Atrial fibrillation (AF) is largely underdiagnosed. Previous studies using deep neural networks with large datasets have shown that screening AF with a 12-lead electrocardiogram (ECG) during sinus rhythm (SR) is possible. However, the poor availability of these trained models and the small size of the retrievable datasets limit its reproducibility. This study proposes an approach to generate explainable features for detecting AF during SR with limited data. We collected 94,224 12-lead ECGs from 64,196 patients from Taipei Medical University Hospital. We selected ECGs during SR from 213 patients before AF diagnosis and randomly selected 247 age-matched participants without AF records as the controls. We developed a signal-processing technique, MA-UPEMD, to isolate P waves, and quantified the spatial and temporal features using principal component analysis and inter-lead relationships. By combining these features, the machine learning models yielded AUC of 0.64. We showed that, even with this limited dataset, the P wave, representing atrial electrical activity, is depicted by our proposed approach. The extracted features performed better than the bandpass filter-extracted P waves and deep neural network model. We provided a physiologically explainable and reproducible approach for classifying patients with AF during SR.