Mental illness is a kind of health condition highly impacts a person's emotions, mind and their behavior. When mental disorder is not diagnosed at its earlier condition it results in depression, anxiety, schizophrenia, autism etc., are extremely widespread today. Recently many machine learning methods have been developed to help experts such as psychologists' and psychiatrists' for making decision regarding mental health of patient based on historical data of the patients. However conventional machine learning methods require optimal performance of feature engineering to obtained better performance and reduce computation complexity. This paper focuses on developing improved multilayer perceptron its objective is to handle over fitting, fine-tuning parameter values by integrating optimized feature selection and adopting mimetic meta heuristic model. LASSO (Least Absolute Shrinkage and Selection Operator) model-based feature selection is used for discovering best feature subset because it controls over fitting and prediction error with the help of ridge regression. In improved MLP instead of using trial and error based random parameter value assignment it inherits the Intelligence of Chicken Swam behavior. The best fittest value is assigned to the parameters of MLP (Multi-Layer Perceptron) to improve the accuracy rate of mental illness disorder detection. The simulation results are conducted on OSMI 2019 dataset and the outcome proved that CSIMLP accomplished highest rate of accuracy in early detection of mental illness disorder compared to other conventional models.