Learning Longitudinal Stress Dynamics from Irregular Self-Reports via Time Embeddings

Louis Simon, Mohamed Chetouani · 2025

Read Learning Longitudinal Stress Dynamics from Irregular Self-Reports via Time Embeddings by Louis Simon, Mohamed Chetouani (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-01569.txt.

Introduction

Phones and wearables can track stress over months, but people report their feelings at irregular times, leaving gaps that break standard models. Louis Simon and Mohamed Chetouani study time embeddings, numeric codes for when each report arrives, as a way to model those sequences. They introduce Ema2Vec, a time embedding built for ecological momentary assessments, brief in-the-moment questionnaires sent during daily life.

The method is tested on a new task, predicting stress over time from those irregular self-reports. It beat standard baselines that slice time into fixed daily windows, and models trained on raw sequences without time-aware codes. Simon and Chetouani conclude that how timing is represented matters as much as the reports themselves when modeling human states from messy, real-world sampling.

Catalog record

CREATORLouis Simon, Mohamed Chetouani
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE38 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-01569

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Learning Longitudinal Stress Dynamics from Irregular Self-Reports via Time Embeddings Louis Simon, Mohamed Chetouani Abstract The widespread adoption of mobile and wearable sensing technologies has enabled continuous and personalized monitoring of affect, mood disorders, and stress. When combined with ecological self-report questionnaires, these systems offer a powerful opportunity to explore longitudinal modeling of human behaviors. However, challenges arise from missing data and the irregular timing of self-reports, which make challenging the prediction of human states and behaviors. In this study, we investigate the use of time embeddings to capture time dependencies within sequences of Ecological Momentary Assessments (EMA). We introduce a novel time embedding method, Ema2Vec, designed to effectively handle irregularly spaced self-reports, and evaluate it on a new task of longitudinal stress prediction. Our method outperforms standard stress prediction baselines that rely on fixed-size daily windows, as well as models trained directly on longitudinal sequences without time-aware representations. These findings emphasize the importance of incorporating time embeddings when modeling irregularly sampled longitudinal data. Keywords Ecological Momentary Assessment, Stress, Time Embedding, Longitudinal Prediction † † footnotetext: Accepted at ACII 2025 - Canberra 1 Introduction Considering the temporal dynamics of human affective states is essential for gaining meaningful insights into mood and related emotional conditions. In the case of automated stress detection, there is a need to develop computational models that can effectively capture these evolving patterns over time, in order to build systems that are both accurate and responsive. However, ecological assessments often result in irregularly spaced self-reports, as shown in Figure 1 . While the delays between reports may carry valuable information, processing such sequences remains challenging for many machine learning models. In the context of student stress prediction, researchers have explored the use of past labels, time to the next deadline, or time to the next label as forms of temporal information. These variables have typically been treated as simple covariates or incorporated into two-stage Bayesian prediction frameworks, as in Mishra et al. 2020 . In our study, we propose to explicitly model temporal trends through the use of time embeddings. To capture delays between Ecological Momentary Assessments (EMAs) and gain insights into the dynamics of self-reports, we draw inspiration from positional encoding and time embedding techniques. Specifically, we leverage Time2Vec embeddings Kazemi et al. 2019 , which are capable of encoding both periodic and non-periodic temporal patterns. We extend this representation to explicitly model longitudinal stress dynamics from irregular self-reports. This novel time embedding, which we call Ema2Vec, is designed to handle irregular EMA intervals and capture stress-related temporal dependencies. Figure 1: Sequence of irregularly spaced self-report : To predict target EMA at time t_{0} , we leverage daily sequences \mathrm{d}_{i} aligned with past self-reports. In this paper, we make the following contributions: • Importance of longitudinal context: We demonstrate that incorporating a longer historical context significantly improves model performance. • Explicit representation of time through time embeddings: We introduce time embeddings to explicitly