Road incidents exacerbate traffic congestion. Knowing where major incidents occur, and how severe they could be can help municipalities and their resources more efficiently. This paper presents a case study about using both structured and non-structured (textual) data for predicting road incidents at the City of Toronto. Road incidents can adversely affect the traffic and exacerbate road congestion. Using machine learning for predicting the probability of incidents, their severity and areas more prone to incident can help municipalities and departments of transportation (DOTs) improve road level of service while allocating their resources more effectively. In this paper, data of incidents in the City of Toronto was used to demonstrate how data analytics can help predict road incidents and their severity. The dataset included both SQL and NoSQL (written comments) data. A data fusion approach was used to merge these two types of data. Machine learning tree-based models were used to predict when and where major incidents occur on two major roads in the city: Gardiner Expressway and Don Valley Park (DVP). The developed models predicted the occurrence of major incidents, using solely SQL data, with an accuracy of 93%. After combing the SQL and NoSQL data for prediction, the model accuracy increased to 97%. Next, models were trained to predict the required time for incident scene clearance on the road. The accuracy of this model for predicting three classes was around 60%. After including the attributes extracted from the textual data, the accuracy increased to 81%. Eventually, several recommendations, especially related to standardized data collection, were made to improve future data collection and management in the City of Toronto. These recommendations can be useful to any municipality and/or DOT who wants to use data analytics in incident and traffic management.