Computational Approaches to Cyber-Syndrome: A Scoping Review with Considerations for Adolescent Mental Health

1. Introduction

The digital revolution is changing human lifestyles, shifting traditional activities such as living, working, socializing, and entertaining into virtual cyberspace [1]. This excessive and problematic online interaction profoundly affects people’s thinking space, particularly their cognitive and affective states. In 2018, Ning et al. proposed the concept of Cyber-Syndrome that captures the holistic impact of cyberspace on human well-being, of which the thinking dimension particularly includes cognitive deficits (e.g., attention deficits, executive dysfunction, memory impairment) and affective disturbances (e.g., mood disorders, emotional dysregulation, and stress-related issues) [2,3,4,5,6].

Adolescents represent a uniquely vulnerable population for the Cyber-Syndrome, given their ubiquitous digital involvement and ongoing neurocognitive development. The World Health Organization estimates that one in seven adolescents globally experiences mental health conditions, with digital behavior emerging as both a contributing factor and a potential window for early detection [7]. Moreover, the consequences of Cyber-Syndrome during adolescence may be more severe and lasting than in adulthood, as maladaptive cognitive and emotional habits formed during this critical window can become entrenched, increasing lifetime risk for mood disorders, attentional problems, and impaired metacognitive development.

Traditional approaches to assessing the cognitive and affective manifestations of Cyber-Syndrome have largely relied on self-report instruments (e.g., Beck Depression Inventory, State-Trait Anxiety Inventory) and laboratory-based cognitive tasks [8]. Although such tools have yielded valuable insights, they are inherently limited by subjective recall, static measurement, and coarse temporal resolution, failing to capture the dynamic, context-dependent nature of real-world mental states [9,10]. In parallel, digital traces generated through daily online activities including linguistic content, interaction patterns, and neurophysiological signals, have emerged as rich and ecologically valid sources for inferring mental states [11,12,13,14]. Using advanced computational methods to analyze these digital data has demonstrated considerable promise in achieving an objective, continuous, and multi-faceted mental state assessment [15,16].

Despite growing interest, the field lacks a synthesis that maps computational approaches for assessing the cognitive and affective dimensions of Cyber-Syndrome through digital behavior analysis. To address this gap, this paper presents a scoping review that charts the conceptual landscape, data sources, computational methods, and adolescent evidence base in this field, identifies gaps and open challenges, and offers guidance for clinical translation. The main contributions are as follows:

• Provide a systematic conceptual framework positioning cognitive and affective manifestations of Cyber-Syndrome within the broader CPST space paradigm, clarifying the theoretical underpinnings of computational mental state inference.
• Review state-of-the-art data sources and features for capturing cognitive and affective states, and examine computational methods ranging from Natural Language Processing (NLP) to Large Language Models (LLMs) for mental state inference.
• Synthesize current evidence and considerations for adolescent populations, and analyze critical challenges as well as future directions for advancing clinical translation and real-world deployment.

The remainder of this paper is organized as follows. Section 2 describes the search strategy and study selection process. Section 3 introduces the conceptual foundations of Cyber-Syndrome within the CPST space framework. Section 4 surveys the data sources and features used for mental state inference, organized into a modality-based taxonomy. Section 5 reviews the computational methods, ranging from text-based analysis and multimodal learning to cognitive computational modeling, temporal state modeling, and personalized modeling. Section 6 analyzes empirical evidence and considerations for adolescent populations. Section 7 examines the digital phenotyping platforms and research institutions driving the field, and Section 8 discusses critical challenges and future directions. Section 9 concludes the work.

2. Search Strategy and Scope

The literature search was conducted in May 2026 across five electronic databases: IEEE Xplore, PubMed, ACM Digital Library, Scopus, and Google Scholar. Search terms combined the constructs of “Cyber-Syndrome” and related digital-health concepts with computational terms (e.g., “digital behavior” AND “mental state assessment”; “affective computing” AND “cognitive impairment”; “machine learning” AND “computational psychiatry”), adapted to each database’s syntax. The search was limited to peer-reviewed journal articles and conference papers published in English within the last ten years. Foundational theoretical and methodological works outside this window were included where required for conceptual grounding.

Studies were included if they (a) proposed or validated computational methods for inferring cognitive or affective states from digital behavioral data, (b) addressed mental health outcomes relevant to the thinking-space dimension of Cyber-Syndrome, or (c) provided theoretical frameworks or systematic reviews connecting digital behavior to mental state assessment. Studies focused exclusively on physical health outcomes, non-digital assessment methods, or purely biological markers without behavioral correlates were excluded.

After removing duplicates, title and abstract screening yielded 412 potentially relevant records. Following full-text assessment, 259 primary studies, methodological papers, and theoretical frameworks were retained in the final scoping corpus, of which 45 specifically targeted adolescent populations. Screening was performed independently by two reviewers, with disagreements resolved by discussion or a third reviewer. Notably, we did not restrict our literature search to the adolescent population. Because the data sources and computational methods reviewed in this paper are, in principle, applicable across age groups, keeping the search unrestricted enabled us to capture the latest advances in the broader field of computational mental state inference. Studies specifically targeting adolescents were subsequently identified and analyzed as a dedicated subset, as discussed in Section 6.

3. Related Concepts

3.1. CPST Space Theory and Cyber-Syndrome

The CPST space theory was first proposed by Ning et al. in 2015, which emphasizes the ubiquitous connection and seamless convergence between physical, social, thinking and cyber spaces [17]. The physical space refers to the traditional material world where humans exist, encompassing bodily states, physiological processes, and physical environments. The social space represents the network of relationships, interactions, and social structures that connect individuals. The thinking space refers to the cognitive and affective domain of individual consciousness, including thoughts, emotions, beliefs, and mental representations, and the cyberspace is the virtual digital environment enabled by information and communication technologies. Under the facilitating role of cyberspace, the boundaries between physical, social, and thinking spaces become increasingly blurred, giving rise to the CPST hyperspace in which human activities, social interactions, mental processes, and digital environments co-evolve and mutually shape one another.

The Cyber-Syndrome refers to the physical, social, and thinking disorders resulting from problematic or overwhelming interaction with cyberspace [2]. The CPST space theory provides a comprehensive lens for understanding Cyber-Syndrome and its multi-dimensional manifestations [18]. In the physical space, they manifest as somatic issues such as myopia, neck pain, and other musculoskeletal problems associated with prolonged screen use. In the social space, they appear as social isolation, weakened interpersonal relationships, and reduced face-to-face communication. In the thinking space, they emphasize how excessive digital engagement translates into measurable changes in cognitive function, emotional experience, and mental health. In this paper, we further categorize the manifestations within the thinking space into two interrelated aspects, the cognitive aspect and the affective aspect, which will be introduced in the following section.

