Obesity and malnutrition have psychological and physiological impacts on the health of older adults. Automated eating gesture detection has the potential to advance the assessment of their dietary intake activity. However, detecting eating gestures in older adults poses heightened challenges due to the variability in their eating behaviors. Some individuals exhibit slower eating habits, while others experience limitations in using one hand. Conversely, certain individuals conform to typical eating patterns observed in the adult population. To address this issue, we propose an automated eating gesture detection system using wrist and head mounted inertial measurement unit (IMU) sensors. An end-to-end approach is developed to detect and segment the time interval of eating gestures by employing a multi-stage temporal convolutional network (MS-TCN). Compared to existing eating gesture detection approaches, the present method is able to segment intervals of detected eating gestures more efficiently. We assess our methodology using one self-collected dataset containing 24 older adults, along with three publicly available datasets: FIC, OREBA, and Clemson. The leave-one-subject-out (LOSO) evaluation shows that our method achieves a segmental F1-score of 0.944 on our dataset. Furthermore, results on the FIC, OREBA, and Clemson datasets consistently indicate that our detection approach outperforms existing sliding window-based algorithms that combine convolutional neural networks and recurrent neural networks (CNN-RNNs).