Standard 12-lead electrocardiography (ECG) is the primary tool for detection and diagnosis of cardiovascular diseases (CVDs). Most wearable ECG devices only provide single limb-lead measurement, limiting their practical use in CVD diagnosis. This study proposes a method of chest-lead ECG reconstruction from a single limb lead using a temporal convolutional network (TCN). The TCN is learned in the variational mode decomposition domain to reduce the non-stationary characteristics of ECG data. Experiments on two public databases suggested that automated diagnosis of CVDs in wearable ECG devices is likely to achieve through the proposed approach.