3.2. Cognitive and Affective Aspects of Cyber-Syndrome

Inspired by previous work, we identify the affective and cognitive aspects of Cyber-Syndrome in the thinking space, each with characteristic manifestations and corresponding computational approaches, as shown in Table 1.

In the cognitive aspect, Cyber-Syndrome manifests through distinct behavioral patterns and is measurable via computational approaches. Attentional deficits are evident through reduced sustained attention, heightened distractibility, and attentional bias toward digital cues [3,19,20]. These can be captured by app-switching patterns, session length variability, and eye-tracking metrics. Executive function impairments include poor impulse control, cognitive inflexibility, and decision-making deficits, and are assessable through keyboard dynamics, task-switching costs, and error rates [4,21,22]. Memory-related symptoms such as working memory overload, prospective memory failures, and source monitoring errors can be inferred from repeated search behaviors, task completion times, and app usage patterns [23,24,25]. Metacognitive disturbances are reflected in diminished self-awareness and discrepancies between self-perception and actual performance, and are often identified through linguistic markers of insight and behavioral-self-report mismatches [26,27]. Impairments in cognitive control, characterized by conflict monitoring deficits and reduced adaptive regulation, can be quantified via error-related signals, post-error slowing, and adaptation to changing task demands [22,28].

The manifestations of Cyber-Syndrome in affective aspects encompass a range of emotional dysregulations that are increasingly assessable through digital phenotyping. For example, mood disorders like depressive symptoms, anxiety, and anhedonia, are frequently detected using sentiment analysis, emotion word usage, and posting patterns from social media or smartphone interactions [13,29,30]. Emotion regulation deficits are marked by heightened reactivity, emotional inertia, and recovery difficulties [31,32,33]. These can be captured through affective variability metrics and the analysis of regulatory strategies embedded in natural language. Reward processing abnormalities including cue reactivity, tolerance, and withdrawal symptoms, are reflected in approach behaviors, consumption patterns, and post-engagement affective states [6,34,35]. Chronic stress and allostatic load are measurable via physiological signals such as heart rate variability and electrodermal activity, alongside circadian disruption and linguistic stress markers [36,37,38]. Social emotions such as social comparison, envy, and fear of missing out are observable through comparative language, engagement patterns, and signals of emotional contagion [39,40]. Collectively, these computational assessment approaches enable a granular, multimodal characterization of affective dysregulation in the context of Cyber-Syndrome.

Before reviewing the computational literature, we clarify the construct status of Cyber-Syndrome and its relation to neighboring frameworks. Cyber-Syndrome was defined as a conceptual construct describing the physical, social, and thinking disorders that arise from problematic cyberspace interaction [2,17,18]. Unlike Internet Gaming Disorder, which is included in the WHO International Classification of Diseases (ICD-11) [41] with operational diagnostic criteria, Cyber-Syndrome currently has no established nosology, diagnostic instruments, or prevalence estimates. Therefore, most computational markers described in Section 4 and 5 were developed and validated for general psychopathology, most commonly depression and anxiety, rather than for Cyber-Syndrome per se [11,12].

4. Data Sources and Features

Computational assessment of cognitive and affective states draws upon multiple data sources, each capturing different aspects of mental functioning [11,12]. Table 2 provides a comprehensive taxonomy organized by modality and psychological relevance.

4.1. Linguistic Markers

Natural language provides evidence for both cognitive and affective processes [13,29]. Digital communication via social media posts, messages, search queries, and typed content generates vast amounts of linguistic data that can be analyzed for cognitive and affective features. We organize linguistic markers along a continuum from surface-level form (lexical and stylistic features) to affective content (sentiment and emotion analysis) and finally to deep cognitive processes (reasoning styles, metacognition, and reality monitoring), moving from what is readily observable to what requires higher-level semantic and pragmatic inference.

4.1.1. Lexical and Stylistic Features

The words individuals choose consciously or unconsciously reflect their cognitive and emotional states. Among the various levels of linguistic analysis, lexical and stylistic features are the most directly accessible and interpretable. These features capture both what people say (vocabulary content) and how they say it (syntactic and typographical style), providing complementary perspectives on underlying psychological states.

Lexical features include word count, vocabulary diversity, part-of-speech distributions, and frequency of function words (e.g., pronouns, prepositions, articles). Stylistic features encompass sentence length, clause complexity, punctuation usage (e.g., exclamation marks, ellipses), and non-standard typography (e.g., capitalization, repeated letters). Relevant research has shown that ellipsis usage reflects speakers’ tendency to optimize processing efficiency by omitting redundant information, serving as a marker of cognitive load and discourse management [42]. Similarly, emoji and emoticon usage provide nonverbal affective cues that can clarify or amplify emotional tone [43]. Dubé et al. found a positive correlation between emotional intelligence and the frequency of emoji usage, indicating that the effectiveness of emojis as nonverbal emotional cues is consistent across different social situations [44].

While classic frameworks such as the Linguistic Inquiry and Word Count (LIWC) offer theory-driven dictionaries that map word usage onto psychological constructs, these lexical markers represent only the first step in linguistic analysis [45,46,47]. More nuanced interpretations of cognitive and affective states require dedicated methods for sentiment and reasoning styles, which we discuss in the following subsections.

4.1.2. Sentiment and Emotion Analysis

Sentiment and emotion analysis directly target the valence and categorical nature of emotional expression, providing a structured approach to mapping affective states from digital footprints [48]. Unlike LIWC’s simple counting of emotion-related words (e.g., “sad”, “happy”), modern sentiment analysis infers context-dependent emotional tone, offering complementary insights into mental health states.

Traditional approaches rely on sentiment lexicons such as SentiWordNet or VADER [49], while modern methods leverage supervised machine learning and transformer-based architectures to achieve higher accuracy in context-dependent sentiment detection [50,51,52]. For example, Zogan et al. developed a hybrid deep learning model that integrates multi-aspect features for explainable depression detection, demonstrating that specific emotion categories such as sadness and anger serve as strong indicators of mental health conditions [53]. Recent advances in transformer-based models, including BERT, RoBERTa, and LLMs, have substantially improved the accuracy and granularity of sentiment and emotion analysis from social media text [54,55,56]. These models capture contextual nuances and implicit emotional cues, and achieve state-of-the-art performance in mental health prediction tasks [13,57,58].

