AI fares better than doctors at predicting deadly complications after surgery

A new artificial intelligence model found previously undetected signals in routine heart tests that strongly predict which patients will suffer potentially deadly complications after surgery. The model significantly outperformed risk scores currently relied upon by doctors.

The federally funded work by Johns Hopkins University researchers, which turns standard and inexpensive test results into a potentially lifesaving tool, could transform decision-making and risk calculation for both patients and surgeons.

"We demonstrate that a basic electrocardiogram contains important prognostic information not identifiable by the naked eye," said senior author Robert D. Stevens, chief of the Division of Informatics, Integration, and Innovation at Johns Hopkins Medicine. "We can only extract it with machine learning techniques."

The findings are published today in the British Journal of Anaesthesia.

A substantial portion of people develop life-threatening complications after major surgery. The risk scores relied upon by doctors to identify who is at risk for complications are only accurate in about 60% of cases.

Hoping to create a more accurate way to predict these health risks, the Johns Hopkins team turned to the electrocardiogram, or ECG, a standard, pre-surgical heart test widely obtained before major surgery. It's a fast, non-invasive way to evaluate cardiac activity through electric signals, and it can signal heart disease.

But ECG signals also pick up on other, more subtle physiological information, Stevens said, and the Hopkins team suspected they might find a treasure trove of rich, predictive data—if AI could help them see it.

"The ECG contains a lot of really interesting information not just about the heart but about the cardiovascular system," Stevens said. "Inflammation, the endocrine system, metabolism, fluids, electrolytes—all of these factors shape the morphology of the ECG. If we could get a really big dataset of ECG results and analyze it with deep learning, we reasoned we could get valuable information not currently available to clinicians."

The team analyzed preoperative ECG data from 37,000 patients who had surgery at Beth Israel Deaconess Medical Center in Boston.

The team trained two AI models to identify patients likely to have a heart attack, a stroke, or die within 30 days after their surgery. One model was trained on just ECG data. The other, which the team called a "fusion" model, combined the ECG information with more details from patient medical records such as age, gender, and existing medical conditions.

The ECG-only model predicted complications better than current risk scores, but the fusion model was even better, able to predict which patients would suffer post-surgical complications with 85% accuracy.

"Surprising that we can take this routine diagnostic, this 10 seconds worth of data, and predict really well if someone will die after surgery," said lead author Carl Harris, a PhD student in biomedical engineering. "We have a really meaningful finding that can can improve the assessment of surgical risk."

The team also developed a method to explain which ECG features might be associated with a heart attack or a stroke after an operation.

"You can imagine if you're undergoing major surgery, instead of just having your ECG put in your records where no one will look at it, it's run thru a model and you get a risk assessment and can talk with your doctor about the risks and benefits of surgery," Stevens said. "It's a transformative step forward in how we assess risk for patients."

Next the team will further test the model on datasets from more patients. They would also like to test the model prospectively with patients about to undergo surgery.

The team would also like to determine what other information might be extracted from ECG results through AI.

Authors, all from the Johns Hopkins School of Medicine and the Whiting School of Engineering, include Anway Pimpalkar, Ataes Aggarwal, Jiyuan Yang, Xiaojian Chen, Samuel Schmidgall, Sampath Rapuri, Joseph L. Greenstein, and Casey O. Taylor.

The work was supported by National Science Foundation Graduate Research Fellowship DGE2139757.