Mental diseases are becoming an increasing issue in the world, and it is necessary to have consistent, strong, and non-invasive methods of monitoring the emotional state. Traditional evaluation procedures tend to be subjective, episodic and self-repor...
Mental diseases are becoming an increasing issue in the world, and it is necessary to have consistent, strong, and non-invasive methods of monitoring the emotional state. Traditional evaluation procedures tend to be subjective, episodic and self-reported, which restricts their efficacy in long-term mental health assistance. The current development of artificial intelligence (AI) has enabled the development of automated emotion recognition systems, but most methods lack robustness because they rely on a single modality, do not personalise to an individual, and cannot adapt to changing emotional patterns. These constraints limit their usage to real-world mental health monitoring settings. The proposed thesis is a flexible and lean multimodal deep learning model of emotion-sensitive mental health monitoring that combines facial expression recognition and speech-based emotion recognition. The multimodal design is based on the complementary relationship between the visual and acoustic signals, where the facial expression can give rich spatial information and the speech can give time and prosody-related emotional qualities. The acquisition of discriminative spatial representations of faces is done through Convolutional Neural Networks (CNNs), whereas the acquisition of time-varying relations in speech signals, including changes in pitch and intensity, is accomplished with the help of Long Short-term-memory (LSTM) networks. The characteristics of the two modalities are combined into a single emotional representation, less vulnerable to noise, modality one-sidedness, and data gaps.To facilitate personalization and long-term flexibility, the framework employs a reinforcement learning (RL)-based adaptive mechanism that dynamically adjusts modality contributions based on user-specific emotional patterns observed over time. It is a dynamic approach, which allows the system to adjust effectively to individual differences and incremental changes in emotion, which is critical in the long-term mental health monitoring. The proposed framework is evaluated by using common benchmark data of emotion recognition of faces and speech. The experimental findings indicate better performance than unimodal and non-adaptive baselines, with a maximum classification accuracy of 93\% and high precision and recall across most emotion categories. Moreover, there is a pragmatic concern like variation in the quality of data, cultural bias, and preservation of privacy that is explored to determine feasibility in the real world. Ethical considerations, such as the consent of the user and the responsible use of AI, are also highlighted to make sure that the system is more of a supportive monitoring tool than a diagnostic solution.