New research led by Texas McCombs offers an AI model that predicts ICU length of stay with explainable results, potentially transforming how hospitals manage critical care resources.
At the height of the COVID-19 pandemic, hospitals across the United States struggled to keep up with the demand for intensive care unit (ICU) beds as patient numbers surged. But even before the pandemic, ICUs faced persistent challenges in maintaining available beds for gravely ill patients.
Artificial intelligence (AI) holds promise in addressing these challenges by predicting the lengths of ICU stays, enabling hospitals to better manage their bed capacity and reduce costs.
Indranil Bardhan, professor of information, risk and operations management and the Charles and Elizabeth Prothro Regents Chair in Health Care Management at Texas McCombs School of Business, is at the forefront of this innovation.
Bardhan and his team have developed an AI model designed to predict ICU stay durations with improved clarity for health care providers — a concept known as explainable artificial intelligence (XAI).
“People were mostly focused on the accuracy of prediction, and that’s an important thing,” Bardhan said in a news release. “The prediction is good, but can you explain your prediction?”
Explaining AI Predictions
Bardhan’s research, conducted alongside doctoral student Tianjian Guo, Ying Ding of UT’s School of Information and Shichang Zhang of Harvard University, aims to make AI more interpretable and useful to ICU doctors.
The team trained their model on a dataset of 22,243 medical records spanning from 2001 to 2012, incorporating 47 different patient attributes, such as age, gender, vital signs, medications and diagnoses.
The model can produce graphs indicating a patient’s probability of being discharged within seven days and detailing which attributes most influence this outcome. For example, the model might show an 8.5% likelihood of discharge within seven days and highlight a respiratory system diagnosis as a primary factor, with age and medications as significant secondary factors.
The study, published in the journal Information Systems Research, found that the model’s predictions were as accurate as other leading AI models, but its explanatory power was superior.
Beyond the ICU
To test the practical application of their model, the researchers surveyed six physicians working in ICUs in the Austin area. Four of the six doctors indicated that the model’s explanations could help them improve staffing and resource management, aiding in more effective patient scheduling.
Despite its promise, the model has a notable limitation: it uses outdated data from prior to 2014, when the health care industry transitioned from the ICD-9-CM to the ICD-10-CM coding system, which offers more detailed and specific diagnostic information.
“If we were able to get access to more recent data, we would have loved to extend our models using that data,” Bardhan added.
However, this model has the potential to be adapted beyond adult ICUs. Bardhan suggests it could also be applicable in pediatric and neonatal ICUs, emergency rooms and even regular hospital units for predicting patient stay durations and optimizing bed management.
Source: UT Austin McCombs School of Business