This thesis demonstrates the potential and benefits of unsupervised learning with Self-Organizing Maps for stress detection in laboratory and free-living environment. The general increase in pace of life, both in the personal and work environment leads to the intensification and amount of work, constant time pressure and pressure to excel. It can cause psychosocial problems and negative health outcomes. Providing personal information about one’s stress level can counteract the adverse health effects of stress. Currently the most common way to detect stress is by the means of questionnaires. This is time consuming, subjective and only at discrete moments in time. Literature has shown that in a laboratory environment physiological signals can be used to detect stress in a continuous and objective way. Advances in wearable technology now make it feasible to continuously monitor physiological signals in daily life, allowing stress detection in a free-living environment. Ambulant stress detection is associated with several challenges. The data acquisition with wearables is less accurate compared to sensors used in a controlled environment and physical activity influences the physiological signals. Furthermore, the validation of stress detection with questionnaires provides an unreliable labelling of the data as it is subjective and delayed. This thesis explores an unsupervised learning technique, the Self-Organizing Map (SOM), to avoid the use of subjective labels. The provided data set originated from stress-inducing experiments in a con- trolled environment and ambulant data measured during daily-life activities. Blood volume pulse (BVP), skin temperature (ST), galvanic skin response (GSR), electromyogram (EMG), respiration, electrocardiogram (ECG) and acceleration were measured using both wearable and static devices. First, a supervised learning with Random Decision Forests (RDF) was applied to the laboratory data to provide a gold standard for unsupervised learning outcomes. A classification accuracy of 83.04% was reached using ECG and GSR features and 76.89% using ECG features only. Then the feasibility of the SOMs was tested on the laboratory data and compared a posteriori with the objective labels. Using a subset of ECG features, the classification accuracy was 76.42%. This is similar to supervised learning with ECG features, indicating the principal functioning of the SOMs for stress detection. In the last phase of this thesis the SOM was applied on the ambulant data. Training the SOM with ECG features from the ambulant data, enabled clustering from the feature space. The clusters were well separated with large cohesion (average silhouette coefficient of 0.49). Moreover, the clusters were similar over different test persons and days. According to literature the center values of the features in each cluster can indicate stress and relax phases. By mapping test samples on the trained and clustered SOM, stress predictions were made. Comparison against the subjective stress levels was however poor with a root mean squared error (RMSE) of 0.50. It is suggested to further explore the use of Self-Organizing Maps as it solely relies on the physiological data, excluding subjective labelling. Improvements can be made by applying multimodal feature sets, including for example GSR.