Source-linked AI summary
Daily Stress Recognition from Mobile Phone Data, Weather Conditions and Individual Traits
Andrey Bogomolov, Bruno Lepri, Michela Ferron, Fabio Pianesi, Alex, Pentland
TL;DR
Stress detection traditionally relies on continuously worn physiological sensors, motivating less obtrusive alternatives. This paper combines smartphone activity, weather conditions, and personality traits in a person-independent multifactorial model for binary daily stress recognition. The model achieves 72.28% accuracy using a 32-dimensional feature space, while the authors identify limitations involving sample setting and unavailable proximity data.
Problem
Stress affects well-being and health, while physiological detection systems require obtrusive sensors to be continuously carried.
Method
The paper uses a person-independent multifactorial model combining smartphone activity, weather conditions, and personality traits for 2-class daily stress recognition.
Results
72.28% accuracy was obtained with a 32-dimensional Random Forest feature representation for the 2-class classification problem.
Takeaways & Limitations
The results indicate that reasonable predictive power requires simultaneous use of smartphone activity, weather, and personality information.
Takeaways & Limitations
The sample came from married graduate students living in one university environment, and proximity data excluded interactions with nonparticipants.
Abstract
from arXiv · showhide
Research has proven that stress reduces quality of life and causes many diseases. For this reason, several researchers devised stress detection systems based on physiological parameters. However, these systems require that obtrusive sensors are continuously carried by the user. In our paper, we propose an alternative approach providing evidence that daily stress can be reliably recognized based on behavioral metrics, derived from the user's mobile phone activity and from additional indicators, such as the weather conditions (data pertaining to transitory properties of the environment) and the personality traits (data concerning permanent dispositions of individuals). Our multifactorial statistical model, which is person-independent, obtains the accuracy score of 72.28% for a 2-class daily stress recognition problem. The model is efficient to implement for most of multimedia applications due to highly reduced low-dimensional feature space (32d). Moreover, we identify and discuss the indicators which have strong predictive power.
1. INTRODUCTION
The paper proposes person-independent daily stress recognition from smartphone activity, weather conditions, and personality traits, motivated by the intrusiveness of physiological sensing. Its multifactorial model uses these complementary information sources and achieves 72.28% accuracy in binary stress classification.
- Physiological stress sensors are intrusive, can be confounded by daily activities, and exhibit important between-person differences.
- The approach combines smartphone activity, weather conditions, and personality traits to recognize daily stress as a 2-class classification problem.
- 72.28% accuracy was obtained for 2-class daily stress recognition using the combined information sources.
- The dataset used to validate the approach was collected over seven months from 111 subjects.
- The study compares the multifactorial model with single-family and pairwise combinations of personality, weather, and smartphone features.
2. RELATED WORK
Prior stress-detection research used physiological, audio, video, and smartphone data, but each alternative has practical, environmental, privacy, or sample-size limitations. The paper positions multifactorial smartphone-based recognition as a distinct approach to daily stress classification.
- Physiological sensing supports stress detection but requires continuously worn sensors for monitoring.
- Audio-based methods depend on sound quality in natural settings and face large individual differences in speech–emotion correlations.
- Video stress-monitoring systems are difficult to employ across varied real-world mobile environments and raise privacy concerns.
- MoodScope recognized daily-average mood with 66% accuracy, but it did not recognize stress states.
- Bauer and Lukowicz detected average behavior modification of 53% in seven students during an exam session, limiting the study's subject coverage.
- Sano and Picard reported 75% stress-recognition accuracy from mobile-phone and wearable features, but their results were preliminary because they used 18 subjects across five days.
3. DATA COLLECTION
The study collected longitudinal smartphone, proximity, survey, personality, and weather data from a diverse graduate-student residence sample. Daily stress labels came from self-reports, and the observed stress variability was greater within individuals than between individuals.
- Data were collected from 117 subjects living in a married graduate-student residency over seven months, including participants from 16 countries.
- Bluetooth scans ran every five minutes to identify proximity with other participating phones.
- The dataset contained 33,497 phone calls, 22,587 SMS messages, and 1,460,939 Bluetooth hits.
- Participants completed evening surveys reporting their daily self-perceived stress level.
- Stress was measured on a seven-item scale from 1, not stressed, to 7, extremely stressed; analyses retained 111 subjects with at least two consecutive weeks of data.
- Within-person daily stress variance was more spread than between-person variance, although between-person variance had higher density.
- Personality was measured with the 44-question Big Five questionnaire using 5-point Likert scales.
- Daily weather features included temperature, pressure, precipitation, humidity, visibility, and wind speed from Boston-area stations.
4. FEATURE EXTRACTION
The paper derives behavioral features from call, SMS, and Bluetooth proximity data, emphasizing phone usage, active behavior, diversity, and regularity. It also applies summary statistics and moving windows to capture temporal variation and possible effects of past events.
- Call and SMS features: 25 call and SMS basic features cover general phone usage, diversity, active behaviors, and regularity.The features include counts, contact diversity, nighttime activity, response behavior, and inter-event timing.
- Call and SMS features: Active-behavior features include nighttime calling, call initiation, SMS response rate and latency, and initiated SMS percentage.A response is counted when a text is sent within one hour after receiving the last text from the contact.
- Regularity: Regularity features measure average and variance of inter-event times for calls, SMS, and combined call-plus-SMS activity.Second-order statistics include mean, median, extrema, quantiles, variance, and standard deviation, with two- and three-day backward windows.
