Emotion classification of social media text, particularly from X (formerly Twitter), presents unique challenges due to its short, informal, and noisy nature. Traditional machine learning and deep learning models often struggle to achieve high accuracy on such data due to high variability, misspellings, and contextual ambiguity. This study aimed to enhance the performance of Long Short-Term Memory (LSTM) models for emotion classification by optimizing hyperparameters using metaheuristic algorithms. The Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Firefly Algorithm were leveraged to maximize the F-score and accuracy, critical metrics for classification tasks. Utilizing a dataset of 23,000 tweets related to a cholera outbreak, LSTM models were trained and optimized to surpass previous accuracy levels. Through experimentation, the Firefly Algorithm-optimized LSTM model emerged as the best-performing approach, achieving 98% accuracy, compared to the baseline LSTM model at 76%, GA-optimized LSTM at 89%, and PSO-optimized LSTM at 90%. The Distributed Evolutionary Algorithm (DEAP) framework was employed to implement the evolutionary algorithms, ensuring flexibility and scalability. This study demonstrates that metaheuristic optimization significantly improves LSTM performance for emotion classification, addressing the unique challenges posed by noisy social media text.