The rapid adoption of smartphones has created opportunities for understanding the emotional state of the user’s through smartphone interactions. This work presents a novel approach using Large Language Models (LLMs) and smartphone notification interaction data for the estimation of the affective state of the user, an area that has not been extensively studied. Recent studies face challenges in generating personalized predictions; instead, they deliver generalized outcomes that are not suitable for an individual’s behavior. Most models failed to incorporate real-time variables, which limits the accuracy and relevance of user’s emotional state predictions. This work applies the use of LLMs to analyze user’s complex, dynamic interactions with smartphone notifications, resulting in a greater understanding of emotions such as happiness, irritation, and frustration. Our work uses real-time smartphone notification data, “NotifyMiner,” and LLMs to deliver context-aware, personalized emotional state predictions. In our experiments, the LLM is trained on interaction data from user’s, which includes response latency, level of engagement, and type of notification, and its predictive capacity is assessed for recognizing various emotional states. We evaluate zero-shot and few-shot embedding techniques regarding their use in LLM to measure and predict user well-being status. The model proved to be successful in identifying various affective states, highlighting its potential for customized alerts, mental health tracking, and enhanced usability features. Through context-aware and personalized predictions, LLMs demonstrate their ability to overcome standard modelling restrictions and analyze complex user behaviors.