in ,

Artificial intelligence fights for acceptance in health care


EARLY in the COVID-19 pandemic, progressive respiratory failure developed in approximately 5% of unvaccinated adults with COVID-19, typically one week after the onset of coryzal symptoms (here and here).

Worldwide, almost two-thirds of patients admitted to intensive care with respiratory failure secondary to severe COVID-19 in 2020 required invasive mechanical ventilation. Requirement for mechanical ventilation carries a high mortality rate and is both labour and resource intensive, so identifying cohorts at high risk for mechanical ventilation is a priority. From a patient perspective, delaying an inevitable intubation increases the risk of sudden respiratory arrest, and unplanned airway management increases the risk of staff infection.

Conversely, avoiding intubation where possible decreases the risk from intubation itself and the sequelae of mechanical ventilation, including ventilator-induced lung injury, nosocomial infection, pressure injuries, and thrombosis. Accordingly, developing tools to accurately predict patients at risk of deteriorating is a priority.

The Short Period Incidence Study of Severe Acute Respiratory Infections (SPRINT-SARI) Australia registry has been prospectively collecting comprehensive data on critically ill patients with COVID-19 admitted to Australian intensive care units (ICUs) from February 2020. Through careful oversight and standardised collection protocols, the strength of the database lies in its consistency and reliability, applied throughout more than 50 hospitals nationwide. Additionally, the data collected are highly granular, including more than 100 demographic, clinical, and laboratory findings per patient-stay.

In October of 2022, the SPRINT-SARI group of investigators published the first study to leverage artificial intelligence/machine learning to identify readily available clinical risk factors for mechanical ventilation in patients with COVID-19 admitted to the ICU using Australian data. The focus of the study was on a “grey-area” cohort of patients who are unwell enough to necessitate admission to the ICU but are yet to require mechanical ventilatory support. The high sensitivity (81%) of machine learning algorithms devised in the investigation enabled the early identification of those at risk of requiring invasive ventilation within three days of ICU admission, providing a powerful tool that aids clinical decision making, and enables more streamlined resource allocation at an organisation level.

Machine learning is a burgeoning research field that has a growing track record of accurately characterising and predicting complex biological phenomena that do not have a priori models. The advantages over traditional (often linear) statistical methods lies in an ability to navigate non-linear interactions within high dimensional data, creating a mechanism through which more accurate predictions can be made.

Machine learning has proven highly effective in “personalised prediction”, such as in the algorithm used by Netflix to individualise recommendations for television shows and movies, or by Facebook (now Meta) to deliver personalised advertisements. The uptake of these powerful algorithms has been slow but steady in patient care, with notable achievements such as a detection system, approved by the US Food and Drug Administration, for diabetic retinopathy in the field of ophthalmology, and automated image analysis for hospital workflow management in radiology.

In a health system strained by the ongoing pressures of staff shortages, rising costs, and lengthy wait times, machine learning approaches provide two key advantages: scalability and effective resource allocation. Today, machine learning algorithms can be deployed to cloud servers where they can be accessed anywhere with an internet connection.

In Australia, this means we can work towards equitable access to artificial intelligence health care technologies in part by increasing internet coverage. Furthermore, the ability to integrate machine learning-driven decision support and telehealth services provides an exciting prospect. Between 13 March 2020 and 31 July 2022, 118.2 million telehealth services were delivered to 18 million patients by more than 95 000 practitioners nationwide.

Looking forward, seamless integration of artificial intelligence into the backend of telehealth services could serve as an adjunct to timely diagnosis and early intervention. With regards to resource allocation, artificial intelligence built into the electronic medical record could aid the triaging process for both primary and tertiary care. Established, accurate models built on large datasets (> 50 000 patients) already exist to predict progression of, for example, diabetes to chronic kidney disease. These have utility in ensuring high risk patients are prioritised and not “lost” within the long waiting lists for specialist care, especially in rural settings.

Despite the clear advantages of artificial intelligence, resistance to its clinical implementation persists, most notably due a lack of machine learning algorithm explicability. The phrase “black box” continues to permeate the narrative surrounding artificial intelligence implementation in health care, capturing the general mistrust in an operation whose inner workings remain imperceptible to the physicians utilising them. There is a fear that relinquishing the responsibility for decision making to this unknowable ally could violate the clinician’s primary responsibility to protect patient welfare and may lead to medico-legal exposure.

