Early and reliable detection of Atrial Fibrillation (AF) from short duration signals remains a significant clinical challenge. In this work, we demonstrate that the temporal dynamics of Block Term Decomposition (BTD) convergence serves as a robust biomarker. Using the vector representation of these parameters in ECG segments of very short duration (only 5 R-R intervals), our Bi-LSTM architecture achieved an F1-Score of 96.33%. This result establishes a new paradigm for AF detection, proving that an analysis focused on the trajectory of the decomposition process dispenses with complex attribute engineering and enables a fast and accurate diagnosis.