next 30 days. As part of that study, we researched practitioner attitudes to the tool and, in doing so, developed some insights into the future needs of Al-enhanced health care.

A well-recognized area of concern in modern western medicine is reducing hospital emergency department visits. Whilst often these patients require the attention of a specialist facility (for myocardial infarction or serious trauma, for instance), there remain several categories that may be managed in primary care, to prevent and/or reduce costly emergency attendances. These include (but are not limited to) chronic diseases such as cardiac failure and diabetes, as well as low acuity problems such as infection. We chose to approach this problem using the availability of digital health, using big data approaches, by extracting and linking records of all those who attended local emergency departments with their general practice data, and using machine learning to develop a prediction model.

The resultant tool, when tested on trial data, was able to accurately predict the risk of ED attendance within 30 days, based on the primary care data alone. In other words, at the time of the GP consultation, a screen popup will inform the GP (and the patient) of the risk (Pearce, McLeod, Rinehart, Patrick, et al., 2019)- It is therefore designed to deliver the information at the time best able to have actions taken to correct the problem. Because of this success, the tool is currently being deployed across 2,000 practices in the east of Australia.

However, the deployment of the tool raised several issues in its delivery (Pearce, McLeod, Rinehart, Whyte, et al., 2019)- Ranging from governance to clinical safety, we had to develop approaches to many new problems along the way. Whilst the tool is accurate 75% of the time, deployment of the program represents a new way of informing the clinical interaction, one that we believe requires a new clinical method, a new way of clinician to think about what it is to deliver care (Pearce, McLeod, Rinehart, Whyte, et al., 2019).

During the tool development, we tried to examine this in more detail. As part of the study, we engaged in a small qualitative assessment of the GP reactions, to further understand the impact on clinical thinking. The aim of this aspect of the project was to collect GP’s perceptions of the risk algorithm scores, the patient attributes and their potential treatment changes based on their own active patients. Implementing the risk scores in a live general practice environment and consequently acquiring feedback on the accuracy and utility of the report was considered an essential step to validate the future use of the alert tool.

Methods

The tool was trialed in ten general practices in eastern Melbourne. We used historical data from within the practice. One hundred patients were identified at each practice, and a selection had the risk rating presented to their usual GP. Participants then rated the tool performance. Seven of the GPs participated in follow-up interviews. Participants were required to review 60 patient records each (equaling to a total 420) with a spread of scores (high, medium and low risk) to ensure adequate evaluation across all score spectrums. Patient information shown for review, when available, included

■ Patient name

■ Demographics: Age, gender, pensioner, DVA, Aboriginal or Torres Strait Islander status

■ Risk Factors: Smoking

■ Diagnoses.

• - Current Diagnoses: Marked as active in GP software or identified in the last visit
• - Historical Diagnoses: Up to ten years.

■ Medications

• - Current medications: Marked as current in GP software or used in the past eight months
• - Historical medications: used in the last 9-24 months.

■ Pathology Tests and Results

• - Current pathology: Tests and results reviewed in the last visit
• - Historical pathology: Tests and results reviewed in the last 12 months.

■ GP Measurements (i.e. BMI, BP, Temperature, weight, etc.)

• - Current Measurements: Tests and results reviewed in the last visit
• - Historical Measurements: Tests and results reviewed in the last five years.

■ Patient risk scores for the 0-30 day and 31-365 day

GP participants answered a range of questions concerning the

■ Accuracy of the risk score

■ Accuracy of the attributes presented for each patient

■ Any changes they would make in patient treatment given the score.