Machine-Learning Tools Predict Post-Op Complications, Surgery Duration

New research shows that machine-learning (ML) tools can use perioperative data to accurately predict post-operative complications and surgery duration.

Surgery and its potential complications create significant burdens on patients, providers, and health systems in terms of health outcomes and associated costs. Research suggests that approximately 7 to 15 percent of major surgery patients are expected to experience a major complication. The press release announcing the research indicates that one-third of hospitals’ expenses are spent on care to prevent these complications, which can be life-threatening.

Accurate risk stratification for patients undergoing surgery is key to preventing them. However, the risk assessment tools that many clinicians rely on can often be limited, either in the scope of what complications they are capable of forecasting or, as the research team notes, how they manage the complexity of perioperative data.

The perioperative period — the time just before and after a patient is in surgery — and the efforts taken during this time to prevent complications generate a massive amount of data. These perioperative data include information related to demographics, history of comorbidities, lab tests, medications, clinical notes, and physiological signals, along with hundreds of other data points.

Using machine learning, researchers and clinicians can gather and use this information to predict surgical complications, but the large number of variables to consider can make model development and validation difficult. The researchers indicated that there are two major pitfalls here: the high dimensionality of the data and the rates of missing information in perioperative data.

High dimensionality refers to the large number of variables in the data, which can vary significantly between patients and datasets. This means that a model may achieve high predictive performance during training, but it may be “brittle” during validation or in the real world because it doesn’t fit those datasets well. This lack of generalizability can lead to significant deterioration and instability in predictive performance, meaning that such a model would not be fit for clinical use.

Rates of missing data present similar issues. According to the press release, perioperative data have extremely variable missing rates across different predictors depending on patient and surgery characteristics. This makes it a challenge to create a model that can account for these data complexities while also maintaining robust performance.

To address these challenges, the research team began by developing a deep-learning (DL) approach that could be used to transform the data without impacting its quality or integrity. The approach, known as clinical variational autoencoder (cVAE), lowers the dimensionality of the data while still capturing the nonlinear relationships among the original perioperative variables.

They then applied cVAE in the development of two ML models: one for forecasting which patients would develop post-operative delirium and one for predicting surgery duration. Using cVAE, the researchers were able to compress 562 clinical variables into 10 variables to make these predictions.

Overall, both models were able to make robust and accurate predictions using these compressed data, or latent space, without suffering the drops in performance typically associated with the use of other autoencoders. The researchers noted that these findings have the potential to significantly impact clinical care in the future.

The research was presented at the Association for Computing Machinery (ACM) SIGKDD International Conference on Knowledge Discovery and Data Mining earlier this month.

This study adds to a growing body of research showing the potential of ML to enhance predictive analytics and address surgical complications.

Last year, researchers at Thomas Jefferson University developed a model for predicting complications such as kidney failure and stroke following surgery.

In June, University of Florida researchers announced a new model, which was found to predict complications as accurately as clinicians.