Using deep learning to diagnose pulmonary hypertension

This editorial refers to ‘A framework of deep learning networks provides expert-level accuracy for the detection and prognostication of pulmonary arterial hypertension’, by G.-P. Diller et al., https://doi.org/10.1093/ehjci/jeac147. Pulmonary hypertension is defined as an increase in mean pulmonary artery pressure ≥25 mmHg at rest, as assessed by right heart catheterization.1 More recently, at the 5th World Symposium on pulmonary hypertension in 2018, it was proposed that the threshold should be lowered to >20 mmHg, based on population studies demonstrating this to be the upper limit of normal and values above 20 mmHg decreasing life expectancy.2,3 Pulmonary arterial hypertension, on the other hand, comprises a very specific group of patients with pulmonary hypertension, namely those with increased pulmonary vascular resistance (>3 Wood units). While this threshold was not altered at the above-mentioned symposium, the upper limit of normality has been shown to be 2 Wood units, whereas higher values decrease survival.1, 2 Direct haemodynamic measurements cannot be performed in all patients due to the invasive nature of the procedure. The diagnosis of pulmonary hypertension therefore usually relies on non-invasive tests, such as echocardiography. Transthoracic echocardiography has only modest sensitivity and specificity for the diagnosis of pulmonary hypertension, and pulmonary artery pressures can be readily over- or underestimated.1,4,5 Consequently, contemporary guidelines recommend that only a probability of pulmonary hypertension be reported on the basis of an echocardiographic examination.1 This probability-based recommendation is predicated not only on the actual, echocardiographically estimated pulmonary artery pressure, but includes a number of indirect signs supporting the presence of pulmonary hypertension.1,2 The symptoms and signs of pulmonary hypertension are often non-specific, but early diagnosis is paramount, since: (i) the condition usually has a progressive course and impacts negatively on survival and (ii) it can often be treated successfully if recognized early.1,6 Echocardiographic screening for pulmonary hypertension requires not only a fair amount of expertise, but also vigilance, and strategies to improve the detection rate in a busy echocardiography laboratory may have clinical utility. Artificial intelligence and ‘deep learning’ (which is a subdivision of machine learning, where artificial neural networks are constructed to model data) have the potential to improve echocardiographic diagnoses by identifying and analysing an array of pertinent imaging parameters.7 Whether deep learning can improve the detection rate of pulmonary hypertension, especially in echocardiography laboratories where expertise in diagnosing this condition is less readily available, is unknown.