Although heart failure with reduced ejection fraction (HFrEF) can benefit from treatment, it remains underdiagnosed due to the limited accessibility of echocardiography. Several studies have shown that artificial intelligent (AI) models of 12 leads ECG can successfully identify ventricular dysfunction. However, such models are hard to deploy for frequent surveillance of the left ventricular ejection fraction (LVEF), especially under the home-based scenario. We aimed to design a prospective study that used a single-lead 30-second ECG recorded and operated by patients to transfer and test an AI model constructed with convolutional neural network (CNN) analysis of standard 12-lead ECG from a retrospective single medical center cohort. In addition, we also try to leverage the short-term heart rate variability (HRV) to improve the sensitivity and specificity of the model. The 12-lead ECGs of sinus rhythm with LVEF measuring within 90 days of the recording date were included. The dataset was split into training and validation sets at a ratio of 7:3 without patient overlapping. To select feasible models of single-lead ECG for identifying LVEF of 45% or lower, we trained the CNN models using single-lead ECG recorded from each lead placement as input. The models with the top 3 performances were transferred to the corresponding lead of the single-lead ECG datasets acquired prospectively to validate their generalizability. Among all single-lead ECG models trained from 56479 recordings of 33010 patients, the AUC of the top 3 ranked models (lead aVR, I, and V5) were 0.89, 0.88, and 0.88, respectively. While aVR is impossible to derive from single-lead placement, we applied the models of lead I, V5, and lead II instead. For the prospective single-lead ECG datasets, 144 patients with an average age of 64 years old were recruited, and the models were tested. No significant drop in the performance of the lead I model was shown (AUC=0.88). Furthermore, adding HRV to the lead I model can improve overall accuracy (AUC=0.92) and sensitivity and specificity of the optimal cutting threshold. With the assistance of AI algorithm, single-lead ECG recorded and operated by patients from a wristband performed well in detecting low LVEF of 45% or lower. This simple approach can provide an inexpensive and convenient tool for the home-based monitoring of LVEF, and the tool's effectiveness warrants further study.