基於光體積變化描記圖個人特徵的壓力檢測模型之動態個人化

Abstract

近年偵測壓力的研究日益受到重視，然而大多著重於捕捉個體差異，無法泛化到未曾見過的樣本上，缺乏跨受試者的能力，因此，如何有效泛化至新樣本，成為重要的課題。本研究系統性模擬PPG資料，控制個體間變異、狀態內變異、狀態間變異、長度變異、人數變異，生成不同參數下的PPG訊號，並且比較以下三種模型的優劣，三種模型分別是通用模型與嵌入個人化特徵模型以及預訓練編碼器個人化特徵模型，除了比較個體內預測精準與否以外，也加入泛化能力上的測試。模擬結果顯示，不論是在何種參數設置底下，嵌入個人化特徵模型表現最為優異，預訓練編碼器個人化特徵模型稍稍落後於前者，代表從資料中學習並嵌入個人化特徵，有助於模型學習。同時除了在模擬資料上實驗，也將模型放置在真實的資料上，測試泛化能力，研究結果也表明，嵌入個人化特徵模型優異於通用模型。同時為驗證其在真實資料上是否也表現一致，本研究亦將上述的通用模型以及嵌入個人化特徵模型用於真實的資料，也驗證在真實資料上也保持相同的結果。

In recent years, research on stress detection has attracted increasing attention. However, most existing studies primarily focus on capturing individual differences and therefore struggle to generalize to unseen subjects, resulting in limited cross-subject generalization capability. Consequently, how to effectively generalize to new individuals has become a critical research challenge. In this study, we systematically simulate photoplethysmography (PPG) data, controlling between-individual variability, within-state variability, between-state variability, signal length variability, and subject population size. Under different parameter configurations, synthetic PPG signals are generated to evaluate and compare the performance of three modeling approaches: a general model, a model with embedded personalized features, and a personalized-feature model based on a pretrained encoder. In addition to evaluating prediction accuracy at the individual level, we further assess the generalization capability of these models. The simulation results demonstrate that, across all parameter settings, the model with embedded personalized features consistently achieves the best performance, followed by the pretrained encoder–based personalized feature model. These findings suggest that learning and embedding personalized characteristics directly from the data substantially facilitates model learning. Beyond experiments on simulated data, the proposed models are also evaluated on real-world datasets to examine their generalization performance. The results consistently show that the model with embedded personalized features outperforms the general model. Furthermore, to verify the robustness of the findings, both the general model and the embedded personalized feature model are applied to real PPG data, confirming that the observed performance trends remain consistent in real-world scenarios.