How Deep Learning and AI Could Fuel the Next Phase of PAH Detection

Two decades ago, a team of French scientists set out to investigate a hard-to-diagnose and hard-to-treat population: patients with pulmonary arterial hypertension (PAH).

Though the disease is life-threatening and life-altering, it was also under-investigated. However, a national database of PAH patients was established in France in the year 2000, and six years later, the investigators were ready to report on the data.

The results made clear a major problem: The nearly 700 people in the registry routinely waited long periods of time—upwards of 27 months—from the time symptoms appeared until the time they were diagnosed with PAH. Once the diagnosis came, most patients had advanced disease. In fact, 75% of registry participants had New York Heart Association class III or IV disease at the time of diagnosis.

The findings underscored the importance of early detection, but they did little to improve early detection. That’s because while tools like chest x-ray, electrocardiogram (ECG), and lung function tests can help physicians identify likely cases of PAH, the gold standard in confirming diagnosis is right heart catheterization, a costly and invasive procedure.

In the years since that study, the therapeutic options for PAH have increased significantly. However, the diagnostic hurdles have remained. That has led new teams of scientists to turn to new tools—artificial intelligence (AI) and deep learning (DL)—in hopes of more accurately identifying patients with PAH.

A 2022 article in the European Heart Journal - Cardiovascular Imaging outlined one such effort. The investigators used 450 patients with PAH, 308 patients with right ventricular dilation without PAH, and 67 healthy controls to train deep convolutional networks to link echocardiographic images with estimated right ventricular systolic pressure in order to detect PAH.

“[W]e hypothesized that tailored DL algorithms would be able to assist with the diagnosis of PAH and provide prognostic information that compares favorably with traditional measurements obtained by expert echocardiographers,” wrote corresponding author Gerhard-Paul Diller, M.D., M.Sc., Ph.D., of the University of Münster, in Germany, and colleagues.

The resulting algorithm achieved an accuracy of 97.6% and a sensitivity of 100% in detecting PAH (on a per-patient basis). Moreover, the team said their model showed the potential to predict patients’ prognoses. However, they also cautioned that their model was not yet ready to replace human experts; instead, they said it could be used to initially screen patients.

This month, a new article in the Journal of Imaging Informatics in Medicine returned to the topic. This time, investigators used a combination of electrocardiography (ECG) and chest x-ray data to train a model to detect patients with elevated pulmonary arterial pressure (PAP).

They utilized data from two hospitals collected over the course of 11 years, parsing ECG and chest x-ray records from more than 30,000 patients each. They then tested the model internally and with an external data set.

The model achieved a negative predictive value (NPV) of 98% on the internal data set and 98.1% on the external data set. In addition to identifying patients with elevated PAP, the authors said their model also was able to predict patients’ risk of left ventricular dysfunction and cardiovascular mortality.

Corresponding author Wen-Hui Fang, M.D., of the Tri-Service General Hospital, in Taiwan, and colleagues, said use of their model could be an important screening tool.

“This model has the potential to be valuable in clinical settings for screening patients with pulmonary hypertension due to its high NPV, allowing for early intervention and improved long-term cardiovascular outcomes,” they wrote.

Fang and colleagues said using parameters like ECG and x-ray alone have been shown to be insufficiently effective at early detection, in part because ECG findings are an “unreliable” screening tool for elevated PAP.

If used in clinical practice, they said, the model could better narrow down which patients warrant more invasive testing, such as right heart catheterization, “potentially resulting in enhanced cardiovascular outcomes,” they said.

Fang and colleagues also cautioned, however, that the “opacity” of deep learning models could be a hurdle, since humans lack the same ability to identify patients with likely elevated PAP. They said further studies should be conducted to investigate the correlation and interpretability of the relationships identified by the models.