This work presents a new method for detection as well as early prediction of sepsis using ratio based features that are derived from the given patient data. The ratios of parameters like length, area and volume are found to be well suited for anomaly detection. Thus, the proposed method begins with patient data preprocessing such as normalization and computes the ratio based features from the selected patient co-variates.In all, 17 features are computed and subsequently they are fed along with given 40 signs to train the two classifier models. The proposed scheme uses one model that captures characteristics for early prediction of sepsis in hospitalised subjects while the other model is meant for detecting sepsis in patients that are already having sepsis. The system combines the optimally trained multivariate Gaussian model for detection and ensemble model of random undersampling (RUSBoost) for early prediction. Genetic algorithm has been used to find the optimal parameter for the multivariate Gaussian model. When evaluated with 2019 PhysioNet/CinC Challenge dataset, the experimental outcomes in Matlab demonstrate the utility score of 40% for 10-cross validation on given training data.