Non-invasive ML methods for diagnosis of congenital heart disease associated with pulmonary arterial hypertension.

Objective: Congenital heart disease with pulmonary arterial hypertension (CHD-PAH), caused by CHD, is associated with high clinical mortality. Hence, timely diagnosis is imperative for treatment.

Approach: Two non-invasive diagnosis algorithms of CHD-PAH were put forward in this review, which were direct three-divided and two-stage classification models. Pre-processing in both algorithms focuses on segmentation of heart sounds into discrete cardiac cycles. Both the dual-threshold and Bi-LSTM (Bi-directional Long Short-Term Memory) methods demonstrate efficacy. In the feature extraction phase, the direct three-divided model integrate time-, frequency-, and energy-domain features with deep learning features. While the two-stage classification model sequentially extracts sub-band envelopes and short-time energy of cardiac cycle. In the classification phase, considering the lack of CHD-PAH data, ensemble learning was widely used.

Main results: An accuracy of 88.61% was achieved with direct three-divided model and 90.9% with two-stage classification model.

Significance: By analyzing and discussing these algorithms, future research directions of CHD-PAH assisted diagnosis were discussed. It is hoped that it will provide insight into prediction of CHD-PAH. Thus saving people from death due to untimely assistance.