Automatic detection of depression in the elderly through Facial Expression Recognition faces a fundamental challenge in the form of domain shift due to skin deformation and facial structural changes due to aging, such as ptosis and deep wrinkles. This study proposes a Two-Phase Transfer Learning framework that integrates high-density facial landmark point extraction (468 points using MediaPipe) with a hybrid spatiotemporal CNN-BiLSTM-VGG19 architecture to address these challenges. Phase I training was conducted on a standard facial dataset to obtain fundamental feature representations, followed by a fine-tuning process in Phase II using a geriatric facial dataset. Experimental results show that the CNN-BiLSTM-VGG19 architecture is highly robust, exploiting deep facial wrinkles as informative texture features. The model successfully achieved 91.42% accuracy on 70-year-old older adults. Furthermore, hyperparameter evaluation confirmed that the Stochastic Gradient Descent (SGD) optimizer combined with a low learning rate of 0.0005 was the most optimal configuration. This balance effectively prevented catastrophic forgetting during domain adaptation, while also achieving a clinical sensitivity recall rate above 96%. Comprehensively, this study demonstrates that the texture-biased CNN-BiLSTM-VGG19 model offers a robust, non-invasive, and highly efficient depression screening instrument for implementation in elderly care facilities.
Wibisono et al. (Thu,) studied this question.