Chronic Low Back Pain (CLBP) afflicts millions globally, significantly impacting individuals’ well-being and imposing economic burdens on healthcare systems. Detecting protective behavior is essential for effective chronic pain management, as it can help prevent pain aggravation and disability. To reduce this burden, we could leverage sensor information and AI techniques to facilitate at-home patient follow-ups. Precisely, by utilizing motion sensors and surface electromyography (sEMG) sensors, we can continuously monitor movement patterns and muscle activity. This paper introduces L-SFAN, a lightweight convolutional neural network (CNN) architecture that innovatively models both spatial and temporal dimensions of multivariate time series to detect protective behavior. L-SFAN uses 2D CNN to capture spatial patterns from the 2D matrix formed by the multivariate time series and uses self-attention to capture long-range temporal dependencies. Temporal average pooling is used to emphasize spatial patterns. On the EmoPain dataset, L-SFAN outperforms state-of-the-art methods while reducing the number of parameters up to 94%, making it lightweight and embedded systems friendly. The ablation study underscores the importance of jointly modeling spatial-temporal information. The competitive performance and efficiency of our proposed method demonstrate its practicality for accessible chronic pain monitoring.