4.1.3. Cognitive Process Markers

Cognitive process markers tap into how a person thinks, which is closely linked to psychological well-being and psychopathology [59]. One key area is the identification of cognitive distortions, including catastrophizing (exaggerating potential negative outcomes), overgeneralization (drawing broad conclusions from a single event), and black-and-white thinking (viewing situations in extreme, all-or-nothing terms) [60]. Computational approaches have increasingly automated the detection of such distortions in text, enabling scalable assessment of maladaptive thinking patterns in naturalistic settings. For instance, dictionary-based methods and fine-tuned LLMs have been used to identify cognitive distortions in social media posts, with studies showing that elevated rates of absolutist and catastrophic language correlate with depression and anxiety severity [61].

Languages also reflect broader reasoning styles, particularly the distinction between analytic and intuitive thinking [62]. Metacognitive markers further extend this analysis by capturing expressions of self-awareness, insight, and reflection that are often impaired in psychiatric conditions [26]. Computational studies have linked reduced metacognitive language in clinical text and social media to poorer insight and diminished reflective functioning in psychiatric populations. Similarly, reality monitoring can be inferred from linguistic features such as perceptual detail, source attributions, and consistency in narrative accounts [27]. Impairments in reality monitoring, commonly observed in schizophrenia and dissociative conditions, provide a promising target for automated language-based assessment.

4.2. Interaction Dynamics

Human-device interactions offer continuous, implicit measures of cognitive and affective states [19,21]. Behavioral traces such as app switching patterns, typing dynamics, touchscreen gestures, and usage duration generate high-resolution temporal data that can be analyzed for markers of attentional control, executive function, emotional regulation, and reward processing.

4.2.1. Keystroke Dynamics

As users interact with digital devices through keyboards or touchscreens, each keystroke generates a sequence of timing events that reflect underlying cognitive processes, motor coordination, and attentional states [63]. Key metrics include typing speed, measured in words per minute or characters per second, which serves as a proxy for processing speed and psychomotor function. For example, the key hold time is associated with fine motor control and may be prolonged under conditions of fatigue or psychomotor retardation. Error rate, typically operationalized as the frequency of backspace or delete usage, provides an indicator of error monitoring and cognitive control. In addition, intra-individual variability across these metrics captures attentional fluctuations and cognitive stability, with increased variability often signaling lapses in sustained attention [64].

Research has demonstrated the performance of keystroke dynamics in detecting a range of mental health conditions and transient cognitive states. Studies show that typing patterns can reliably distinguish individuals with cognitive impairment from healthy controls, with slowed typing speed and prolonged flight time serving as salient markers [65]. Stress has also been found to modulate keystroke behavior, particularly through increased variability and altered flight time patterns [66]. In the context of depression, individuals exhibit characteristic changes in keystroke dynamics, including reduced typing speed, increased hold time, and higher error rates [67]. These findings underscore the potential of keystroke dynamics as an unobtrusive behavioral marker for real-time monitoring of cognitive and affective states.

4.2.2. Mouse and Touch Dynamics

Mouse and touchscreen interactions capture the continuous motor output of users as they navigate digital environments, providing a rich complement to keystroke-based assessments [68]. These interaction modalities generate high-resolution spatiotemporal data that reflect psychomotor function, cognitive conflict, and emotional arousal. For example, movement speed reductions have been associated with psychomotor retardation in depression and fatigue. Movement curvature has been linked to cognitive uncertainty and response conflict, with more curved trajectories emerging during periods of indecision or competing response tendencies [69]. Click or tap pressure, where measurable on pressure-sensitive devices, may index physiological arousal and emotional intensity, with increased force often accompanying heightened affective states [70]. Scroll patterns, including scroll speed and regularity, reflect reading engagement and comprehension, as well as attentional investment in content consumption.

Empirical studies have demonstrated the utility of mouse and touch dynamics in detecting cognitive and affective disturbances. Research has shown that individuals experiencing stress exhibit slower movement speed and increased trajectory curvature during pointing tasks, reflecting the impact of anxiety on motor planning and execution [71]. In clinical populations, impaired fine motor control captured through touchscreen gestures has been linked to executive dysfunction and psychomotor slowing [72]. Furthermore, the combination of mouse dynamics with other behavioral streams, such as click patterns and scrolling behavior, has enabled the identification of cognitive load states and attentional lapses with high temporal resolution.

4.2.3. App-Switching and Multitasking Patterns

App-switching patterns capture the dynamics of task management in real-world digital environments, revealing how individuals allocate attention across competing activities. Switch frequency, or the rate at which users transition between applications, serves as an indicator of attentional stability [19]. Task completion, assessed through whether initiated activities are brought to completion, provides a behavioral measure of goal-directed persistence and executive control. Sequential patterns, or the ordered structure of app transitions, reveal cognitive routines and task hierarchies, while interruption patterns offer insights into the integrity of attentional control mechanisms.

Accumulating evidence links app-switching and multitasking behaviors to cognitive and mental health outcomes. Heavy multi-taskers have been shown to exhibit poorer attentional filtering and increased susceptibility to distraction, suggesting that fragmented digital engagement may both reflect and reinforce attentional vulnerabilities [73]. In clinical samples, individuals with attention-deficit or hyperactivity disorder demonstrate higher app-switching rates and shorter session lengths compared to healthy controls [74]. Moreover, longitudinal analyses of smartphone usage patterns have revealed that increased switching frequency and reduced task completion predict greater symptom severity in depression and anxiety [75]. Specifically, Zhu et al. focused on the association between smartphone multitasking and depressive symptoms among college students in a multi-time point longitudinal cohort, and reported that smartphone multitasking behavior is significantly associated with an increased risk of depression [76]. These behavioral markers, when captured passively and continuously, enable real-time inference of cognitive control states and hold promise for identifying critical transitions in mental health status.

4.3. Temporal Patterns and Rhythms

The temporal patterns and rhythms people engage with digital devices reveal fundamental aspects of circadian function, sleep, and mental state dynamics [77,78]. Temporal features such as usage timing distributions, active and inactive periods, and daily regularity indices generate continuous streams of chronometric data that can be analyzed for markers of circadian rhythm stability, sleep-wake patterns, and the temporal structure of mood and cognitive states.

