The relationship between premature atrial complexes (PACs) and cardiovascular diseases (CVD) remains unclear, and current data-driven PAC detectors, based on beat classification, achieve low sensitivity hindering PACs monitoring research. Moreover, the results interpretability of these PAC detectors is not addressed, resulting in reduced usability and hampering the adoption by cardiology experts. Explainable AI (XAI) techniques allow the unveiling features' relevance that influence PAC detection, thereby improving the trustworthiness of decision-making in CVDs diagnosis. This study shows the explainability analysis of a PAC detector developed by the authors (avg accuracy 0.985, avg sensitivity 0.954, avg specificity 0.981, avg PPV 0.627) that uses random forest for multiclass classification of Normal (N), Supraventricular (S), and ventricular (V) beats based on ECG-extracted heart rate variability (HRV) and morphological features. We employ SHAP (SHapley Additive exPlanations) as an XAI technique to assess the influence of the different features with the ultimate goal of building surrogate simpler models that enhance the interpretability of PAC detectors. The explainability results indicate that features related to the RR intervals are the most influential in predicting N and S classes. Furthermore, morphological features extracted from QRS complex along with beats information make the highest contribution for prediction class V. By training random forest and decision tree classifiers on these features, we intend to build surrogate models using only 14 out of the original 85 features. The surrogate random forest experiences a significant decrease in performance only for class S detection in PPV of 0.063 (p<0.05), while the surrogate decision tree decreases significantly (p<0.05) the NPV for class N by 0.23, the sensitivity and PPV for class S by 0.10 and 0.19 respectively, and the PPV for class V by 0.22. We demonstrate a PAC surrogate fully explainable model could be derived from the decision tree despite a minor performance loss.