Sentiment classification is a method of determining and classifying emotions represented in the text. Comprehending the underlying emotions of sensitive content, like suicide notes, can provide valuable insights into the mental health of an individual, making this task especially important. Efficient sentiment classification can help researchers, policymakers, and mental health practitioners devise more effective interventions and support plans. Despite the importance of this task, there are not enough datasets available to train models to identify different emotions in suicide notes. A variety of emotions are being represented in the dataset we are using. They are happy, sad, neutral, love, hate, and proud. In this study, we will train the model in three phases. First, we will extract features from the input sentences using Bidirectional Encoder Representations from Transformers (BERT). This will be followed by sequence modeling performed using the Bidirectional Long Short-Term Memory (Bi-LSTM) network. In the third phase, to emphasize more attention on the important segments of the sentence, we will use a Multi-head Attention mechanism. This would help increase the overall classification efficiency as it combines the strength of three different components. Resulting with Precision, Recall, and F1-Score all at 81%, this approach proves to be an effective one. This will help study the emotional content of the message written in suicide notes and will surely help in promoting initiative toward mental health in both personal and professional contexts.