Source-linked AI summary
Automatic Stress Detection in Working Environments from Smartphones' Accelerometer Data: A First Step
Enrique Garcia-Ceja, Venet Osmani, Oscar Mayora
TL;DR
Occupational stress is difficult to measure objectively and has health and organizational consequences. This study used smartphone accelerometer features and statistical models in unconstrained workplaces, achieving 71% accuracy with user-specific models and 60% with similar-users models.
Problem
Occupational stress affects workforce health and organizations, while subjective self-reporting makes psychological-state measurement difficult.
Method
The study extracted time- and frequency-domain features from a single smartphone accelerometer and classified self-reported stress levels using statistical models.
Results
71% overall accuracy was achieved for user-specific models and 60% for similar-users models.
Takeaways & Limitations
Single-accelerometer smartphone data showed potential for stress detection in unconstrained working environments, including similar-users modeling when individual labelled data were scarce.
Abstract
from arXiv · showhide
Increase in workload across many organisations and consequent increase in occupational stress is negatively affecting the health of the workforce. Measuring stress and other human psychological dynamics is difficult due to subjective nature of self- reporting and variability between and within individuals. With the advent of smartphones it is now possible to monitor diverse aspects of human behaviour, including objectively measured behaviour related to psychological state and consequently stress. We have used data from the smartphone's built-in accelerometer to detect behaviour that correlates with subjects stress levels. Accelerometer sensor was chosen because it raises fewer privacy concerns (in comparison to location, video or audio recording, for example) and because its low power consumption makes it suitable to be embedded in smaller wearable devices, such as fitness trackers. 30 subjects from two different organizations were provided with smartphones. The study lasted for 8 weeks and was conducted in real working environments, with no constraints whatsoever placed upon smartphone usage. The subjects reported their perceived stress levels three times during their working hours. Using combination of statistical models to classify self reported stress levels, we achieved a maximum overall accuracy of 71% for user-specific models and an accuracy of 60% for the use of similar-users models, relying solely on data from a single accelerometer.
I. INTRODUCTION
The paper addresses occupational stress by detecting self-reported stress levels from smartphone accelerometer behavior in real workplaces. It reports 71% accuracy for user-specific models and 60% for similar-users models using a single accelerometer.
- Occupational stress is associated with health problems and reduced organizational performance, creating a need for continuous workplace monitoring.Reported consequences include cardiovascular, musculoskeletal, immunological, and mental-health problems, as well as lower productivity and increased absenteeism.
- Smartphones can monitor behavior with less discomfort and observer effect because they are familiar, widely adopted devices.The study focuses on smartphone sensing in real working environments without restricting phone use.
- 30 subjects from two organizations reported perceived stress three times during working hours during the study.
- 71% overall accuracy was achieved with user-specific models, compared with 60% using similar-users models.Both approaches classified self-reported stress levels using statistical models and a single triaxial accelerometer.
II. RELATED WORK
Prior automatic stress-detection studies used both controlled laboratory and unconstrained real-life settings. This work emphasizes unconstrained, out-of-lab detection with unknown stressors and only a smartphone accelerometer.
- Previous studies can be grouped into in-lab experiments with controlled protocols and unconstrained experiments during subjects’ daily routines.
- This work collected data outside the laboratory in an unconstrained environment with unknown stressors.
- The study explored detecting acute stress levels using only a smartphone accelerometer.The sensor was selected as non-visual and non-auditory, reducing privacy concerns and interference with daily routines.
- Similar-users models were developed when individual models were infeasible because of scarce labelled data.
III. DATA COLLECTION
Behavioral and subjective stress data were collected from Samsung Galaxy SIII Mini smartphones with participant consent. Stress was assessed repeatedly using a burnout-derived questionnaire.
- Behavioral data were collected using built-in sensors of Samsung Galaxy SIII Mini smartphones and stored by a study application.Although several contextual data types were recorded, this work analyzed only the specified sensor data.
- Participants completed a questionnaire derived from the clinically validated Oldenburg Burnout Inventory.
- Perceived stress was reported three times during working days on a 5-point scale.Assessments occurred in the morning, afternoon, and at the end of the workday.
A. Participants
The study analyzed 30 healthy participants over eight weeks in ordinary work settings, excluding weekends. Data completeness varied because some surveys were skipped and users differed in collection days.
- 30 healthy subjects provided sensor and self-reported stress data for eight weeks, excluding weekends.
- Participants used the phones from morning through the end of the workday without restrictions on usage.
- The average number of collection days per user was 29±6 because of user compliance issues.Some surveys during the day were occasionally skipped.
- All participants consented to data recording and were told that their data were anonymous and for research purposes only.
IV. PRE-PROCESSING
The study transforms raw accelerometer signals into 34 time- and frequency-domain features, then summarizes them around survey responses and groups stress into three levels.
- Feature extraction: 34 features were extracted from raw accelerometer data across time and frequency domains.Features include axis statistics, magnitude measures, signal energy, entropy, Fourier components, and spectral peaks.
