Artificial intelligence (AI) offers a possible solution, says Indranil Bardhan, professor of information, risk, and operations management and Charles and Elizabeth Prothro Regents Chair in Health Care Management at Texas McCombs. AI models can predict the lengths of time patients will spend in the ICU, helping hospitals better manage their beds and, ideally, cut costs.
But although AI is good at predicting length of stay, it’s not so good at describing the reasons, Bardhan says. That makes doctors less likely to trust and adopt it.
"People were mostly focused on the accuracy of prediction, and that’s an important thing," he says. "The prediction is good, but can you explain your prediction?"
In new research, Bardhan makes AI’s outputs more understandable and useful to ICU doctors, an approach called explainable artificial intelligence (XAI).
With McCombs doctoral student Tianjian Guo, Ying Ding of UT's School of Information, and Shichang Zhang of Harvard University, Bardhan designed a model and trained it on a dataset of 22,243 medical records from 2001 to 2012.
The model processes 47 different attributes of patients at the time they’re admitted, including age, gender, vital signs, medications, and diagnosis. It constructs graphs that show a patient’s probability of being discharged within seven days. The graphs also depict which attributes most influence the outcome and how they interact.
In one example, the model calculates an 8.5% likelihood of discharge within seven days. It points to a respiratory system diagnosis as the main reason, and to age and medications as secondary factors.
Running their model against other XAI models, the researchers found its predictions were just as accurate, while its explanations were more comprehensive.
To test how useful their model might be in practice, the team surveyed six physicians at Austin-area ICUs, asking them to read and evaluate samples of the model’s explanations. Four of the six said the model could improve their staffing and resource management, helping them better plan patient scheduling.
The model has one major limitation, Bardhan notes: the age of the data. In 2014, the industry’s medical coding system changed from ICD-9-CM to ICD-10-CM, adding much more detail in diagnosis coding and classification.
"If we were able to get access to more recent data, we would have loved to extend our models using that data," he says.
His model need not be limited, however, to adult ICUs. “You could extend it to pediatric ICUs and neonatal ICUs," Bardhan says. "You could use this model for emergency room settings.
"Even if you're talking about a regular hospital unit, if you want to know how much or how long a patient is likely to need a hospital bed, we can easily extend our model to that setting."
Tianjian Guo, Indranil R Bardhan, Ying Ding, Shichang Zhang.
An Explainable Artificial Intelligence Approach Using Graph Learning to Predict Intensive Care Unit Length of Stay.
Information Systems Research, 2024. doi: 10.1287/isre.2023.0029