It impacts about one-fifth of the 3 million hospital patients seeking acute care each year in Canada. Because of it, the average length of a hospital stay in the United States has increased by 20 per cent in 2022 compared to 2019. In the United Kingdom, it costs approximately 100 million pounds per year. It has also impacted health care systems and patients in many other countries, such as Germany, Sweden, Scotland, Denmark, France, New Zealand, Australia, and South Korea.

It is a phenomenon called delayed hospital discharge.

One McMaster University researcher is helping to advance the management and prevention of delayed discharge using data analytics, artificial intelligence (AI), and optimization. His research — which can be applied globally — has the potential to reduce costs and hospital waiting times and improve patients’ health.

The effects of delayed discharge on health care 

Delayed discharge is a designation given to hospital patients who no longer need the level of care a hospital provides but still require additional care and support. However, their care journey is interrupted because the homecare support or long-term care they need is unavailable. As they wait to be discharged, the patient’s health has been shown to deteriorate, putting them at risk for adverse health outcomes such as readmission and death. At the same time, no other patient is able to use the hospital bed the delayed discharge patient occupies.

“The way delayed discharge impacts health care systems is almost universal,” says Manaf Zargoush, researcher and associate professor of Health Policy and Management at the DeGroote School of Business. “It has detrimental impacts not only on patient outcomes, increasing the risk of health and functional decline, but also on the entire health care system. It affects the efficiency and availability of care for other patients because of hospital overcrowding, which leads to increased wait time for emergency care and non-emergency surgeries.”

Using big data and AI to make informed health care decisions 

Part of the reason health care systems encounter challenges such as delayed discharge, says Zargoush, is because they lag in employing data to make predictions for improving decisions. To combat the issue, he uses big data and artificial intelligence to predict patient outcomes in different scenarios and optimize health care decisions. For example, 18 years of rich data on almost all delayed discharge patients in Ontario inform Zargoush’s research in two projects.

“While the cost, data, and other parameters differ from one geographical location to another, the data-driven framework and procedure we have developed and proposed in our studies remain the same,” he explains. “By retraining the same model with the new data and solving the same problem with new parameters depending on that geographical location, our framework can capture region-specific patterns leading to informed decision-making anywhere. This is the very promise of AI-driven decision-making, which is going to revolutionize our health care in almost all aspects.”

So, what do the results of Zargoush’s delayed discharge research look like in motion? In a hospital health care setting, for example, a computer can use Zargoush’s algorithm with a patient’s record to inform various care and transition decisions. Policymakers can also use the research to prioritize budget allocation and long-term decision-making.

“The complex models are working behind the scenes while the proposed solutions are very simple to implement in something like an app or computer software,” he says. “It’s as simple as health care professionals using their phones and computers to improve their decision-making. We’ve shown that this kind of decision-making outperforms many other decision-making paradigms, including first come, first serve.”

Delayed discharge, says Zargoush, is just one area of health care where artificial intelligence can be applied to make more informed decisions. Another is optimizing medication prescriptions for chronic disease management. Currently, Zargoush is working on two projects to help physicians make better prescription decisions for diabetic and hypertension patients.

“Patients are complex, which means different medications work differently for different patients,” he explains. “Personalized medicine is the opposite of a one-size-fits-all practice currently in place in many situations. We’re looking for the prescription decisions that are the best at the individual level depending on characteristics such as age, gender, other diseases a patient has, and all the medications that they are using. This personalized health care is made possible through using AI applied to electronic health records data.”

Addressing the Dark Side of Artificial Intelligence

However, Zargoush adds, it’s important to note that this research is not meant to replace physicians.

“This is not even possible, given the complex nature of human beings,” he says. “But I do see this misunderstanding sometimes. Instead, I view these methods as decision-support tools that help physicians, health care managers, and policymakers to make better choices.”

In addition to the concern that AI will replace health care providers, Zargoush also addresses the potential for unfair or unethical use of AI. In another recent study, for example, he has developed tools and methods to apply AI-based resource allocation in health care in a fair manner with respect to equity-seeking groups.

“Most of the time, the data we collect are inherently biased,” he explains. “So, if we train artificial intelligence on a biased data set, the resulting AI-trained predictions will be biased too. To this end, we have found ways to train our AI-based decision-making algorithms in a way to ensure fairness. This is the responsible way of data science. If rich data is available, there’s almost no limit to where artificial intelligence can be applied to improve how we do things in health care and almost all other fields, and we can do this responsibly.”

Note: In this article, artificial intelligence is synonymous with machine learning.