4.3.1. Circadian Rhythms

The timing of digital interactions provides insights into circadian function and sleep patterns. Activity timing reflects the alignment of an individual’s behavioral rhythm with the endogenous circadian clock. Periods of extended inactivity, particularly overnight, serve as reliable proxies for sleep timing and duration, enabling objective, longitudinal assessment without the need for self-report. Social jetlag, operationalized as the discrepancy in sleep timing between weekdays and weekends, captures the misalignment between biological and social time. Recent works have demonstrated that elevated social jetlag is consistently associated with mood disorders, including depression and bipolar spectrum conditions [79]. Circadian disruption, manifesting as irregular, fragmented, or phase-shifted activity rhythms, represents a transdiagnostic marker of mental health disturbance, with well-documented links to mood instability, cognitive impairment, and diminished well-being [80].

Empirical studies leveraging passively collected smartphone data have demonstrated that circadian rhythm metrics are sensitive to both clinical status and subclinical symptom fluctuations. Individuals with depression exhibit later activity onset, reduced daytime activity, and increased fragmentation of rest–activity cycles compared to healthy controls. In bipolar disorder, circadian disruption often precedes and predicts the onset of mood episodes, suggesting that temporal patterns may serve as early warning signals. Furthermore, the objectivity and continuous nature of passively derived circadian metrics make them particularly valuable for longitudinal monitoring, offering an ecologically valid complement to clinic-based assessments. When integrated with other behavioral and physiological streams, circadian features contribute to a multidimensional characterization of mental health that captures both stable trait-like patterns and dynamic state fluctuations.

4.3.2. Temporal Variability and Complexity

Beyond average patterns of activity timing, the variability and complexity of behavioral rhythms provide critical insights into mental health and cognitive functioning. Day-to-day variability in activity timing, duration, and intensity reflects the stability of daily routines. Individuals with mood disorders, particularly bipolar disorder, exhibit marked fluctuations in activity patterns that mirror their affective instability. Entropy measures, which quantify the predictability or randomness of temporal sequences, offer a sophisticated lens for characterizing behavioral dynamics. Reduced entropy has been observed in depression and anxiety, while increased entropy may characterize manic states [81].

Fractal scaling, which captures long-range temporal correlations in behavioral time series, provides another important dimension of temporal complexity. Healthy physiological and behavioral systems exhibit fractal scaling, characterized by self-similar fluctuations across multiple timescales that balance stability and adaptability. Disruptions in fractal scaling have been documented across a range of mental disorders, including depression, bipolar disorder, and schizophrenia, with deviations toward either excessively rigid or overly random dynamics signaling compromised regulatory capacity [82]. These complexity metrics, derived from passively collected digital behavioral data, offer a powerful approach for quantifying the integrity of self-regulatory systems. When tracked longitudinally, changes in temporal variability and complexity may serve as early indicators of impending mental health deterioration, enabling timely intervention.

4.3.3. Seasonal and Long-Term Patterns

Temporal patterns in digital behavior extend beyond daily and weekly cycles to encompass seasonal and multi-year trajectories, offering insights into the longer-term dynamics of mental health. Seasonal variation in activity timing, sleep patterns, and digital engagement reflects the influence of changing photoperiod and environmental factors on mood and behavior. Reduced activity levels, later sleep onset, and diminished social engagement during winter months are characteristic of seasonal affective disorder, and passive monitoring of digital behaviors can capture these seasonal shifts with high fidelity. For individuals with mood disorders, seasonal patterns may manifest as predictable periods of symptom exacerbation or remission, and identifying these seasonal rhythms can inform personalized treatment planning and preventive interventions.

Long-term trends captured over months or years reveal the trajectory of mental health conditions, including the progression of illness, response to treatment, and risk of relapse. Gradual shifts in activity timing, such as persistent phase delay, may signal the emergence of depressive episodes in vulnerable individuals. Similarly, sustained changes in social interaction patterns, app usage diversity, or communication frequency may reflect evolving clinical status. The accumulation of longitudinal digital behavioral data enables the construction of individualized baselines against which deviations can be detected, supporting a personalized approach to mental health monitoring. When combined with clinical assessments and self-report, these long-term temporal patterns enhance our ability to understand illness trajectories, evaluate intervention effectiveness, and identify critical windows for preventive care.

4.4. Neurophysiological Signals from Wearables

Consumer wearable devices increasingly provide continuous physiological data relevant to mental state inference [36,37]. These devices enable passive, longitudinal monitoring of autonomic nervous system activity and brain function, offering objective biomarkers that complement self-report and behavioral measures.

4.4.1. Heart Rate and Heart Rate Variability

Heart rate and heart rate variability (HRV) are among the most widely studied physiological markers in digital mental health. Average heart rate and beat-to-beat intervals index sympathetic arousal and stress levels, while HRV metrics capture autonomic nervous system balance [83]. Reduced HRV has been consistently associated with cognitive control deficits, particularly in tasks requiring attentional regulation and response inhibition [84]. In the affective domain, low HRV is a robust correlate of anxiety disorders, depression, and emotional dysregulation, suggesting diminished capacity for adaptive self-regulation [85].

4.4.2. Electrodermal Activity

Electrodermal activity (EDA) captures changes in skin conductance driven by sympathetic nervous system activation, providing a sensitive index of arousal and emotional responding [86]. Tonic EDA, measured as baseline skin conductance level, reflects sustained arousal states, while phasic EDA captures event-related fluctuations in response to specific stimuli or cognitive demands [87]. The frequency and amplitude of EDA peaks serve as indicators of acute stress, emotional reactivity, and surprise. In mental health applications, EDA has been used to detect stress responses, panic attacks, and emotional dysregulation episodes, with wearable sensors enabling continuous monitoring in real-world settings.

4.4.3. Electroencephalography

Recent advances in electroencephalography (EEG) devices have made continuous brain activity monitoring increasingly feasible [88]. These lightweight, dry-electrode systems capture key neural oscillatory patterns relevant to cognitive and affective states. For example, Theta activity (4-8 Hz) is associated with cognitive load, sustained attention, and memory encoding processes. Alpha activity (8–12 Hz) reflects cortical inhibition and relaxation states. Beta activity (12-30 Hz) correlates with active concentration, alertness, and motor preparation. Frontal alpha asymmetry, defined as the difference in alpha power between left and right frontal regions, has been extensively studied as a marker of depression risk and trait negative affect [89].

