The expression of human epidermal growth factor receptor-2 (HER2) has important implications for pathogenesis, progression, and therapeutic efficacy of breast cancer. The detection of its variation during the treatment is crucial for therapeutic decision-making but remains a grand challenge, especially at the cellular level. Here, we develop a machine learning-driven surface-enhanced Raman spectroscopy (SERS)-integrated strategy for label-free detection of cellular HER2. Specifically, our method allows the extraction of cell-rich spectral signatures utilized for identification and classification of cancer cells with distinct HER2 expression with a high accuracy of 99.6%. By combining label-free SERS detection and machine learning-driven chemometric analysis, we are able to perform longitudinal monitoring of therapeutic efficacy at the cellular level during the treatment of HER2+ breast cancer, which aids in the subsequent decision-making and management. This work provides a promising technique capable of performing dynamic label-free spectroscopic detection for therapeutic surveillance of diseases.