- Feature extraction: Features were extracted using non-overlapping windows of 128 samples, corresponding to 25.6 seconds.
- Stress labels: Stress responses were grouped into low stress for scores 1–2, medium stress for score 3, and high stress for scores 4–5.
- Survey alignment: For the second and third surveys, features from the preceding 2 hours were summarized using their mean, maximum, and minimum, producing 102 final features.
- Survey alignment: Table II reports the total number of observations associated with each stress level for the second and third surveys.
V. EXPLORATORY DATA ANALYSIS
Exploratory analyses examine weekday stress patterns and whether extracted features distinguish stress levels. Entropy appears promising, especially for separating low from high stress, while high-versus-medium separation is weaker.
- Weekday patterns: Tuesday had the highest reported stress level, while stress decreased toward its minimum on Friday.
- Weekday patterns: Differences in average stress across weekdays were not significant, as indicated by overlapping standard error bars and an analysis of variance test.
- Feature distributions: Entropy median differences were clearly visible between high and low stress and between medium and low stress.
- Feature distributions: Entropy showed a smaller but still noticeable median difference between high and medium stress.
- Statistical comparisons: Most features differed significantly across stress-level pairs for most users, except Peak Magnitude.The analysis used a Mann Whitney U test with α = 0.01 and Bonferroni correction.
VI. STATISTICAL MODELS
The study uses Naive Bayes and Decision Trees to classify stress levels, with forward feature selection used to identify feature combinations that improve accuracy.
- Classification models: Naive Bayes and Decision Trees were selected as the two classification models for predicting stress levels.
- Feature selection: Forward Feature Selection adds predictors sequentially and retains the variable that most improves accuracy at each step.
A. Model Schemes
The paper compares user-specific, general, and similar-users schemes for stress prediction, using behavioral similarity to select training users. Similar users are identified by clustering vectors that encode pairwise stress-level feature differences.
- User-specific models train and evaluate separately for each user, whereas general models use data from other users under leave-one-person-out evaluation.
- Similar-users models train a test user’s predictor using observations from a subset of users with similar behavioral patterns.
- Behavior vectors encode, for each feature, median differences between every pair of stress levels: low-medium, low-high, and medium-high.
- K-means clusters other users’ behavior vectors, then selects the cluster whose centroid is closest to the test user’s vector.
- The test user’s similarity vector is computed from a random observation subset that is discarded during evaluation to avoid over-fitting.
- The silhouette index selects the number of clusters; an example has index 0.32 for k = 2, compared with 0.2 for k = 3.
B. Ordinal Classification
The paper treats stress level as an ordered response and adapts classification and evaluation to account for the severity of prediction errors. It also examines clustering quality when using three stress classes.
- The silhouette index for k = 3 is 0.2, while the plotted lines represent behavior vectors grouped into clusters.
- Stress classes are ordinal because low < medium < high, so prediction errors can differ in severity.
- The ordinal approach converts a k-class problem into k-1 binary problems, estimates each class probability, and predicts the class with maximum probability.
- The approach was applied with Naive Bayes to incorporate ordering information into stress classification.
- Ordinal evaluation includes MSE, MAE, linear correlation, and ACCn, which allow error severity, tolerance, association, or near-correct predictions to be assessed.
VII. EXPERIMENTS AND RESULTS
The experiments compare user-specific, general, and similar-users models across classifiers and evaluation measures. User-specific models perform best overall, while similar-users models improve on general models but remain a less personalized alternative.
- The experiments evaluate user-specific, general, and similar-users model schemes using cross-validation tailored to each scheme.User-specific models use 5-fold cross-validation; general and similar-users models use leave-one-person-out cross-validation.
- The evaluation reports Pearson correlation, Spearman correlation, and Accuracy within 1 (ACC1).
- All non-random classifiers had similar overall performance for user-specific models, while the Random classifier achieved ACC1 = 0.81 because ACC1 is optimistic.
- General models performed much lower than user-specific models, with non-random correlations stronger than Random but still weak.
- Similar-users models increased accuracy over general models by 8% for Naive Bayes and 5% for the Decision Tree.
- The comparison includes general models, similar-users models, and user-specific models.
VIII. CONCLUSIONS
This work takes a first step toward detecting stress in unconstrained working environments using smartphone accelerometer data and statistical models. User-specific models performed best, while similar-users models offered an intermediate approach when labelled data were scarce.
- Accelerometer data were converted into time-domain and frequency-domain features for statistical classification of self-reported stress levels.The study relied on data from a single accelerometer sensor.
- The data were collected in an unconstrained environment with unknown stressors, supporting evaluation in real working conditions.
- 71% overall accuracy was achieved with user-specific models, which performed best but require more labelled data.These models target each specific user.
- 60% overall accuracy was achieved with similar-users models, positioned between general and user-specific models.They use a small amount of labelled data to identify similar users.
- The follow-up work will extend data collection to several months, include more users, and analyze phone-handling situations.These extensions aim to provide finer-grained insight into user behaviour.