Parkinson's disease (PD) is the second most common neurodegenerative disease of central nervous system (CNS). Till now, there is no definitive clinical examination that can diagnose a PD patient. However, it has been reported that PD patients face deterioration in handwriting. Hence, different computer vision and machine learning researchers have proposed micrography and computer vision based methods. But, these methods possess two main problems. The first problem is biasedness in models caused by imbalanced data i.e. machine learning models show good performance on majority class but poor performance on minority class. Unfortunately, previous studies neither discussed this problem nor took any measures to avoid it. In order to highlight the biasedness in the constructed models and practically demonstrate it, we develop four different machine learning models. To alleviate the problem of biasedness, we propose to use random undersampling method to balance the training process. The second problem is low rate of classification accuracy which has limited clinical significance. To improve the PD detection accuracy, we propose a cascaded learning system that cascades a Chi2 model with adaptive boosting (Adaboost) model. The Chi2 model ranks and selects a subset of relevant features from the feature space while Adaboost model is used to predict PD based on the subset of features. Experimental results confirm that the proposed cascaded system shows better performance than other six similar cascaded systems that used six different state of the art machine learning models. Moreover, it was also observed that the proposed cascaded system improves the strength of conventional Adaboost model by 3.3% and reduces its complexity. Additionally, the cascaded system achieved classification accuracy of 76.44%, sensitivity of 70.94% and specificity of 81.94%.