Deep Learning Model Classifies and Detects Heart Disease

Researchers have developed an AI model that uses chest X-rays to simultaneously detect cardiac functions and valvular disease, according to a paper submitted to The Lancet Digital Health. It may be the first study to develop and validate a deep learning-based classification model using digital x-rays from multiple institutions to prevent 'overfitting,' or errors in interpreting new data.

The model, developed by scientists at Osaka Metropolitan University, can categorize cardiac functions and heart diseases faster than traditional methods, streamlining the diagnostic process. The platform could improve the efficiency of cardiac diagnoses, support traditional echocardiography, and help fill in when specialized technicians aren’t readily available. 

The scientists used multi-institutional data collected from 22,551 chest radiographs associated with 22,551 echocardiograms from nearly 17,000 patients. The mean age was 72 years old with the gender evenly divided. The dataset was collected from three facilities between 2013 and 2021 for training, validation, and internal testing. External testing was conducted at a fourth site. The chest radiographs served as the input data and the echocardiograms were used as the output data. The AI model was trained to associate features connecting both datasets.

Stay updated. Subscribe to the AI Business newsletter.

The deep-learning model was able to accurately categorize six types of valvular heart disease, which can help detect cardiovascular issues. Pulmonary regurgitation, mitral regurgitation, mitral stenosis, aortic stenosis, aortic regurgitation, and inferior vena cava dilation were distinct, based on the digital x-rays.

The team developed a multilabel deep-learning model using EfficientNet as a feature extractor. Seventeen labels were chosen as classifiers. Each classifier was connected to a SoftMax activation function, followed by a cross-entropy loss function. The model was trained on the basis of ImageNet pretrained parameters and tuned with the training dataset using five-fold cross-validation.

Results of the AUC (Area Under the Curve) - a key metric of classification model performance - were promising, especially for detecting left ventricular ejection fraction. 

“In addition to improving the efficiency of doctors’ diagnoses, the system might also be used in areas where there are no specialists, in night-time emergencies, and for patients who have difficulty undergoing echocardiography,” said the study’s lead author Dr. Daiju Ueda, from the Department of Diagnostic and Interventional Radiology at the Graduate School of Medicine of Osaka Metropolitan University.