Occupational stress is increasingly present in our society. Usually, it is detected too late, resulting in physical and mental health problems for the worker, as well as economic losses for the companies due to the consequent absenteeism, presenteeism, reduced motivation or staff turnover. Therefore, the development of early stress detection systems that allow individuals to take timely action and prevent irreversible damage is required. To address this need, we investigate a method to analyze changes in physiological and behavioral patterns using unobtrusively and ubiquitously gathered smart office data. The goal of this paper is to build models that predict self-assessed stress and mental workload scores, as well as models that predict workload conditions based on physiological and behavior data. Regression models were built for the prediction of the self-reported stress and mental workload scores from data based on real office work settings. Similarly, classification models were employed to detect workload conditions and change in these conditions. Specific algorithms to deal with class-imbalance (SMOTEBoost and RUSBoost) were also tested. Results confirm the predictability of behavioral changes for stress and mental workload levels, as well as for change in workload conditions. Results also suggest that computer-use patterns together with body posture and movements are the best predictors for this purpose. Moreover, the importance of self-reported scores' standardization and the suitability of the NASA Task Load Index test for workload assessment is noticed. This work contributes significantly towards the development of an unobtrusive and ubiquitous early stress detection system in smart office environments, whose implementation in the industrial environment would make a great beneficial impact on workers’ health status and on the economy of companies.