Researchers and physicians at the University of Pittsburgh and UPMC have successfully developed a machine learning model to predict patients at high-risk for complications after surgery. Their study, published in JAMA Network Open, highlights the potential of using artificial intelligence (AI) to identify and provide appropriate care to patients who may face post-surgery complications.
Before the COVID-19 pandemic, complications occurring within 30 days after surgery were the third leading cause of death worldwide, resulting in approximately 4.2 million deaths annually. Identifying patients at high-risk for these complications before their surgeries can not only save lives but also reduce healthcare costs.
Dr. Aman Mahajan, chair of Anesthesiology and Perioperative Medicine at Pitt School of Medicine and director of UPMC Perioperative and Surgical Services, emphasized the significance of improving the overall health of patients before surgery through prehabilitation. However, identifying high-risk patients can be challenging for busy clinicians, who must navigate an abundance of health data and frequently conduct additional testing and assessments. To address this hurdle, the researchers aimed to develop an easy-to-use model that utilizes existing electronic health records (EHR) to provide an automated and accurate risk assessment quickly.
To create the model, Dr. Mahajan, alongside Dr. Oscar Marroquin, UPMC’s chief healthcare data and analytics officer, and their teams trained an algorithm using the medical records of over 1.25 million surgical patients. The model focused on mortality and major cerebral or cardiac events, such as strokes or heart attacks, that occurred after surgery. Subsequently, the model was validated using data from an additional 200,000 surgical patients at UPMC.
After the validation process, the model was deployed across 20 UPMC hospitals. Each morning, the program scans the EHR of scheduled surgery patients and identifies those deemed high-risk. This notification enables clinical teams to coordinate care more effectively and implement prehabilitation strategies, such as encouraging healthier choices or referring patients to the UPMC Center for Perioperative Care. Consequently, the risk of complications is reduced. Moreover, clinicians can run the model on-demand at any time.
To assess the model’s performance against the industry standard, the American College of Surgeon’s National Surgical Quality Improvement Program (ACS NSQIP), Dr. Mahajan and his team conducted a comparison. While the ACS NSQIP is widely used in hospitals across the United States, it necessitates manual input of patient information and cannot make predictions if data are missing. The researchers found that their model outperformed the ACS NSQIP in identifying high-risk patients.
Dr. Marroquin highlighted that the team specifically designed the model with healthcare workers in mind. By completely automating the prediction process and allowing for educated predictions even with missing data, the model minimizes additional workload for clinicians, offering them a reliable and useful tool.
Moving forward, Dr. Mahajan and his team plan to further refine and develop the model. Their goal is to train the program to predict the likelihood of sepsis, respiratory issues, and other common complications that often prolong patients’ hospital stays following surgery.
The successful deployment of this machine learning model showcases the potential of AI in healthcare and its ability to support clinicians in providing optimal care to high-risk patients. By leveraging existing data in electronic health records, the model streamlines the identification and management of patients who would benefit from prehabilitation, ultimately improving surgical outcomes and reducing the global burden of post-surgery complications.
