Student Stress Factors Dataset Overview

• Student Stress Factors Dataset is a survey-based resource capturing stress via psychometric scores like the PSS and additional academic, environmental, social, and behavioral variables.
• It comprises data from distinct studies, including a Bangladesh survey with three stress classes and an AIIMS-guided questionnaire with binary stress classification.
• The dataset’s diverse preprocessing pipelines, feature engineering, and modeling approaches highlight its potential for robust, context-aware stress prediction.

Student Stress Factors Dataset denotes a survey-based data resource for modeling student stress from structured psychological, academic, environmental, social, behavioral, physiological, and demographic variables. In its most fully specified recent form, it consists of responses from 1,100 undergraduate and graduate students across three universities in Bangladesh, collected in two waves during April 2023 and November 2023, with stress levels derived from the 10-item Perceived Stress Scale (PSS) and contextual covariates distributed through Kaggle (Ovi et al., 1 Aug 2025). A plausible implication is that the term is not yet a uniquely standardized identifier in the literature, because a separate machine-learning study also used the same label for a different questionnaire dataset built from 28 stress-and-well-being items administered to 843 college students, with 825 records retained after cleaning (Singh et al., 2024).

1. Dataset identity and scope

Recent work uses the name “Student Stress Factors Dataset” for closely related, but non-identical, survey collections. One version is the dataset referred to as Dataset 1 in a context-aware machine-learning framework; it is based on the PSS together with 11 additional items targeting academic, environmental, social, and lifestyle factors, and it was collected from college students by convenience sampling through university mailing lists and classroom announcements (Ovi et al., 1 Aug 2025). Another version appears in a workshop-based study of stress and non-stress classification; it contains 28 questionnaire items validated under the guidance of experts from the All India Institute of Medical Sciences (AIIMS) Raipur and was administered to students aged 18–21 years old (Singh et al., 2024).

Dataset usage	Reported composition	Access
Survey dataset in a context-aware framework (Ovi et al., 1 Aug 2025)	1,100 students; PSS + 11 additional items; two 2023 waves	Kaggle
AIIMS-guided questionnaire dataset (Singh et al., 2024)	843 students; 28 questions; post-cleaning $n=825$	Kaggle and GitHub

This naming overlap matters methodologically. The Bangladesh dataset is organized around stress-level stratification from a standardized PSS score, whereas the AIIMS-guided dataset frames the problem as binary stress versus non-stress classification from a broader stress-and-well-being questionnaire. The phrase “student stress factors” therefore refers both to a specific public dataset and to a broader survey design tradition in which stress is modeled through covariates that extend beyond symptoms alone.

2. Feature space and measurement design

The Bangladesh dataset organizes its variables into psychological, academic, physiological, environmental, behavioral, social, and demographic categories (Ovi et al., 1 Aug 2025). The reported schema includes PSS_Score, Anxiety_Level, Depression_Score, Study_Hours, Assignment_Load, Exam_Frequency, GPA, Sleep_Quality, Physical_Activity, Nutrition_Index, Chronic_Illness, Noise_Level, Campus_Facilities, Time_Management, Procrastination_Score, Family_Support, Peer_Support, Social_Engagement, Screen_Time, Relaxation_Time, Age, and Gender. The measurement scales are heterogeneous by design: PSS_Score ranges from 0–40, multiple self-report variables use 1–5 Likert responses, and several contextual variables are direct numeric counts or durations such as hours studied per day, pending assignments, and social outings per week.

In the AIIMS-guided questionnaire dataset, the feature space is broader in item count but narrower in explicit covariate semantics (Singh et al., 2024). It contains 31 columns: Age, Gender, Q1 through Q28, and Stress_Label. The 28 questionnaire items cover seven domains: Emotional Well-being, Physical Well-being, Academic Performance, Relationships & Social Environment, Leisure & Relaxation, Stress Symptoms, and Coping & Non-Stress Indicators. All responses are mapped to a 5-point Likert scale from 1 = Not at all to 5 = Extremely.

These schemas show two different design philosophies. The Bangladesh dataset makes the factor structure explicit by naming each feature as a stress-related construct or contextual variable. The AIIMS-guided dataset instead foregrounds questionnaire coverage and downstream classification. In both cases, stress is not treated as an isolated symptom; it is embedded in a multivariate description of student life that includes workload, sleep, social support, health, and self-regulation.

3. Stress labels and reported distributions

In the Bangladesh dataset, stress-level classes are derived directly from PSS_Score using standard cutoffs (Ovi et al., 1 Aug 2025). Low stress is defined as $\text{PSS\_Score} \leq 13$, moderate stress as $14 \leq \text{PSS\_Score} \leq 26$, and high stress as $\text{PSS\_Score} \geq 27$. The reported class counts are 342 low-stress samples, 486 moderate-stress samples, and 272 high-stress samples. Select descriptive statistics include $\mu = 21.5$, $\sigma = 7.8$ for PSS_Score, $\mu = 4.2$ hours and $\sigma = 1.6$ for Study_Hours, and $\mu = 3.1$, $\sigma = 1.1$ for Sleep_Quality.