4.4.4. Multi-Sensor Integration

Modern wearable platforms increasingly integrate multiple physiological sensors, enabling multimodal assessment of neurophysiological states [37]. Smartwatches combine photoplethysmography (PPG) for heart rate monitoring, accelerometry for movement detection and EDA sensors. Smart rings offer continuous PPG and accelerometry along with skin temperature measurements, providing robust sleep and circadian monitoring. Multimodal wristbands integrate PPG, EDA, accelerometry, and temperature into a single form factor, supporting simultaneous assessment of cardiovascular, electrodermal, and motor activity. These integrated platforms facilitate the fusion of complementary physiological signals, enabling more robust inference of mental states.

5. Computational Methods

5.1. Mental state inference based on text

As the primary vehicle for individual thought, emotion, and cognitive expression, text data contains rich semantic features and latent psychological cues, serving as an indispensable foundation for precise and non-invasive mental state inference. NLP has emerged as a primary tool for inferring mental states from digital text [15,16].

5.1.1. Traditional Text Representation Methods

Traditional text processing and feature extraction methods have laid an important foundation in research on psychological trait recognition and mental health screening. The following section primarily explores the specific applications, effectiveness, and limitations of lexicon-based methods (e.g., Linguistic Inquiry and Word Count, LIWC), the Bag-of-Words (BoW) model, and Term Frequency-Inverse Document Frequency (TF-IDF) in capturing psychological representations. Lexicon-based methods, particularly the theory-driven LIWC framework, quantify psychological constructs via predefined dictionaries [45]. Recent studies demonstrate that integrating LIWC with topic modeling or specialized lexicons (e.g., cultural or clinical domain-specific dictionaries) effectively uncovers distinct age-specific linguistic patterns [90], and predicts depression severity with high stability [91]. BoW model extracts word frequency vectors for inductive text classification [92,93]. A study demonstrated that combining BoW with classification models can achieve an accuracy of up to 86% in depression recognition [94]. TF-IDF identifies idiosyncratic linguistic markers by evaluating word relevance [95]. Research shows that TF-IDF-based logistic regression effectively identifies depressive symptoms, thereby enhancing the efficiency of mental health screening [96].

5.1.2. Word Embeddings

Word embeddings utilize continuous vectors to represent word semantics [97]. Static methods such as Word2vec and GloVe assign a fixed representation to each word [98,99]. A study on Twitter-based depression detection evaluated three word embeddings (Word2Vec, GloVe, and FastText) across eight machine learning classifiers [100]. It reveals that the proposed GloVe-based stacked ensemble model achieved 97% accuracy, significantly outperforming individual classifiers and traditional BoW approaches. Kunder et al. investigated mental health text classification and evaluated the impact of Word2Vec-based text vectorization coupled with deep learning architectures to detect emotional states such as stress, anxiety, and sadness. The results demonstrated that integrating Word2Vec embeddings significantly enhanced model training efficiency and classification accuracy [101].

5.1.3. Transformer-based contextual embeddings

Transformer-based contextual embeddings (such as BERT, RoBERTa, and ELECTRA) leverage surrounding context to capture nuanced semantic features [54,55], enabling the accurate identification of subtle linguistic markers associated with depression [102]. Recent work has further demonstrated the effectiveness of BERT-based models for social media-based depression detection [13,53]. Specifically, a study focusing on depression detection from social media texts employed BERT and RoBERTa on the Reddit depression dataset hosted on the Kaggle platform, validating the superior performance of deep pre-trained language models in social media depression recognition [103]. Additionally, Salameh et al. leveraged RoBERTa to capture deep contextual semantic features and integrated a Convolutional Neural Network (CNN) for hierarchical feature extraction, achieving an accuracy of 96% on the Reddit depression dataset [104].

5.1.4. LLMs

LLMs such as GPT, LLaMA, and Claude show great promise for text-based mental health detection [57,105]. Leveraging zero/few-shot learning or targeted fine-tuning, these models efficiently infer mental states [56,58]. A study based on 428 emotional diary entries from 91 participants applied LLMs to depression risk detection. The results demonstrated that the fine-tuned GPT-3.5 achieved an accuracy of 90.2%, while the zero-shot GPT-3.5 reached a balanced accuracy of 84.4% [106]. Additionally, LLMs possess excellent generative explainability, providing intuitive natural language explanations that significantly enhance model interpretability and trustworthiness in medical scenarios [13,107]. Another study indicated that fine-tuned large models achieved an accuracy of 96% in depression detection from social media posts, outperforming traditional approaches [58].

5.2. Mental State Inference Based on Multimodal Data

5.2.1. Attention Mechanism

As the core computational module of the Transformer architecture, the self-attention mechanism precisely models global dependencies by dynamically assigning importance weights to each element within a sequence [54,108]. Its derived variants have been widely applied to multimodal modeling across various scenarios (such as text, images, speech, video, and EEG signals), significantly expanding the application boundaries of the model. In multimodal psychological inference, cross-modal attention enables dynamic, directional interactions by allowing one modality to attend to another [109]. To address the prevalent pain points in audio-text multimodal depression detection, a study proposed a text-guided cross-modal feature reconstruction and decomposition framework. This framework achieves the decoupling of shared and private features between audio and text modalities, outperforming the state-of-the-art (SOTA) methods of the same period [110]. A newly proposed synchronous multimodal Transformer can effectively fuse electroencephalogram and clinical interview data, and capture deep inter-modal correlations through bidirectional cross-attention, thus achieving improved accuracy in depression detection [111,112]. Furthermore, hierarchical attention constructs weights across multiple levels to capture multi-scale structural characteristics [113]. Another study integrated three modalities: text, images, and user auxiliary information. Through a dual-layer attention mechanism—operating both intra-modally and inter-modally—the model automatically screens for key depression-related information, achieving a classification accuracy of 92.84% on the test set [114]. In the field of multimodal psychological and affective computing, co-attention mechanism is widely utilized to address feature mismatch issues across heterogeneous data [115]. A study proposed the ViLBERT model, which pre-trains joint image-text representations through an innovative two-stream architecture and co-attention mechanism, successfully transforming cross-modal alignment into a transferable, generalized capability and achieving remarkable success across multiple classic vision-language tasks [116].

