Detecting diseases from social media has diverse types of applications such as disease spread detection and health monitoring, etc. Traditionally, language models (LMs) have been adopted for this purpose and they have achieved significant performance on this task. However, there is still ongoing research on improving the discriminating representations of LMs. To this end, in this paper, we propose a novel fine-tuning method that unifies Contrastive Learning (CL) and language modeling. Specifically, we propose a self-augmentation method that augments hidden representations of the model with its own representations. The proposed method consists of two branches, where the first branch is a traditional LM learning the features for the given data whereas the second branch is augmented with first branch representations to learn features encouraging generalization. CL further improves representations by pulling original and augmented versions closer and pushing other samples away. Our approach improves performance over the traditional fine-tuning approach on 3 NLP datasets of binary, multi-label, and multi-class classification tasks consisting of social media posts related to various diseases. Notably, it improves the F1-score up to \(2.48\%\) on baseline and \(2.1\%\) on state-of-the-art methods.