Deception detection has been extensively studied using vision-based features in domains such as crime, finance, and social interaction. However, little attention has been given to how deception manifests visually in self-reported well-being-a critical area for behavioral health, where inaccurate reporting may affect treatment outcomes and the therapeutic alliance. While clinicians often rely on visual cues such as gaze, facial expressions, and body language to assess deception, these cues remain underutilized in AI-based deception detection in wellbeing scenarios. This study explores vision-based features of deception in the well-being domain and compares them with those from three other domains: biography, academics, and crime. Using mock interview data, we extract facial landmarks, body gestures, and facial action units (AUs) using four feature selection methods. We then visualize and analyze the spatial distribution of features associated with truthful and deceptive responses. Results show that well-being features are generally fewer and more localized-particularly around the nose ridge-with unique presence of eye landmarks and limited hand gestures. In contrast, biography and academics show broader facial and body engagement, while crime displays no differentiation between truth- and deception-related features and lacks emotional AU combinations. AUs associated with joy (AU 6 and AU 12) appear consistently across well-being, academics, and biography, suggesting some domain-agnostic cues. Overall, our findings indicate that most visual features relevant to deception are domain-specific. This highlights the importance of contextaware approaches in deception modeling and supports the development of more reliable, human-centered AI tools for wellbeing assessment and mental health applications.

We present a first of its kind pilot study investigating distinct features presented in lip movement captured through WiFi channel state information (CSI) as four research volunteers read several emotion-charged text samples. We pursue the tasks of emotion and identity recognition with these data via feature-level fusion, extracting features from both the time and frequency domains. While the extracted frequency-domain features, i.e., zero-crossing rate and fundamental frequency, are commonly associated with audio and speech recognition related applications, we found statistical features, such as mean, median, skew, and kurtosis, most suitable for capturing salient information in CSI data. Specifically, classifying the emotional states of speakers (i.e., anger, joy, fear, love, surprise, and sadness) and the identity of speakers themselves, we achieved 96.4% and 57.9% accuracy for the identity and emotion recognition tasks, respectively.

Abstract Indoor localization is still nowadays a challenge with room to improve. Even though there are many different approaches that have evidenced as effective, most of them require specific hardware or infrastructure deployed along the building that can be discarded in many potential scenarios. Others that do not require such on-site infrastructure, like inertial navigation-based systems, entail certain accuracy problems due to the accumulation of errors. However, this error-accumulation can be mitigated using beacons that support the recalibration of the system. The more frequently beacons are detected, the smaller will be the accumulated error. In this work, we evaluate the use of the noise signature of the rooms of a building to pinpoint the current location of a low-cost Android device. Despite this strategy is not a complete indoor localization system (two rooms could share the same signature), it allows us to generate beacons automatically. The noise recorded by the device is preprocessed performing audio filtering, audio frame segmentation, and feature extraction. We evaluated binary (determining if the ambient sound recording belonged to a specific room) and multi-class (identifying which room an ambient noise recording belonged to by comparing it amongst the remaining 18 rooms from the original 19 rooms sampled) classification methods. Our results indicate that the two Stacking techniques and K-Nearest Neighbor (KNN) machine learning classifier are the most successful methods in binary classification with an average accuracy of 99.19%, 99,08%, and 99.04%. In multi-class classification the average accuracy for KNN is 90.77%, and 90.52% and 90.15% for both Voting techniques.

One of the strongest predictors of success in post-secondary education is student engagement. Unfortunately, people with psychiatric disabilities are less engaged in their campus communities. This work-in-progress paper details the disclosure-based self-stigma reduction program, Up To Me, which is developed to increase inclusion and engagement of people with mental illness on college campuses by teaching strategies to weigh costs and benefits of disclosing one's mental illness. Further, we elaborate on the program's evaluation mechanisms, which involve both self-reported and passively recorded smartphone sensor data. The latter reflects a unique merging of behavioral and computer sciences that serves to facilitate behavioral modeling using artificial intelligence as an objective measure of Up to Me outcomes. Similar to data collection for some activity and biometric recognition applications, we employ a publicly available and free-to-use smartphone sensor reading app to correlate self-reported well-being with Up to Me participant behaviors. We anticipate that the behavioral data gathered via smartphones will substantiate self-report data on Up to Me outcomes.

This paper presents a proof-of-concept demonstrating the feasibility of global (non-person specific) mood classification using smartphone data. We employed a publicly available dataset consisting of six weeks of phone activity data. It included call, SMS, and app events along with up to five self-reported mood entries per day for 27 subjects. While existing efforts have explored person-specific and one-vs-one mood classification models, we show that a global, multiclass mood prediction model is achievable with 65% mood classification accuracy. Our global model aims to protect the privacy of smartphone users, especially since existing research employ a mobile app to track the specific daily actions of users to infer their levels of valence and arousal. Our findings show that features representative of app, call, and text messaging patterns and previous levels of valence and arousal may be most useful for mood detection. After evaluating all features using four different feature selectors, we found that the salient feature set only resulted in a 2% degradation in performance compared to the use of all features. As our results are data-dependent, our future research will involve data collection on a much larger scale to further evaluate the feasibility of privacy-preserving mood classification. Since this work is focused on smartphone devices, our work could lead to privacy-preserving objective measurement of mood for counseling services and user-friendly mHealth applications.