Model Evaluation Improvements for Multiclass Classification in Diagnosis Prediction

We are living in an age in which we are invaded by the amount of available data. These data are increasing in an exponential way. The art of making sense of all the data represent an issues nowadays. Moreover, the ability to deal with different types of these data require new approaches in the field of exploratory analysis. Therefore the extraction of relevant information, the discovery of relations between data and the ability to generalize to new data represent a continuous challenge. Exploratory data analysis becomes an impressive area of concern for certain domains such as education, healthcare, biology, economics, geography, geology, history or agriculture. Particularly, the purpose of this paper is related to medicine and psychology. Some machine learning advantages are being investigated in order to improve a treatment, a diagnosis of a patient. This paper, presenting a work in progress, discusses an approach to a relevant supervised learning method from the art of machine learning field: classification. Various aspects are considered, as preprocessing of the input data; selection of the model applied to the data; evaluation of the model; improving the performance of a model, selection of the most relevant features to be included in the model and also learning a model that is able to perform well on new data [1]. The computed metrics for performance evaluation of a model are also highlighted. The data sets (mixed data) used in our analysis are data from medical field (kidney and lung disease: pulmonar-renal syndrome) and also are suitable for multiclass classification. In this paper, the selected models are ensembles of decision trees such as Random Forest and Gradient Boosted Regression Trees. The model evaluation, the model improvements and feature selection ultimately lead to building models able to generalize to new data with a high value of accuracy. All these represent an added value in fields such us medicine and psychology, where a physician or a psychologist may use pattern and information as input in the treatment of a patient.