In the AIIMS-guided dataset, the label is binary rather than ordinal (Singh et al., 2024). A composite Stress score is computed by summing responses across the 28 items, and participants whose total score exceeded the clinical cutoff were assigned Stress, with the remainder assigned Non-stress. After cleaning, the dataset contains 825 records, with Stress_Label distributed as 42% non-stress and 58% stress. Reliability analysis reported Cronbach’s $\text{PSS\_Score} \leq 13$0, and the top five predictors by Pearson’s $\text{PSS\_Score} \leq 13$1 versus Stress_Label were Q5 (Feeling overwhelmed), Q12 (Trouble sleeping), Q19 (Difficulty concentrating), Q3 (Irritability), and Q22 (Mood swings).

Dataset variant	Label construction	Reported distribution
Bangladesh survey dataset (Ovi et al., 1 Aug 2025)	Low, moderate, high stress from PSS_Score cutoffs	342 / 486 / 272
AIIMS-guided questionnaire dataset (Singh et al., 2024)	Binary Stress_Label from 28-item summed score and expert cutoff	42% non-stress, 58% stress

A common source of confusion is to treat these labels as interchangeable. They are not. One dataset models ordinal stress severity from a named psychometric instrument, while the other constructs a binary endpoint from a broader stress-and-well-being questionnaire. Reported classification performance must therefore be interpreted in light of the label ontology rather than as a direct ranking of algorithms across equivalent tasks.

4. Preprocessing, validation, and modeling

The Bangladesh dataset is accompanied by an explicitly staged preprocessing and modeling pipeline (Ovi et al., 1 Aug 2025). Missing numeric values are handled by mean imputation, categorical values by mode imputation, Gender and Chronic_Illness are one-hot encoded, and Likert items are treated as numeric. The paper reports both Min–Max scaling to $\text{PSS\_Score} \leq 13$2 and Z-score standardization where used. Feature selection is performed with SelectKBest using an F-test, with the optimal $\text{PSS\_Score} \leq 13$3 chosen via 5-fold cross-validation on SVM, and with RFECV using a linear SVM and 5-fold stratified cross-validation. Dimensionality reduction is implemented with PCA preserving 90%, 95%, or 99% of total variance. The data are split with a stratified 75% train / 25% test partition using random_state=42, and hyperparameter tuning uses 10-fold stratified cross-validation. Base classifiers are SVM, Random Forest, Gradient Boosting, XGBoost, AdaBoost, and Bagging; ensemble strategies include hard voting, soft voting, weighted hard voting, weighted soft voting, and stacking. The best reported result on this dataset is 93.09% accuracy with weighted hard voting (Ovi et al., 1 Aug 2025).

The AIIMS-guided dataset follows a more compact preprocessing pipeline (Singh et al., 2024). Duplicate records are removed, Gender is label encoded to $\text{PSS\_Score} \leq 13$4, Likert responses are mapped to integers 1–5, and all input features are normalized with min–max scaling to $\text{PSS\_Score} \leq 13$5. Any record with one or more missing responses is dropped, which removes approximately 2% of cases, and extreme outliers outside the allowed response or age ranges are automatically filtered. The paper evaluates Decision Trees, Random Forest, Support Vector Machines, AdaBoost, Naive Bayes, Logistic Regression, and K-nearest Neighbors, and reports that Support Vector Machines reach 95% accuracy for Stress (Singh et al., 2024).

Taken together, these workflows show that the Student Stress Factors label is associated with two distinct methodological regimes: one emphasizes explicit feature engineering, multi-class learning, and ensemble design; the other emphasizes questionnaire screening, binary classification, and conventional supervised baselines. Both are tabular-learning settings, but they embody different assumptions about how stress should be discretized and predicted.

5. Position within the student-stress data ecosystem

The Student Stress Factors Dataset belongs to a broader ecosystem of student-stress data resources that differ in modality, temporal resolution, and ground-truth construction. The Student Mental and Environmental Health (StudentMEH) Fitbit dataset contains raw Fitbit exports, weekly segments, and labels for depression, anxiety, and stress inherited from CES-D-10, STAI, and PSS-4 surveys aligned to the nearest week; stress is defined as high perceived stress for $\text{PSS\_Score} \leq 13$6 (Lopez et al., 22 Jan 2026). In that setting, stress-related modeling depends not only on the label definition but also on sensor modality and aggregation window. Reported guidance includes calories + AdaBoost at 4 h with $\text{PSS\_Score} \leq 13$7 for stress and the warning to avoid concatenating all modalities because single-modality models often outperform multimodal ensembles in that dataset (Lopez et al., 22 Jan 2026).

