As stress is linked to numerous emotional and physical conditions, its timely detection and proper management is important for our health. Convolution neural network (CNN) has been shown to be promising in stress detection because it could automatically capture the discriminant information regarding physiological change from heart rate variability (HRV), usually derived from electrocardiogram (ECG) signals. This study proposed a two-step training method to improve the acute stress detection performance through optimizing the probability threshold of a CNN. The results showed that the average error rate was significantly reduced from 17.3 +/- 9.2% to 9.2 +/- 5.7% after probability threshold optimization, and the classification results were more balanced between stress and rest data. This study presented a simple method to improve stress detection performance using CNN without additional data, rendering benefits for the practical application of HRV-based stress measurement.