Multimodal stress recognition using facial thermography and pulse oximetry: de-identified dataset and analysis code

De-identified numerical dataset and complete analysis code for the study"Multimodal stress recognition using facial thermography and pulse oximetry". Facial thermography and fingertip pulse oximetry were recorded from 48 adultsduring a single fixed-order laboratory session (baseline, neutral speech,Stroop, mental arithmetic with speaking, recovery). 43 participants remainafter quality-control and protocol exclusions, giving 10,280 analysed 10-secondwindows. The deposit contains per-frame thermal region-of-interest summaries, pulse-rateand SpO2 readings, participant and phase metadata, questionnaire responses,derived window features, participant-disjoint fold definitions, per-windowpredictions, aggregate results, and the full pipeline needed to reproduce thereported models, control analyses, statistical tests, tables and figures. No thermal or visible face imagery and no raw photoplethysmography waveformsare included, so heart-rate-variability indices cannot be derived. Because theprotocol ran in a fixed order for every participant, elapsed session time isstrongly confounded with the stress label: elapsed time alone classifies it at0.997 balanced accuracy. Phase labels are protocol-level targets, not clinicalground truth. Data are licensed CC BY 4.0; code under code/ is MIT.