Emotion recognition to support personalized therapy in the elderly: an exploratory study based on CNNs

Emotions are pivotal in human life, and the ability to recognize and interpret emotions accurately holds significant value, particularly for populations such as the elderly who face challenges in expressing their emotions. Emotion recognition systems serve as valuable tools in various scenarios, including the customization of therapies, with the potential to enhance their effectiveness and precision.

To contribute in this context, we present a approach based on traditional Convolutional Neural Networks for emotion recognition in facial expressions. Methods: The proposed CNN model is composed of six convolu- tional layers and was trained, validated and tested with a comprehensive and robust dataset, formed by combining and joining the FER-2013, Chicago Face, Yale Face and KDEF databases. After model training, we apply Haar Cascade to static images of elderly people to identify the face and our CNN model to classify emotions.

The result obtained during the training/validation stage was 90.77% accuracy, demonstrating robustness with a diverse data set. However, accuracy dropped to 69.05% during the testing stage, aligning with the literature on generalization challenges. Despite this, sensitivity and specificity remained strong at 0.8882 and 0.9823 respectively, indicating reliable detection of true positives and negatives. When comparing this performance with studies in the lit- erature, our model stands out for having better results. Furthermore, the proposed model successfully classified emotions in images of elderly subjects.

The outcomes of our study are highly promising. The development of intel- ligent emotion recognition systems for the elderly represents a viable alternative for enhancing the quality of life for this population and their support networks.