How does TrustmeWatcher combine workplace micro-sensing with explainable well-being feedback?

The application joins sensing, self-report and feedback in one workflow

Workplace sensing studies combine long-running behaviour traces with self-reports, but the tools that collect those data often sit apart from the interface that returns results. TrustmeWatcher is the application built for the TRUST-ME project to connect this work. It reuses ActivityWatch's OS-level watchers for computer-activity collection and adds its own application layer, so the person who produces the data can review the recorded activity, control ambient sensing, and receive model estimates and explanations in the same application.

The paper's scope is deliberately narrow: only ActivityWatch data serves as model input. A camera, an eye tracker, audio and recorded highlights are used to provide study context, and because the camera and eye tracker capture sensitive contextual data, participants get a visible and reversible way to pause or resume sensing. Placing that control beside returned feedback makes consent an ongoing choice rather than an installation-time setting.

Check-ins and video questionnaires stay separate labels

The study collects two forms of self-report. Participants complete periodic nine-question check-ins on a StreamDeck during work, and they complete video questionnaires after reviewing recorded highlights. The first is an immediate response close to the experience; the second is a later response grounded in a concrete episode from the day.

TrustmeWatcher preserves the response source and timestamp when it aligns each self-report with the preceding activity context. That matters because the two channels are analysed together without being treated as interchangeable labels. The nine Likert-scale items cover mood, emotion intensity, tiredness, work enthusiasm, work immersion, physical comfort, social experience, work stress and productivity, and responses from both sources were combined for model training.

Six state scores and well-being are built from questionnaire items

The prediction pipeline uses seven regressors to estimate six state scores and overall well-being. Targets are calculated from questionnaire responses: mood valence is mood plus three, arousal is emotion intensity, restfulness is six minus tiredness, stress management is six minus work stress, productivity is the productivity item, engagement averages work enthusiasm and work immersion, and overall well-being averages valence, engagement and stress management. Tiredness and stress are reversed so higher scores always indicate a better state. Physical comfort and social experience are collected but are not prediction targets.

Cronbach's alpha was used to assess which state scores were consistent enough to combine for overall well-being. On that basis valence, engagement and stress management were included, while fatigue and arousal were excluded. Each target uses the same 25-feature vector, summarizing the preceding 60 minutes together with 24-hour and seven-day context, and describing learned activity-state shares, application-category ratios, focus fragmentation, keyboard and pointer activity, and changes from recent history. Features are standardized separately for each participant using statistics fitted on training data only; a participant not present during training uses global training-set statistics. For each target, the regressor with the lowest validation MSE is selected.

Activity features add information beyond each person's typical response

Questionnaire days were ordered chronologically within each participant and divided approximately 80/10/10 into training, validation and test sets, with one complete day excluded at each boundary to reduce temporal overlap. Instances were not shuffled, to avoid data leakage. Models were selected using validation MSE and evaluated once on the held-out test set.

The comparison that matters is against the personal-mean baseline, which is calculated from each participant's own training and validation responses and therefore tests whether activity features add information beyond that person's typical answer. On the 0–6 response scale, the personal-mean baseline reaches 1.0 validation MSE and 0.85 test MSE, while the model reaches 0.85 validation MSE and 0.78 test MSE. The global-mean baseline, calculated separately for each target from training and validation responses, reaches 1.4 and 1.3, and the random baseline, drawing scores uniformly from the 0–6 scale, reaches 5.9 and 5.8. The margin over the personal mean is the smallest of the three comparisons, so the gain is real but modest.

Predictions become bands and counterfactuals, not raw numbers

TrustmeWatcher presents the six state scores as semantic cards. Scores below 2 map to the low band, scores from 2 to below 5 to moderate, and scores from 5 to 6 to high. Labels follow each construct, so a high stress-management score appears as low stress and a high restfulness score as low fatigue. Selecting a card opens the counterfactual view.

In that view the user chooses a target score on the shared 0–6 scale. Starting from the current 60-minute computer-activity allocation, the search swaps time between personal distraction, media, communication, development, writing and research in five-minute units: a swap removes five minutes from one category and adds those five minutes to another. The system evaluates every valid allocation and returns the one whose model estimate is closest to the selected target. This zero-sum search adopts the framing of counterfactual explanation used in prior work, and it describes alternative allocations of recorded time rather than a demonstrated way to change how someone feels.

The field record covers 17 knowledge workers over nine months

The TRUST-ME data-collection study followed knowledge workers in their everyday work setting rather than in a single laboratory session. The current archive contains pseudonymized records from 17 participants and spans July 2025 to March 2026. The amount of recorded self-report data varies with each participant's work schedule: individual records cover between 10 and 28 distinct workdays, with a median of 18.

All 17 participants contributed computer-activity context and check-ins; nine also contributed video questionnaire responses. The aligned records feed an external model-development pipeline, and the deployed model later uses recent and longer-range activity features to produce daily estimates on the user's device. The trained model runs locally, and the dashboard presents its predictions in semantic bands.

Explanation comprehension and trust calibration remain untested

The reported evaluation covers prediction performance and data-collection coverage. Explanation comprehension, trust calibration and interpretation of model scenarios require analysis of repeated participant interactions, so the paper does not yet show whether the counterfactual view helps users understand or act on a prediction.

The input data also bound what the estimates can mean. Computer activity omits meetings and off-screen work, and application categories cannot establish intent, so estimates reflect recorded context rather than complete accounts of work or well-being. The paper's own framing is that automatically measured and self-reported productivity can diverge, so neither should stand in for the other.

The second phase turns the model into a participant-facing study

The TRUST-ME study has two phases. The first collected computer activity and repeated self-reports from 17 participants for model development and evaluation. The second, currently being prepared, will use TrustmeWatcher as its participant-facing application.

That shift is the paper's real priority. The first phase produced a model whose test MSE of 0.78 beats a personal-mean baseline of 0.85 on the same 0–6 scale, but the second phase is where the application's other components are meant to be exercised: activity review from hours to years, self-report collection, privacy control, local prediction, and a counterfactual interface for examining explanation use across repeated encounters. For readers weighing this work, the decisive question is not whether the model edges out a personal mean, but whether repeated exposure to bands, counterfactuals and pause controls changes how participants interpret predictions about their own working day.

Privacy control is built into the feedback loop

Privacy Control lets users pause or resume the camera and eye tracker used by the study, and the application shows whether that sensing is active. Pausing turns the camera and eye tracker off. The paper argues such control is particularly important in workplace settings, where power asymmetries can make consent less freely given, and that it may help reduce concerns about surveillance.

This is a design rationale rather than a measured outcome in the reported evaluation. It nonetheless shapes how the rest of the system should be read: the same interface that returns well-being estimates also carries the commands that stop ambient sensing, so participants can keep reviewing computer activity and receiving daily feedback while the camera and eye tracker are paused.

Sources used in this answer