Hurricane Sandy was one of the deadliest and costliest of hurricanes over the past few decades. Many states experienced significant power outage, however many people used social media to communicate while having limited or no access to traditional information sources. In this study, we explored the evolution of various communication patterns using machine learning techniques and determined user concerns that emerged over the course of Hurricane Sandy. The original data included ~52M tweets coming from ~13M users between October 14, 2012 and November 12, 2012. We run topic model on ~763K tweets from top 4,029 most frequent users who tweeted about Sandy at least 100 times. We identified 250 well-defined communication patterns based on perplexity. Conversations of most frequent and relevant users indicate the evolution of numerous storm-phase (warning, response, and recovery) specific topics. People were also concerned about storm location and time, media coverage, and activities of political leaders and celebrities. We also present each relevant keyword that contributed to one particular pattern of user concerns. Such keywords would be particularly meaningful in targeted information spreading and effective crisis communication in similar major disasters. Each of these words can also be helpful for efficient hash-tagging to reach target audience as needed via social media. The pattern recognition approach of this study can be used in identifying real time user needs in future crises.