Evaluating the Reproducibility of Physiological Stress Detection Models
summary
Ninety participants across four controlled studies using two sensors (Polar heart-rate and Empatica E4) test cross-study stress models. Models trained on one dataset often fail on another because classification thresholds vary widely and distort precision and recall. Methods that account for temporal prior stress improve F1 consistently. An unsupervised clustering approach fits per-participant probabilities to two centers at 0 and 1 to choose thresholds without test labels and outperforms training-set thresholds and often matches optimized test thresholds.
AI-written summary · about
related talks
06mPredicting Subjective Measures of Social Anxiety from Sparsely Collected Mobile Sensor Data
22mA Classification Model for Sensing Human Trust in Machines Using EEG and GSR
32mIrina Rish, Research Staff, IBM T.J. Watson Research Center
01h 01mModel validation using deep generation of stress data
05mMaking Sense of Sleep: Multimodal Sleep Stage Classification in a Large, Diverse Population ...
35mML: Beyond confusion matrix, model performance analysis
27mHARSHA URLAM & SWETA SINGH at Cypher 2017
53mHumAInity Innovation Challenge Finals 2020