- Diversity: Diversity features quantify contact distribution using entropy, unique-contact counts, and contact-to-interaction ratios for calls and SMS.Higher entropy corresponds to interacting more evenly and frequently with many contacts; Miller-Madow correction is applied to entropy estimates.
- Proximity features: Bluetooth proximity data are filtered using RSSI >= 0 and organized into general proximity, diversity, and regularity features.The proximity feature set also uses Miller-Madow correction for entropy calculation.
5. METHODOLOGY
Daily stress recognition is formulated as person-independent binary classification using smartphone-derived behavioral data, weather, and personality traits. The methodology reduces the feature space, compares several classifiers, and selects a Random Forest model using subject-separated validation and model-selection metrics.
- Problem formulation: The task labels stress scores <= 4 as not stressed and scores > 4 as stressed.The resulting classes contain 63.84% not-stressed and 36.16% stressed cases.
- Validation: Training and testing data are split 80%/20% while preventing the same subjects from appearing in both sets.Leave-one-subject-out validation trains 111 models on 110-subject subsets and averages performance on the held-out subjects.
- Feature selection: Feature ranking and subset selection reduce approximately 500 initial features to 32 features.The reduced space is intended to make the approach efficient for mobile and multimedia applications.
- Classifier selection: The study evaluates Random Forest, GBM, linear and Gaussian-kernel SVMs, and Neural Networks, selecting Random Forest as the best solution.Only Random Forest performance results are reported thereafter.
- Random Forest model: Random Forest models combine decision-tree predictors whose decision boundary is characterized by a margin function.The margin compares average votes for the correct class with the highest average vote for another class, while generalization error is defined from negative margins.
- Model selection: The model uses 112 trees as the estimated optimum under 10-fold cross-validation and selects the final model using Cohen’s κ.Cohen’s κ corrects classification agreement for agreement expected by chance.
- Robustness and assumptions: Resampling with replacement and cross-validation are used to limit overfitting and address possible missing logged calls, SMS, and Bluetooth proximities.Confounding variables are identified but retained during training and testing.
6. EXPERIMENTAL RESULTS
The recognition model was evaluated with accuracy, κ, sensitivity, and specificity, including held-out testing and 10-fold cross-validation. Its test accuracy was 72.28%, and metric distributions were stable across folds.
- 72.28% test accuracy was achieved by the random forest recognition model, compared with 90.68% accuracy on the training set.
- Accuracy, κ, sensitivity, and specificity were used to evaluate the approach.
- Performance-metric distributions did not vary substantially across 10 cross-validation folds, indicating good generalization despite heterogeneous data and resampling noise.
- The experiments compared the multifactorial approach with models using single feature families and pairwise feature combinations.
7. DISCUSSION
The discussion finds that combining personality, weather, and smartphone-activity features is central to performance, while examining which individual predictors contribute most to daily stress recognition.
- 72.28% accuracy was obtained by the Random Forest model using all three feature sets and 32-dimensional feature vectors.
- Single feature sets and pairwise combinations did not outperform the majority classifier, whose accuracy was 63.84%.
- The reported 72.28% accuracy exceeded a prior 53% result, although direct comparison was considered potentially inadequate because the tasks differed.
- All personality traits contributed significantly, including Agreeableness and Openness to Experience alongside traits emphasized in earlier social-psychology studies.
- Temperature was associated with stress, while humidity, visibility, and wind speed also had significant predictive effects.
- Of the 32 selected features, 11 were proximity features, while 6 came from call data and 6 from SMS data.
- SMS predictors included reply latency, defined as median response time, and the amount of incoming and outgoing SMS communication.
8. IMPLICATIONS AND LIMITATIONS
The paper presents mobile-phone-based stress recognition as an unobtrusive, widely available tool with potential applications in clinical, workplace, and self-monitoring contexts. Its limitations include a homogeneous living environment and missing proximity data for nonparticipants.
- Implications: 72.28% accuracy was achieved by combining personality traits, social relationships, and weather data for daily stress recognition.The system is described as cost-effective, unobtrusive, widely available, and reliable.
- Applications: The approach could support therapists by recording longitudinal daily stress, identifying recurrent stressors, and informing treatment adjustments.
- Applications: Mobile applications could increase stress awareness, issue warning-threshold alerts, and suggest stress-management or relaxation techniques.
- Applications: The system could be extended to workplaces for early detection of stress-related conflicts and contagion and for supporting balanced workloads.
- Limitations: The sample consisted of married graduate students living in one campus environment, limiting environmental representativeness despite participants spanning 16 countries and all continents.The study also lacked proximity data for interactions with people outside the participant group.
9. CONCLUSION
The conclusion evaluates daily stress recognition using smartphone activity, weather conditions, and personality traits. Results indicate that combining all three sources supports predictive power and generalization, while the approach remains subject to the study’s stated limitations.
- Conclusion: The study models daily stress recognition from smartphone activity, weather conditions, and personality traits as a 2-way classification problem.
- Conclusion: All three data types are necessary for reasonable predictive power because removing any one reduces performance below the baselines.
- Conclusion: Accuracy and κ distributions indicate robustness and generalization power for the multifactorial approach.
- Conclusion: The results provide evidence that individual daily stress can be reliably predicted while requiring simultaneous consideration of personal, environmental, and stable individual characteristics.