Researchers in Denmark and the U.S. have used deep neural networks to develop complementing models for predicting complications likely to arise in patients who’ve had surgeries of all kinds.
In a study report published in the August edition of The Lancet Digital Health, the team states their models overall outperformed preceding surgical risk-prediction tools while remaining robust—i.e., resistant to bias—when patient populations changed.
Lead author Alexander Bonde, MD, of the University of Copenhagen and colleagues trained multilayer and nonlinear neural networks on data from almost 4.7 million patients in an American College of Surgeons database that collects data from 722 hospitals in around 15 countries.
Searching the database for all adults who had surgery over a seven-year period ending in 2018, the researchers carved out another 1.17 million patient cases for a validation dataset and close to 14,000 for a test set.
Meanwhile the researchers developed three increasingly complex deep neural network models.
The complexity came in the form of progressively more variables in algorithm inputs.
The third and most sophisticated model had some 76 input features.
To try out the resulting neural networks on a discrete population, Bonde and co-authors pulled data on all patients who were treated at a single midwestern teaching hospital. They excluded these cases from the base dataset and held them aside for giving their model a final test.
To gauge the model’s reliability when challenged by changes in the underlying patient population, they used it only on cases involving emergency surgery.
Noting that they did not exclude rare procedures from the project—which likely would have inflated their performance metrics—Bonde et al. report that that the mean areas under the receiver operating characteristic curve of each of their models “outperformed previously published performance metrics, with a direct correlation between increasing model complexity and performance. Our models also retained predictive power despite substantial changes” in the underlying patient population.
Further, they comment,
Our algorithms might be used by clinicians to help guide future preoperative, intraoperative and postoperative risk management, serving as an important step towards personalized medicine in surgery.”
The team calls for a clinical trial to identify whether deep learning can, in practice, help cut post-surgery complication rates.
The study’s corresponding author is Kartik Varadarajan, PhD, of Harvard. Senior author is Martin Sillesen, MD, PhD, of the University of Copenhagen.
The study is available in full for free.