5.2.2. Multimodal LLMs

Multimodal LLMs (MLLMs) provide a groundbreaking paradigm for multimodal mental state inference by extending the powerful reasoning and generative capabilities of text-based LLMs to diverse behavioral modalities [117,118,119]. A study proposes MLlm-DR, capable of understanding vocal and visual inputs and supporting interpretable depression diagnosis by integrating a lightweight LLM with a lightweight query module (LQ-former) [120]. To address the limitation that existing learnable queries lack affective semantic constraints, a study proposes the Prototype-as-Prompt framework, which maps audio-visual representations onto a fixed set of multimodal affective prototypes and leverages them as soft prompts to guide LLMs in multimodal emotion analysis [121]. A study proposed the LLaVA model, which connects a vision encoder to a large language model (LLM) and is fine-tuned on machine-generated image-text instruction data, enabling general-purpose visual and language understanding alongside impressive multimodal chat capabilities [122]. A study proposed the BLIP-2 model, which introduces a lightweight Querying Transformer module to successfully bridge a frozen image encoder and a large language model (LLM), enabling general-purpose vision-language understanding and zero-shot image-to-text generation with minimal trainable parameters [123]. A study proposes the Emotion-LLaMA model, which achieves seamless fusion of audio, visual, and textual input information through dedicated encoders designed for affective tasks. Built upon an improved, instruction-tuned LLaMA architecture, the model significantly enhances its capabilities in emotion recognition and emotion reasoning [124].

5.2.3. Contrastive Learning

In multimodal data processing, Contrastive Learning, as a powerful self-supervised representation learning paradigm, plays a core role in bridging the heterogeneity gap between different modalities and achieving cross-modal semantic alignment in the feature space [125]. To address the difficulties in capturing subtle contextual semantic features and the problem of imbalanced class distribution in datasets for the Emotion Recognition in Conversation (ERC) task, a study proposes a prompt-based hybrid supervised contrastive learning model, which outperforms 15 state-of-the-art ERC models on the MELD and IEMOCAP datasets [126]. Another study proposed a cross-modal alignment method based on contrastive learning to transfer emotion-semantic-rich EEG features to easily collected and low-artifact ECG signals, achieving promising performance in emotion recognition [127]. Zhu et al. constructs a graph cognitive modeling framework integrated with contrastive representation learning. Based on student interaction data, this model can accurately complete student behavior classification and clickstream prediction, improving intelligent education systems and reducing the risk of student dropout [128].

5.2.4. Multimodal Fusion Strategies

Multimodal fusion strategies primarily encompass early, late, intermediate, hybrid, and attention-based approaches. Early fusion concatenates raw inputs to preserve source correlations [129], whereas late fusion combines independent model predictions at the output layer to mitigate modality heterogeneity [130]. For deeper interaction, intermediate fusion integrates high-level feature representations within mid-level network layers [131]. One study proposed two audio-text multimodal depression detection neural networks based on early and late fusion strategies, validating the superiority of the early fusion approach in the depression detection task [132]. Additionally, another study systematically compared the performance of three strategies—early, intermediate, and late fusion—and confirmed that intermediate fusion achieved optimal performance when processing multimodal electroencephalogram (EEG) and eye-tracking data [133]. Additionally, hybrid fusion employs multiple fusion points for progressive, multi-stage feature integration [134], while attention-based fusion dynamically scales modality weights according to context to enhance the robustness of psychological state inference [109]. A study focused on the field of multimodal depression detection proposed a hybrid fusion strategy that combines intra-modal early fusion with inter-modal late fusion. A corresponding feature extraction fusion module and an attention-based decision fusion module were designed, enabling accurate prediction of depression severity from audio and video modalities [135].

5.3. Cognitive Computational Modeling

The key advantage of cognitive modeling is explanatory power—identifying which specific cognitive or affective processes are altered [136,137]. Traditionally, cognitive models have been applied to laboratory task data. Recent work extends them to real-world digital behavior [10,136]. For example, users’ social media engagement can be viewed as a reward learning process where likes and comments serve as social rewards that continuously influence individuals’ posting behaviors [138], and users’ choices and switches between different applications essentially reflect ongoing cost-benefit computations occurring in the brain [19]. In addition, the latency of message responses can effectively demonstrate the trade-offs that individuals make between immediate and delayed rewards [139].

5.3.1. Reinforcement Learning

Reinforcement learning (RL) formalizes how individuals learn from rewards and punishments [140]. In the field of computational psychiatry, this model has been widely applied to uncover the underlying mechanisms of various psychiatric disorders [141]. Research indicates that depression is generally characterized by decreased reward learning rates [142], while anxiety disorders are often accompanied by abnormally heightened punishment sensitivity [143], and individuals across multiple psychopathological conditions also exhibit substantial alterations in the exploration-exploitation trade-offs [144]. A study focusing on a systematic review of depression and RL confirmed that RL-related parameters can serve as endophenotypes for depression, providing a novel perspective for pathological analysis [145]. Additionally, a meta-analysis revealed that the pathology of depression is not an abnormality of a single system but rather a synergistic effect of low reward approach sensitivity and high threat avoidance sensitivity [146].

5.3.2. Drift Diffusion Models

Drift Diffusion Models (DDMs) decompose response times into latent decision processes [147] and explain decision mechanisms through four core parameters: drift rate (efficiency of evidence accumulation), boundary separation (the speed-accuracy trade-off), non-decision time (perceptual and motor latency), and starting point (initial bias). These parameters are effectively estimated from naturalistic digital behavior, thereby providing quantitative support for psychological state inference [148]. One review fully elaborates on the core parameters of DDM and systematically summarizes the aberrant patterns of DDM parameters across five brain disorders, confirming that DDM can serve as a quantitative tool for cognitive deficits in decision-making [149]. Another systematic review traces the developmental history, core theoretical framework, and mechanistic basis of key DDM parameters, summarizing its extensive applications across diverse fields and addressing controversies regarding the model’s identifiability and flexibility [150]. Additionally, a study demonstrates that reward positivity is associated with the evidence accumulation efficiency for high-reward options, further validating that DDM can precisely delineate the specific sub-processes through which depression impacts reward-based decision-making [151].

5.4. Temporal State Modeling

Mental states are inherently dynamic, evolving over multiple timescales [77]. Accordingly, analyses based merely on a single timescale tend to obscure the complete profile of psychological dynamics. The construction of Multi-scale models enables researchers to effectively integrate and capture the inherent interactions and feedback mechanisms among different timescales. This approach facilitates a systematic understanding of the full lifecycle of mental state evolution [152].

