Objective: During mechanical ventilation with a respirator, the establishment of an artificial airway can impair patients’ ability to expel secretions, easily leading to sputum accumulation. In current clinical practice, physicians primarily rely on auscultation of breath sounds to determine the need for sputum suction. This study proposes a transfer learning-based method for breath sound feature recognition and autonomous determination of sputum suction timing. Methods: An electronic stethoscope was used to collect breath sounds from the main airways of clinically ventilated patients before and after sputum suction, with the pre-suction sounds indicating the need for suction. The collected data underwent high-pass filtering and wavelet soft-threshold denoising, followed by the extraction of log-Mel spectrograms. A VGGish model pretrained on the Audio Set dataset was then employed to extract feature vectors from these spectrograms, which were subsequently classified using a support vector machine to determine whether suction was required. Results: Experimental results showed that the precision, recall and F1 score for recognition of breath sounds requiring sputum suction were 86.73%, 93.06% and 89.78%, respectively. Conclusion: The proposed breath sound recognition method based on transfer learning effectively determines the timing of sputum suction and shows significant clinical potential.