一種面向情緒壓力分布外檢測的多任務(wù)跨模態(tài)學習方法

打開文本圖片集

中圖分類號：TP391 文獻標志碼：A 文章編號：1001-3695（2025）06-018-1734-08

doi：10.19734/j. issn.1001-3695.2024.10.0457

Multi-task cross-modal learning approach for out-of-distribution detection of emotional stress

Wan Yichen 1，2 ， Xing Kai 1，2? ，Liu Yu3，Yang Hui4，Xu Junhan12，Yuan Yanxue 1，2 （20 （1.SchoolofomputerSiee &Tcholog，Uniersitofcince&TholfCina，Hefe6，Cha.SuzhostuefrAd vancedResearch，Unierstyofience&echologfina，SuzouJangsu5o，hina；3NnjingDrumowrHsptal，jing 210008，China；4.SchoolofLife Sciences，NorthwesternPolytechnical University，Xi’an71oo72，China）

Abstract：Recent research indicates that emotional stress detectionsystemsbasedon PPG signalscan bea potential convenient solution.However，PPG-based methods usuallyinduce severe OOD issues when detecting stressin previouslyunseen subjects duetosignificantvariations inPPGsignalsacross individuals.Toaddressthischallnge，thispaperproposedarossmodal stressdetection model basedonmulti-task learning（CSMT）.Byintroducing ECG signalreconstructionand multiple cardiovascular feature prediction asauxiliarytasks toenhancethefeatureextractioncapabilityofPPG signals，theproposed methodperformedcollaborativeoptimizationofPPG-based stress detectioninhigh-dimensionalrepresentationspace，thereby learning robuststressdetectionrepresentationsacrossindividuals.Experimentalresultsonthe WESADdatasetdemonstrate that inleave-one-subject-out validation tests，CSMT achieves best accuracy and F1 scores compared to existing methods in both thre-class （neutral/stress/amusement）andbinary（stress/non-stress）clasificationtasks，meanwhileefectivelymitigating theOODgeneralization probleminstressdetection.Theablation experimentsfurthervalidatetheefectiveness ofCSMTin enhancing model generalization capability.

Key Words：multi-task learning；photoplethysmography（PPG）； stress detection; out-of-distribution（OOD）issues

0 引言

情緒壓力是一種復雜的生理心理反應(yīng)，長期或慢性的情緒壓力會對人體的呼吸系統(tǒng)、心血管系統(tǒng)及內(nèi)分泌系統(tǒng)等產(chǎn)生嚴重的不良影響[1]，增加患焦慮癥、抑郁癥、心臟病以及注意力障礙等身心健康問題的風險[2]。（剩余20369字）