With the intensification of population aging, cardiovascular diseases (CVDs) have become a concern for people [1]. The development of mobile health technology has effectively addressed the conflict between healthcare resources and medical demands. However, due to considerations, such as portability and battery life, portable electrocardiogram ECG monitors often have a limited number of lead channels. Some researchers have designed monitoring systems with a limited number of leads [2]. CVDs rely on the standard 12‑lead ECG for diagnostic interpretation. This reliance creates a contradiction between portable design and clinical importance, highlighting the need to explore methods for recovering 12‑lead ECG data from a limited number of lead signals [3,4].

Relevant studies [5,6] have demonstrated the vector space characteristics of ECG physiological signals and provided a theoretical basis for reconstructing three‑lead signals into 12‑lead signals. Lead reconstruction is a regression problem, and the methods commonly used to address it include linear regression, which is computationally straightforward [7]. Other relevant studies [8,9] have shown that nonlinear methods yield better reconstruction results. Support vector machines [10], regression trees [11], and deep learning networks [[12], [13], [14], [15], [16]] have been applied in ECG lead reconstruction.

To meet the fundamental requirements of ECG vector space, we employ leads I, II, and V2 for lead reconstruction Limb leads can be reconstructed perfectly due to their inherent linear relationships on the frontal plane. The focus of this study is on using neural networks to reconstruct precordial leads.