This wariness extends to patients too, who may be unnerved by the power afforded to “non-human” machines in the determination of their management. Further complicating matters is that a trade-off exists between model complexity and explicability. The complex algorithms that are best equipped to extract and leverage data trends are often the most difficult to understand.

Importantly, efforts are growing to bridge this gap in trust, with the development of third-party explanatory systems that are as creative and intelligent as the algorithms themselves. In 2020, a novel explanatory modelling system for improved interpretability of non-linear machine learning models was released, using an approach based on game theory to rationalise the contributions made by varying input features to the predictions being made.

This system was utilised by the SPRINT-SARI group to clearly identify which features were most contributory in their artificial intelligence system for predicting invasive ventilation in critically unwell patients with COVID-19. The goal was to provide physicians with an insight into the algorithm rationale, garnering clinician confidence and boosting uptake. The algorithm transparency achieved by this novel approach provides hope for an improved working relationship between health care professionals and artificial intelligence-based tools employed in the hospital setting going forward.

Ultimately, predictive algorithms are only as useful as the data on which they are trained, meaning that utility is strictly limited by the demographic and disease-related features of the cohort in question.

In Australia, the majority of large-scale clinical research is undertaken at tertiary centres in a metropolitan setting. This grossly under-represents national health service delivery, with only 12.8% of the population admitted to hospital in 2021–2022 compared with 83.6% who visited a GP. The development of widespread data-linkage platforms and community databases such as PATRON, run by the University of Melbourne, are changing this trend, opening doors to large-scale artificial intelligence systems that can be implemented in primary care. For GPs, artificial intelligence models developed on such data could provide a plethora of diagnostic and prognostic aids, enabling more streamlined patient flows and earlier escalation of deteriorating patients to specialist or emergency care.

Before moving to universal adoption, a key consideration will be the generalisability of artificial intelligence models, given the infrastructure discrepancies that make data collection more challenging in regional practice. There is no guarantee that artificial intelligence tools developed from data in metropolitan centres could be generalised to patients from rural and/or Indigenous backgrounds.

Future quantitative research must take into consideration the proportionately lower primary care, and higher tertiary care accessed by Aboriginal and Torres Strait Islander populations compared with non-Indigenous Australians.

Artificial intelligence is ground-breaking but cannot serve as a quick fix for systematised and deeply embedded disparities in our health care system. Consequently, within this nascent health–technology nexus, we as a nation have an opportunity and responsibility to ameliorate rather than worsen the unacceptable health inequalities faced by our Indigenous and rural population. There exists an imperative to establish appropriate methods that will lead to systematic, routinely collected, accurate, population-based data in regional and remote areas, as well as strengthen the data collection practices already in place in metropolitan settings.

A recent industry report into Australia’s health technology sector asserted that “the future is already here; its’s just not very evenly distributed”. A key question going forward will be how best to address this challenge and achieve wide distribution of scalable artificial intelligence in the health sector. Our modern data analytics capabilities driven by rapid artificial intelligence advancement remain underutilised, slowing the development of sophisticated predictive tools that could permit early risk stratification for COVID-19 and other diseases.

Progress necessitates improved clinician and patient awareness for the utility of broad, systematic data collection, and appreciation for the complex tools available to harness them. Concurrently, data scientists have a responsibility to address the opacity of artificial intelligence and enable clinicians to peer “under the bonnet” of these intricate machines.

The resultant multidisciplinary collaboration is capable of fuelling the political and economic engines required to propagate technological advancement in the health care sector, advancing us towards the ultimate goal of cost-effective, individualised and truly holistic care for Australians.

Dr Roshan Karri is a surgical resident in the intensive care unit at the Royal Melbourne Hospital.

Associate Professor Mark Plummer is the Head of Research and Innovation in the intensive care unit at Royal Adelaide Hospital and the University of Adelaide.

 

 

 

The statements or opinions expressed in this article reflect the views of the authors and do not necessarily represent the official policy of the AMA, the MJA or InSight+ unless so stated.

Subscribe to the free InSight+ weekly newsletter here. It is available to all readers, not just registered medical practitioners.

If you would like to submit an article for consideration, send a Word version to mjainsight-editor@ampco.com.au.

 



Source link

What do you think?

This TikTok-famous Puffer Vest Is Also an Amazon Best-seller — and It’s on Sale Now – Yahoo! Voices

Demand for remote jobs outpaces supply