Early detection of mental health disorders plays an important role when it comes to treatment and patient outcomes. Traditional diagnostic methods include clinical interviews and self-report questionnaires, often suffer from biases and also lack scalability. Nowadays, machine learning techniques are offering an alternative approach by identifying suitable indicators of mental health conditions such as anxiety, depression and mood disorders. The approach uses diverse data types including speech, social media activity and psychological signals. This study reviews the effectiveness of machine learning models in detecting the symptoms of mental health disorders. Furthermore, an experimental case study using BERT is performed on a diverse dataset. Also, the potential of machine learning for future research is highlighted. The findings revealed that there is a need for more diverse datasets to enhance the model generalization across populations from diverse demographics, ultimately contributing to more accessible mental health diagnosis.