Razavi et al. study a different operationalization of student stress based on smartwatch data and ecological self-reports (Razavi et al., 2023). Their dataset contains 54 college students, approximately 40 days of continuous 1 Hz heart rate and hand-acceleration data, and 60 s windows centered on self-reported “stress taps.” The reported label distribution is 3,497 stress windows versus 29,475 non-stress windows, and XGBoost achieves an AUC of 0.64 with 84.5% accuracy. The most important features are the standard deviation of hand acceleration, standard deviation of heart rate, and minimum heart rate (Razavi et al., 2023). Here, “stress factors” are not questionnaire covariates but short-horizon physiological signatures.

StudentLife provides yet another framing: 23 Dartmouth students, 1,183 labeled days, and daily EMA responses on a 1–5 perceived-stress scale collapsed into three classes, with passive sensing from activity, audio, conversation, phone state, GPS-derived aggregates, and sleep variables (Shaw et al., 2019). The CALM-Net model combines an LSTM-based auto-encoder with per-student multitask heads and reports a 45.6% improvement in F1 score over the state of the art (Shaw et al., 2019). This line of work shifts the emphasis from population-level tabular screening toward personalized temporal modeling.

Other datasets define student stress factors at the level of cohorts, events, or discourse. An Oura Ring study of 103 Japanese university students over up to 28 months identifies cyclical stress biomarkers during exams, New Year’s, and job hunting season through in-the-wild wearable monitoring and linear mixed-effects models (Neigel et al., 2024). A Reddit-based dataset of student-related depression posts isolates 19,552 COVID-period posts and identifies stress-factor categories such as Online Education Challenges, Employment Concerns / Job Loss, Family Issues, Social Isolation / Missing Friends, and Daily‐Activity Disruption (Thukral et al., 2020). A weekly panel study of first-year physics students in Germany yields 3,241 PSQ observations and 5,823 declared stressors, with exercise sheets, physics lab courses, math courses, exam preparation, and exams emerging as dominant stress sources across the semester (Lahme et al., 2024). EmpathicSchool, by contrast, captures acute laboratory stress through facial video, EDA, BVP, IBI, HR, skin temperature, and accelerometry, with 20 s sliding windows and NASA-TLX-derived interval labels (Hosseini et al., 2022).

At the largest scale, PISA 2022 has been used to model the relationship between stress-related indicators and academic performance for 613,744 students across 81 economies, with consistent negative associations between anxiety-related constructs and performance across continents (Ghazanchyan et al., 30 May 2026). Relative to such global data, the Student Stress Factors Dataset is more granular in its student-life covariates but much narrower in sampling frame. Its main value lies in combining psychometric stress scores with directly named academic, behavioral, social, and environmental variables in a form that is immediately usable for supervised learning.

6. Limitations, privacy, and interpretive issues

The survey-based Student Stress Factors datasets are comparatively easy to preprocess and share, but their inferential scope is bounded by sampling design and label construction. The Bangladesh dataset uses convenience sampling across three universities and two timepoints, and the AIIMS-guided dataset is derived from a workshop cohort of students aged 18–21 years old (Ovi et al., 1 Aug 2025, Singh et al., 2024). A plausible implication is that transportability across institutions, countries, and degree structures is dataset-specific rather than guaranteed.

A second limitation is semantic heterogeneity in what counts as a “stress factor.” In the survey datasets, stress is a function of questionnaire responses and named contextual covariates. In the Fitbit dataset, weekly Monday–Sunday segments inherit the closest survey label and stress is defined through PSS-4 high perceived stress (Lopez et al., 22 Jan 2026). In Razavi et al., stress is the moment of a self-report tap expanded to a 60 s window (Razavi et al., 2023). In the first-year physics panel, stress is perceived stress on the PSQ plus free-text stressor coding (Lahme et al., 2024). In EmpathicSchool, stress is a task- and interval-level label derived from NASA-TLX reports (Hosseini et al., 2022). In the Reddit study, stress factors emerge from LDA topics and manual interpretation of posts (Thukral et al., 2020). These are related constructs, but they are not identical targets.

Privacy and governance become more stringent as student-stress data become more multimodal and longitudinal. The StudentMEH Fitbit study anonymizes records via random IDs unlinked to personal identifiers and recommends IRB approval, informed consent, opt-in/opt-out procedures, and approved repositories for de-identified CSVs (Lopez et al., 22 Jan 2026). The Oura Ring study stores data under pseudonymized user IDs, encrypts data at rest, restricts access to institutional credentials, and aggregates analyses across more than 20 participants to prevent re-identification (Neigel et al., 2024). These practices are especially relevant when the factor space extends from questionnaire items to passive sensing, longitudinal behavioral traces, or social media.

A final misconception is that more modalities or more variables necessarily yield better models. The Fitbit study explicitly reports that concatenating all modalities should be avoided because single-modality models often outperform multimodal ensembles in that dataset (Lopez et al., 22 Jan 2026). Conversely, the survey-based Student Stress Factors variants achieve high reported accuracy with tabular inputs alone (Ovi et al., 1 Aug 2025, Singh et al., 2024). The central methodological lesson is therefore not that one modality dominates, but that stress prediction performance depends on the alignment among sampling frame, label definition, feature semantics, and temporal granularity.