5.4.1. Vector Autoregressive Model

Vector Autoregression (VAR) models effectively capture complex multivariate temporal dependencies [153], utilizing Granger causality to identify directional behavioral influences and impulse response functions to simulate shock propagation. A study enrolling 56 patients with anxiety disorders employed multilevel vector autoregressive (mlVAR) network analysis to investigate pre- to post-treatment changes in inter-symptom associations. The study confirmed that Cognitive Behavioral Therapy (CBT) significantly reduced symptom network density and severed the pathological concurrent association between negative emotion and avoidance behavior, thereby elucidating the core therapeutic mechanism of CBT for anxiety disorders [154]. A study addressing the core clinical challenge of high comorbidity between anxiety and depression fitted a mlVAR model to analyze the day-to-day dynamic interactions between symptoms. It revealed that the directional predictive effect of depressive symptoms on anxiety symptoms was significantly stronger than the reverse predictive effect of anxiety symptoms on depressive symptoms [155].

5.4.2. Temporal Neural Network Models

Advancements in deep learning offer robust computational frameworks for temporal state modeling, supporting the fine-grained fusion of multimodal sequential data [14,156] and enabling automated, end-to-end extraction of time-varying features from raw data [157]. To capture sequential dynamics, RNN variants like LSTM and BiLSTM are widely used to model text structures and behavioral interaction patterns, effectively securing bidirectional temporal dependencies [19,53,158,159]. Existing studies indicate that BiLSTM networks can be applied to model text sequences from clinical interviews and fused with audio and video modalities to achieve depression detection and severity assessment [160]. Concurrently, LSTM networks can be employed to model the temporal dynamics of social media data, thereby enhancing depression recognition performance by integrating text, images, and posting timestamps [161]. Furthermore, a hybrid CNN-LSTM model has been adopted to model behavioral sequences in videos, enabling the joint recognition of depression and anxiety [162]. Alternatively, Temporal Convolutional Networks (TCNs) serve as an efficient substitute, offering prominent parallelization advantages and interpretable receptive fields for temporal feature extraction [163]. The successful application of TCNs in multimodal emotion recognition demonstrates that TCNs possess inherent advantages in modeling long-range temporal contexts and multi-scale dynamic features [164].

5.4.3. State-space modeling

As mental states are often not directly observable, state-space models are further employed to infer the underlying latent mental states from observable digital behaviors [165]. Among them, Hidden Markov Models (HMMs) are applied to conduct staged modeling for discrete latent states with Markovian transition characteristics [166]. A study confirmed that the total dwell time in the socially active state derived from the model can serve as an objective digital biomarker for social functional impairment in neuropsychiatric disorders [167]. Another study has verified that the HMM integrated with subjective well-being, emotion regulation and coping styles can more effectively predict individuals’ prospective risk of depression [168]. Jiang et al. proposed the Hidden Markov Model with Multivariate Autoregressive observation, which can effectively extract neuronal dynamic features from EEG signals and interpret the pathogenesis of brain disorders from the perspective of state transitions [169].

Kalman filters excel at tracking continuous latent states within linear dynamics. A study applied this method to depression-related Ecological Momentary Assessment (EMA) data, successfully characterizing emotional dynamics, environmental reactivity, and state controllability under depressive conditions [170]. Other studies aim to construct an accurate galvanic skin response detection system that introduces the Kalman filter algorithm for noise reduction to effectively identify and analyze stress levels [171].

5.5. Personalized Modeling

Substantial individual differences in baseline cognitive and affective functioning necessitate personalized approaches [172].

5.5.1. Meta-Learning

The application of meta-learning represents an advanced paradigm in personalized modeling—utilizing a "learning to learn" mechanism to achieve fast adaptation to an individual’s unique affective or cognitive profile [173,174]. A study proposes a novel meta-learning framework that captures generalized emotional feature cues and undergoes customized fine-tuning tailored to individual annotation styles, thereby effectively enabling personalized speech emotion recognition [175]. Another study proposed an integrated multimodal emotion recognition framework augmented with meta-learning, which models the contextual dependencies among audio, video and electroencephalography signals and greatly improves the model’s cross-task adaptability [176]. To address the heterogeneity of neural signals, Chen and Fang et al. proposed a model-agnostic meta-learning algorithm for EEG-based emotion recognition that achieves state-of-the-art cross-subject accuracy through one-shot adaptation, significantly enhancing the generalizability of future affective brain–computer interfaces [177].

5.5.2. Multi-task Learning

The concept of multi-task learning (MTL) approaches personalization by treating the predictive modeling of each unique individual or distinct subpopulation as an independent yet correlated learning task [178,179]. Focusing on personalized text-based emotion recognition in multi-task scenarios, a study proposed a novel personalized deep learning architecture that performs emotion recognition from the user’s own perspective by computing human subjective biases and learning user-specific vector representations [180]. Another study proposed a deep multi-task learning framework that can simultaneously accomplish sentiment analysis and emotion recognition [181]. Zhao et al. proposed a selective feature-sharing multi-task network for facial expression recognition and synthesis, which effectively mitigates task interference and achieves state-of-the-art performance through data augmentation [182].

5.5.3. Transfer learning

Transfer learning provides a highly efficient mechanism for personalized modeling—transferring generalized psychological, behavioral, or physiological representations from a large source population to a data-scarce target individual [183,184]. To address individual differences in cross-subject EEG emotion recognition, a study proposed a hybrid transfer learning strategy (DFF-Net) integrating domain adaptation and few-shot fine-tuning, effectively enhancing recognition accuracy [185]. Another study proposed a multi-source data sentiment analysis framework using a swarm intelligence algorithm to optimize deep transfer learning, effectively enhancing classification accuracy and generalization for diverse emotional categories [186].

5.5.4. Digital Twins

An emerging concept is that of affective digital twins, personalized computational models that simulate an individual’s cognitive and affective processes [187]. This model possesses three major characteristics: continuous updating, counterfactual simulation and privacy protection capabilities. A study systematically proposed the framework of the Digital Cognitive Twin, constructing an adaptive cognitive health ecosystem through multimodal sensor fusion and AI-driven adaptive gamification [188]. Another study proposed digital twin frameworks that construct virtual replicas of patients by integrating multi-dimensional data and artificial intelligence technologies, which can be used for early diagnosis, risk assessment, personalized treatment simulation, and real-time monitoring in mental health care [189]. Ashokbhai et al. proposed a domain-agnostic methodological framework for a Cognitive Digital Twin (CDT) that augments, rather than replaces human decision-making by explicitly modeling cognitive workload and affective state [190].