represent temporal information, enabling more effective and lightweight modeling of temporal dynamics. • Improvement of Ema2Vec over Time2Vec: Ema2Vec, specifically designed to capture the dynamics of irregular self-reports, outperforms Time2Vec, confirming the advantage of using a task-specific temporal representation. • Forecasting without sensing data: We show that the learned model is able to forecast future levels of self-reported stress based solely on historical self-reports, without the sensing data emitted during the day prior to the target EMA onset. This paper is organized as follows: Section 2 surveys past work on machine learning models for stress prediction and time embedding methods. Section 3 introduces the problem of longitudinal stress prediction along with the Ema2Vec time embedding module. Section 4 describes the StudentLife dataset, the training process, as well as the stress forecasting evaluation task. Section 5 presents results on both longitudinal prediction and forecasting. Finally, Section 6 concludes the paper by discussing the limitations of our approach and outlining directions for future work. 2 Related Work 2.1 Ecological Momentary Assessment The widespread use of smartphones and wearable devices has allowed research to continuously track behavioral signals in a environmental context. Researchers have extensively used smartphones and smartwatches to model stress, mood, and emotion based on phone log, position, and physiological signals. Data collection varies from few days Kang et al. 2023 to several month Ben-Zeev et al. 2017 , target different population, e.g., students Wang et al. 2014 , Kang et al. 2023 or hospital workers Mundnich et al. 2020 . These passive sensing apparatus are often paired with mobile applications specially developed for the gathering of Ecological Momentary Assessment (EMA) Shiffman et al. 2008 measures in the form of self-reported questionnaires, allowing for in-context probing of mental states. EMA surveys can be used to query users about their current affect Miller et al. 2009 , stress Wang et al. 2014 , sleeping behaviors Triantafillou et al. 2019 , or mood disorders Ebner-Priemer and Trull 2009 . EMA-based studies are particularly suited for longitudinal data as they provide repeated measures under various context which can ultimately help to identify situations inducing changes. In our study, we focus on the prediction of self-reported levels of stress. Researchers have been focusing on how different modalities, architectures, personalization techniques, and time windows can improve the prediction of self-reported stress levels using mobile sensing data. For example, Yu and Sano Yu and Sano 2023 proposed a semi-supervised model with active sampling to predict stress levels from physiological signal, leading to an improvement of the binary accuracy by 7.7% to 13.8% across three dataset. Bonafonte et al. Bonafonte et al. 2025 conducted a thorough analysis on passive sensing modalities contribution and showed the importance WiFi features, which encode mobility patterns, and phone log features in binary and 3-class stress level classification. Beyond modalities, the choice of model architecture plays a crucial role. While many contributions rely on Multi-Layer Perceptrons (MLP), Long Short-Term Memory (LSTM) networks, or traditional machine learning models, researchers have also explored alternative architectures. For example, Harit et al. Harit et al. 2024 proposed a Spatio-Temporal Graph Neural Network (STGNN) based model, demonstrating the potential of innovative architectures in this domain. Moreover, because of data heterogeneity and self-reports’ obvious subjectivity, a significant effort towards model personalization is at the core of many studies. Jaques et al. Jaques et al. 2017 developed a multitask learning MLP and a Gaussian Process regression model with domain adaptation to enable personalized stress prediction in a population of 200 students. Shaw et al. Shaw et al. 2019 proposed a Multitask AutoEncoder model named CALM-Net and achieved an gain of 39.4 % of the weighted-F1 compared to a non-personalized model. The aforementioned studies typically employ sensing sequences that capture context from the immediate past, spanning a day, or extending over longer periods of one to two weeks. Considering a longer temporal context can be beneficial, as stress induces changes at varying rates across different modalities. For instance, physiological signals may respond to stress within minutes or hours, while behavioral patterns captured through phone logs may evolve more gradually over several days. A major challenge in this context lies in distinguishing relevant from irrelevant information over extended periods of time. In our work, we address this by developing a more adaptive approach: we extract

Read the complete text