Machine-Learning Tool Accurately Detects Environmental Enteropathy, Celiac Disease

A machine-learning algorithm achieved a 93.4% accuracy in detecting environmental enteropathy and celiac disease in children, according to the findings of a study published in JAMA Network Open.

The algorithm had a false-negative rate of 2.4% and automatically learned microlevel features in duodenal tissue.

“If we can use these cutting-edge technologies and ways of looking at data through data science, we can get answers faster and help these children sooner,” said Sana Syed, M.D., an assistant professor of pediatrics at the University of Virginia School of Medicine.

Researchers obtained more than 3,100 segmented images from 121 hematoxylin-eosin-stained duodenal biopsy glass slides from 102 patients. Primary study physicians labeled the slides as environmental enteropathy, celiac disease or control based on histological and clinical findings.

The research team collected the environmental enteropathy slides from Aga Khan University Hospital in Karachi, Pakistan (29 slides, 10 patients) and the University of Zambia Medical Center (16 slides, 16 patients). The researchers had 34 celiac disease slides from 34 patients and 42 control group slides from 42 patients that came from the Biorepository and Tissue Research Facility at the University of Virginia.

The researchers also collected biomarkers from blood, urine and fecal samples of patients with environmental enteropathy from Pakistan and Zambia. The research team used the biomarkers to purpose an algorithmic framework to correlate the numerical metadata with biopsy features. But since the researchers obtained limited and variable biomarkers, biological inferences could not be made from their results.

So the research team developed a correlation algorithm to test in a larger and more comprehensive data set.

The researchers proposed a convolutional neural network based on AlexNet to classify biopsy images. A deep neural network model trained on the entire data set.

The models achieved 92.1% cross-validation per-image prediction accuracy and 93.4% per-patient accuracy. Overall, the model had a false-negative rate of 2.4%.

“The machine-learning algorithm can provide insights that have evaded human eyes, validate pathologists’ diagnoses and shorten the time between imaging and diagnosis…” wrote Wende Hope, communications manager of the mechanical and aerospace engineering department at University of Virginia.

Get the best insights in digital health directly to your inbox.

Related

Machine Learning Accurate in Identifying Patients at Risk of Developing Psychosis

Risk Prediction Model Helps Identify Risk of Developing Lung Cancer

AI Model Analyzes EHR Data, Identifies Child Leukemia Patients for Studies