Many children announce their arrival in the delivery room with a sharp cry. When a newborn automatically takes its first breath, its lungs expand, its blood vessels dilate, and its entire circulatory system reorganizes for life outside the womb. However, this process does not always go as planned.

Some infants, especially those who are seriously ill or born prematurely, have pulmonary hypertension. Pulmonary hypertension is a serious disease in which the arteries to the lungs remain narrow after birth or become blocked again in the first days or weeks after birth. This restricts blood flow to the lungs and reduces the amount of oxygen in the blood.

Severe cases of pulmonary hypertension must be detected and treated as quickly as possible. The sooner treatment begins, the better the newborn’s prognosis. However, making the correct diagnosis can be difficult. Only an experienced pediatric cardiologist can diagnose pulmonary hypertension based on a comprehensive ultrasound examination of the heart.

“Detecting pulmonary hypertension is time consuming and requires cardiologists with very specific expertise and years of experience, skills that tend to be available only in the largest pediatric clinics. “Yes,” says Professor Sven Wellman, Medical Director of Neonatology at KUNO. St. Hedwig’s Clinic, part of the Hospital of the Knights of St. John, in Regensburg, Germany.

Researchers from a group led by Julia Vogt, who runs the Medical Data Science Group at ETH Zurich, recently collaborated with neonatologists from the KUNO Clinic St. Hedwig to provide reliable support for diagnosing diseases in newborns. We developed a computer model to do this. Their results are now International Journal of Computer Vision.

Making AI reliable and explainable

The ETH researchers started by training their algorithm on hundreds of video recordings from heart ultrasounds of 192 newborns. This dataset includes videos of a beating heart taken from various angles, as well as diagnoses by experienced pediatric cardiologists (presence or absence of pulmonary hypertension) and disease severity (from “mild” to “moderate”). It also includes an assessment of severity (“severity”).

To determine whether the algorithm was successful in interpreting the images, the researchers then added a second dataset of ultrasound images from 78 newborns that the model had never seen before. . The model suggested the correct diagnosis in approximately 80% to 90% of cases and was able to determine the correct level of disease severity in approximately 65% ​​to 85% of cases.

“The key to using machine learning models in medical settings is not just their predictive accuracy, but also whether humans can understand the criteria the models use to make decisions,” Vogt said.

Her model allows this by highlighting parts of the ultrasound image that are the basis for classification. This allows doctors to see exactly which areas or characteristics of the heart and blood vessels the model has determined to be suspicious. Pediatric cardiologists examined the dataset and found that the model was looking at the same features as them, even though it was not explicitly programmed to do so.

Diagnosis is done by humans

This machine learning model could potentially be extended to other organs and diseases, such as diagnosing cardiac septal defects and valvular heart disease.

It may also be useful in areas where there are no experts. Standardized ultrasound images are taken by medical professionals, and the model can provide a preliminary risk assessment and an indication of whether a specialist should be consulted. Healthcare facilities with access to highly qualified professionals could use this model to reduce workload and help achieve better and more objective diagnoses.

“AI has the potential to significantly improve healthcare. The key issue for us is that the final decision should always be made by a human, a doctor. We just need to provide support to make sure that as many people as possible get the best care possible,” Vogt said.