Integral membrane proteins play important functional roles in biological systems. In order to understand the mechanism of their functions, knowledge of their 3D structures is essential. However, due to inherent experimental difficulties, the 3D structures of only a limited number of integral membrane proteins are known. In the absence of 3D structure, prediction of transmembrane (TM) helix topology is extremely important. Although the performances of currently available TM helix prediction methods are high, neither can they predict all TM helices nor are they free of over-prediction. A clear understanding of the failures and successes of the current TM helix prediction methods is necessary before better prediction methods can be proposed. In this work we perform an analysis of a known set of TM helices in proteins in terms of their average hydropathy values, amino acid preferences and lipid accessibility. Specifically we focus on TM helices that are correctly predicted and those that are missed by typical TM helix prediction methods. Compared to correctly-predicetd helices, difficult-to-predict (missed by prediction methods) helices are found to be mostly hydrophilic and they show unique amino acid propensity - over-representation of Trp, Tyr and Phe, and, under-representation of Glu, Asp and Lys. However, the correctly-predicted as well as difficult-to-predict helices show no correlation with lipid accessibility. Our results will help understand the limitations of the current TM helix prediction methods and will be useful in proposing new TM helix prediction strategies.