0
0
At the height of the COVID-19 pandemic, hospitals faced severe shortages of intensive care unit (ICU) beds. Even before the pandemic, managing ICU bed availability was a persistent challenge, especially with an aging population. According to data, 11% of hospital stays involve ICU care, further straining resources.
Artificial intelligence (AI) may offer a solution, says Indranil Bardhan, professor at Texas McCombs School of Business. Bardhan and his team have developed an AI model that predicts the length of ICU stays, helping hospitals optimize bed management and reduce costs.
Improving Predictive Models
While AI can accurately predict how long a patient might stay in an ICU, it often struggles to explain the reasoning behind its predictions. This lack of transparency makes doctors hesitant to trust and implement AI recommendations.
“People were mostly focused on the accuracy of prediction, and that’s an important thing,” says Bardhan. “The prediction is good, but can you explain your prediction?”
To bridge this gap, Bardhan and his team introduced an approach called explainable artificial intelligence (XAI). Working with researchers from UT’s School of Information and Harvard University, they designed a model trained on 22,243 medical records from 2001 to 2012.
How the Model Works
The AI model considers 47 different attributes of patients upon admission, including age, gender, vital signs, medications, and diagnosis. It then constructs graphs that predict the likelihood of discharge within seven days, highlighting key factors influencing the outcome. For example, in one scenario, the model calculated an 8.5% probability of discharge within a week, with respiratory system diagnosis as the primary contributing factor and age and medications as secondary factors.
The researchers found that their model’s predictions were as accurate as existing XAI models but offered more comprehensive explanations.
Practical Applications and Challenges
To evaluate its practical impact, the team surveyed six ICU physicians in Austin, Texas. Four out of six doctors believed the model could help with staffing and resource management, enabling better patient scheduling.
However, Bardhan acknowledges a key limitation: the model was trained on older data. In 2014, the healthcare industry transitioned from the ICD-9-CM coding system to ICD-10-CM, introducing greater diagnostic detail.
“If we were able to get access to more recent data, we would have loved to extend our models using that data,” he says.
Despite this, Bardhan believes the model could be applied beyond adult ICUs, including pediatric and neonatal ICUs, emergency rooms, and even standard hospital units to predict patient bed usage.
Future of AI in ICU Management
The study, titled An Explainable AI Approach Using Graph Learning to Predict ICU Length of Stay, was published in Information Systems Research. With further development, AI-powered scheduling could significantly enhance ICU efficiency and improve patient outcomes.
Disclaimer: While AI models show promise in improving ICU bed management, their effectiveness depends on data accuracy and physician adoption. AI should be used as a supplementary tool rather than a replacement for medical expertise.
