Keywords
AI, Stress, Dataset, Participants, Process

Technologies
Python, RandomForest, XGBoost, SVM, GroupKFold

Stress is increasingly recognised as a serious health concern, yet most detection methods rely on self reporting which is slow, subjective and impractical in real world settings. This project replicates and extends existing research by Campanella et al., developing a machine learning pipeline that classifies stress from wristband sensor data (heart rate and skin conductance) collected from 29 participants, opening the door further to real-time stress detection.

Three main methodological extensions are introduced. First, the set of meaningful pieces of information extracted from the raw signals was increased from 27 features to 39. Second, the model is calibrated to each individual’s own resting baseline, rather than treating every individual's relaxed to stressed thresholds the same. Third, data from the same person never appears in both training and testing, eliminating a false inflation of results.

Using different sliding window configurations, the extended pipeline consistently improved performance, achieving accuracy of up to 89%, a 12.5 percentage point improvement over the original pipeline, suggesting that a reliable stress detection wearable is within reach.