Video games are often viewed as resources for coping, emotional regulation, and resilience during adversity. However, little is known about gamers' experiences when both everyday life and gaming-related resources are disrupted. This interpretive phenomenological study explored the lived experiences of 15 Iranian gamers during the Iran–Israel missile attacks through semi-structured interviews. Data were analyzed using Diekelmann et al.‘s hermeneutic phenomenological method, with Conservation of Resources (COR) theory as an interpretive framework. Four themes emerged: collapse of the gaming lifeworld, digital solidarity and relational reconnection, reclaiming agency through gamer identity, and reconstructing everyday life and future horizons. The findings indicate that the coping value of gameplay was uneven rather than uniformly absent, while the social and symbolic dimensions of gaming continued to provide connection and support for several participants. These findings add nuance to existing research by suggesting that the significance of gaming during severe crises may extend beyond gameplay itself to the relationships, identities, and communities that form around it.
Artificial intelligence (AI) is increasingly used for psychological support, yet the psychological and relational conditions associated with this use are poorly understood. Internalized shame is a documented barrier to human-delivered mental health care, but its association with AI use, which involves no human interlocutor, has not been examined. This cross-sectional study tested whether internalized shame accounts for the association between perceived social support from family, friends, and significant others and self-reported AI use for psychological support among Korean college students. A sample of 512 students completed the Multidimensional Scale of Perceived Social Support, the Internalized Shame Scale, and a single item on the frequency of AI use for psychological counseling purposes in the past six months. A second-order structural equation model was estimated with weighted least squares with mean and variance adjustment, treating all indicators as ordinal and adjusting for gender, age, academic year, and subjective socioeconomic status. Lower perceived social support was associated with more frequent AI use (total β = −.14). This association was statistically accounted for by internalized shame (indirect β = −.11, 95% CI [−.17, −.06]), and the conditional direct association was near zero. In contrast to the deterrent association documented for human-delivered care, internalized shame was positively associated with AI use. The findings highlight internalized shame as a correlate of AI use for psychological support and a focus for research on how psychological barriers to help-seeking differ across support channels.
Problematic smartphone use (PSU) is an important behavioral health concern, but it remains unclear whether wearable-derived physiological and activity signals can discriminate psychometrically defined PSU beyond participant-specific temporal regularities. This six-month repeated-measures proof-of-concept study examined 20-day trajectories of heart rate, sleep, and physical activity in relation to Ko's Smartphone Addiction Test (KSAT)-defined PSU status.Eighty-five first-year university students were retained after quality control; 79 contributed at least one eligible 20-day sequence, yielding 1512 repeated windows. A bidirectional long short-term memory (Bi-LSTM) model was evaluated first under the original random window-level framework and then under five-fold participant-grouped evaluation with training-participant-only preprocessing. Logistic regression and random forest were evaluated on the same participant-disjoint outer folds, and an untuned unidirectional LSTM was added as an architecture-simplification sensitivity analysis. SHapley Additive exPlanations (SHAP) and Integrated Gradients (IG) were used only to characterize fitted-model attributions.The window-level Bi-LSTM showed high apparent discrimination (AUC = .973; AP = .904), but this framework allowed overlapping windows from the same participant across partitions and used pre-partition scaling. Under participant-grouped evaluation at the participant-assessment level, AUCs were .479 (95% CI .355-.610) for the Bi-LSTM, .378 (.253-.512) for logistic regression, .563 (.444-.689) for random forest, and .545 (.429-.659) for the unidirectional LSTM; all confidence intervals included .50. XAI analyses showed nonuniform temporal and directional attributions, but these patterns were model- and sample-dependent.High window-level discrimination did not translate into robust participant-disjoint generalizability, and the recurrent models did not clearly outperform non-sequential comparators. The main implication is methodological: overlapping-window evaluation can substantially overstate model performance when repeated observations from the same individuals are not separated. Larger independent cohorts, temporally matched outcome measures, and prespecified participant-level validation are required before wearable-based PSU models can support prognostic or clinical use.
