People often express thoughts and emotions through text, which can include sarcasm. Sarcasm detection is a difficult task because of noisy social media labels and a lack of high-quality datasets. It is necessary to identify sarcastic texts that affect mental health. The proposed novel method, Convolutional Neural Network and Long-Short Term Memory CNN+LSTM, which utilizes selectively frozen LSTM input layers, enhances sarcasm recognition to identify potential signs of depression, offering a unique approach for more precise mental health evaluation. For identification of sarcastic text, the News headline dataset and Reditt dataset were used together. For identification of sarcastic text, datasets were taken from Kaggle. To evaluate the efficiency of different approaches on sarcastic text, the suggested hybrid approach was compared with state-of-the-art methods like machine learning ML and deep learning DL. The ML and DL methods like logistic regression, decision tree, support vector classifier, multinomial Naïve Bayes, XG Boost, CNN, and LSTM were compared. The hybrid dataset was applied to all ML and DL methods. The proposed method outperformed, with an accuracy of 84%, recall of 83%, precision of 84%, and F1 score of 83% on the hybrid dataset, which is higher than the accuracy obtained from traditional ML methods. The hybrid method is well suited for the complex text. It performs well for detecting sarcastic language features.