Every year, approximately 4.2 million individuals worldwide lose their lives due to surgical complications that occur within 30 days following a procedure. These complications not only result in loss of life but also contribute to increased healthcare costs.
In an effort to address this issue, a team of researchers from UPMC and the University of Pittsburgh has developed and implemented a machine learning algorithm capable of identifying patients at high risk of post-surgical complications. Recently, UPMC published research in JAMA Network Open validating the efficacy of this tool, demonstrating that it outperforms the industry standard.
To train the algorithm, the research team utilised medical records from over 1.25 million patients who had undergone surgical procedures at UPMC hospitals between 2011 and 2019, explained Dr. Aman Mahajan, the lead author of the study. The model was trained to recognize patients at risk of mortality within 30 days after surgery, as well as to flag those with a high likelihood of major cerebral or cardiac events, such as strokes or heart attacks.
An important aspect of the algorithm's design is its training on patient data that reflects the demographic profile of the state. Dr. Mahajan emphasised that UPMC is a large health system with diverse hospitals serving different populations, representing a broad geographic and demographic range.
The AI model also takes into account patients' social determinants of health, as these factors often influence a patient's recovery after an operation, added Dr. Mahajan.
To validate the model, UPMC deployed the algorithm to predict the risk for over 206,000 patients scheduled for surgery between June 1, 2019, and May 30, 2020. The health system then observed the model's predictions and found that it was accurate over 95% of the time in predicting both mortality and cerebral/cardiac events.
The research team also compared their model to the industry standard, the American College of Surgeons' National Surgical Quality Improvement Program (ACS NSQIP). While hospitals across the country use ACS NSQIP, it relies on clinicians manually inputting patient data and cannot make predictions if any information is missing. The study demonstrated that UPMC's model was superior in identifying high-risk patients compared to ACS NSQIP.
Following validation, UPMC implemented the algorithm across 20 of its hospitals. The tool reads electronic medical records each morning for scheduled surgery patients and identifies those at high risk of complications. When clinical teams receive this notification, they can take precautionary measures such as delaying the procedure or adjusting medications. Clinicians can also run the model at any time, including right before a patient visit, according to Dr. Mahajan.
For instance, consider a patient scheduled for colon cancer surgery. Their risk profile may vary based on their medical history and life circumstances, including comorbidities, medications, access to healthy food, and previous medical events. By running the risk score, the surgeon's office can quickly assess the likelihood of success relative to other medical conditions, enabling them to make informed decisions.
Moreover, the model promotes shared decision-making between patients and providers. Patients can engage in discussions with their healthcare providers about whether surgery is the best option or if it would be beneficial to postpone the procedure temporarily to address other health issues.
Currently, UPMC's model predicts the risk of mortality, major cardiac events, and cerebral events. In the future, the health system aims to train the algorithm to anticipate the likelihood of complications like sepsis, respiratory issues, and others that commonly require prolonged hospitalisation after surgery.
Click here to read the original news story.