Classification models have proven useful for predicting clinical interventions and patient outcomes. One of the key issues that affect the predictive ability of supervised learning frameworks in the healthcare scenario is imbalance in data sets. In addition, non-uniform data collection processes in clinical scenarios lead to poor quality data sets. We designed a novel approach to predict Intensive Care Unit (ICU) transfers based on a weighted-similarity measure for patients outside of ICUs. The approach uses similarity between patient vital signs as input features for training the model. To address the data quality issues, we demonstrate the use of various up-sampling and down-sampling techniques to handle imbalanced data sets and train a classifier on a re-sampled data set. The data set used for testing the approach is derived from the MIMIC III database. We compare our results with the clinically accepted methodology to capture patient's health state, assisting in clinical decision making. Our model outperforms the standard methodology used in clinical decision making in standard scoring metrics such as F1-score, False Positive Rate and Mathew's Correlation Coefficient [MCC].