6. Evidence and Considerations for Adolescent Populations

Adolescence is a critical period for rapid cognitive development and the maturation of emotional regulation mechanisms, making adolescents highly sensitive to changes in the external environment, especially the shift towards digital lifestyles. The data sources and computational methods reviewed in Section 4 and 5 are, in principle, applicable to all age groups. However, applying them to adolescents requires careful adaptation. This section explores how the general computational methods discussed above can be applied to adolescents, referring to the research summarized in the Appendix Table A1.

6.1. Patterns across Adolescent Studies

First, the evidence base spans a wide developmental window, from early adolescence (ages 10-14 in [191,192]) through late adolescence and young adulthood (ages 18-25 in [36,193]). Large-scale general-population cohorts [194,195] coexist with a substantial cluster of university-student samples [39,46,155,196,197,198] and a growing number of smaller clinical and feasibility studies concentrated in the 13-18 age band [191,199,200,201,202,203,204,205,206,207,208,209,210,211]. Second, the table shows a pronounced methodological divide. Questionnaire- and scale-based self-reports dominate the survey-oriented work [39,40,196,200,201,212], while passively sensed smartphone data collected through applications such as AWARE [206], Beiwe [211], EARS [208], and mindLAMP [213] anchor a separate cluster of sensor-based studies [202,203,205,207,209,214,215]. EMA-based affective tracking [155,207,208,216], wearable physiological sensing [193,204,206], keyboard or touchscreen traces [191,210], and speech-based analysis [217] appear in a smaller but growing number of studies, illustrating an ongoing shift toward continuous, multimodal digital phenotyping. Third, the computational methods used in adolescent samples are correspondingly diverse, ranging from classical statistical approaches such as regression, specification curve analysis, and multilevel modeling [9,33,40,191,194,195,200,203,212] to machine learning and deep learning pipelines [218], speech-based architectures [217], multimodal deep learning [199], and personalized modeling strategies [206,207]. Finally, several studies confirm that personalized and idiographic models consistently outperform generalized approaches in adolescent populations, reflecting the high inter-individual variability characteristic of this developmental stage.

6.2. Developmental Specificity

The table also documents that age itself moderates several marker-symptom associations in ways that would be missed in adult-only studies. The relationship between social media use and life satisfaction, for instance, is most strongly negative among younger adolescents and weakens with increasing age [194]. Keystroke and touchscreen features collected from adolescent devices show associations with depressive symptoms measured by the PHQ-9 [210], and linguistic markers derived from mobile communication including the use of first-person singular pronouns predict depression risk in adolescents [209]. Similarly, studies comparing age groups reveal different linguistic patterns in how adolescents and adults express depression [90]. The dynamics of positive and negative affect also vary developmentally: inertia and variability of emotional states follow distinct patterns as a function of clinical depression and depression risk in adolescent samples [216], and reward-processing metrics derived from cognitive models predict child and adolescent depressive outcomes [151]. Beyond symptom associations, adolescent-specific risk pathways are also documented in the table, including the association between problematic smartphone use and anxiety in early adolescents [219] and the mechanisms linking social media and smartphone use to self-harm within clinical adolescent cohorts [220]. These age-related findings caution against the assumption that computational tools developed and validated only in adult populations will transfer to adolescents. Rather, their validity in this age group cannot be presumed without age-stratified revalidation.

7. Platforms and Institutions

The practical implementation of computational methods for mental state inference depends heavily on a robust ecosystem co-constructed by pioneering research institutions and specialized digital phenotyping platforms. These academic centers and software frameworks have fundamentally reshaped the paradigm of psychiatric assessment, transitioning it from a fragmented, subjective self-report modality into a continuous, objective behavioral sensing paradigm with high ecological validity. This transition is realized by leveraging ubiquitous personal digital devices to capture high-fidelity, multimodal digital footprints. These data streams encompass passive hardware sensor logs, naturalistic smartphone interaction patterns, and active cognitive tasks alongside ecological momentary assessment (EMA) data.

To map and synthesize the core forces driving the advancement at the frontier of this field, this section conducts a comprehensive evaluation across two pivotal dimensions: research institutions and collaborative platforms. Table 3 provides a granular technical analysis of mainstream open-source collaborative software platforms, while Table 4 focuses on leading global research laboratories and institutions. The methodologies of these pioneering institutions and the capabilities of open-source platforms complement each other, laying a solid ecological foundation for the paradigm shift in digital psychiatry from an experience-driven to a data-driven approach.

2pt

2pt

8. Challenges and Future Directions

8.1. Measurement Reliability and Mechanistic Modeling

Establishing reliable ground truth for mental states remains a fundamental challenge [245]. Self-reported measures correlate only moderately with clinical assessments and are subject to recall bias — limitations that are especially pronounced in adolescents, whose introspective ability is still developing. Clinical diagnoses impose categorical labels on dimensional phenomena, and digital data timestamps often fail to align with symptom assessments, creating temporal mismatches. While most current research emphasizes prediction over explanation [136], a promising path forward lies in theory-driven computational modeling that estimates parameters of cognitive processes (e.g., reinforcement learning parameters, drift-diffusion rates) from real-world digital behavior [136,137]. This mechanistic approach can bridge behavioral observation and targeted intervention, and is particularly valuable for adolescents, where developmental changes require separating normative maturation from emerging pathology.

8.2. Individual Variability and Personalized Modeling

Digital behavior data contains substantial measurement noise, and dramatic individual differences in baseline usage patterns mean absolute values of digital metrics carry less meaning than deviations from personal norms [172,246]. For adolescents, this challenge is compounded by rapid developmental change, where baselines shift naturally over months. Personalized modeling approaches — including multi-task learning, domain adaptation, and the emerging concept of affective digital twins — offer a path forward by constructing individualized computational models that simulate each person’s cognitive and affective processes [187]. These models continuously update with new data, enable counterfactual simulation of interventions, and can reside on-device for privacy, making them especially suitable for long-term adolescent mental health monitoring.

8.3. Privacy, Consent, and Ethical Safeguards

Mental health data ranks among the most sensitive personal information, and digital footprints can inadvertently reveal intimate psychological details [247,248]. For minors, these concerns are heightened: informed consent requires both parental permission and adolescent assent, and regulatory frameworks such as COPPA and GDPR-K impose additional requirements [249]. Privacy-preserving technologies including federated learning, differential privacy, and on-device processing are essential enabling tools [250,251,252]. These allow model training across decentralized data without raw data sharing, ensuring that analysis remains local with only anonymized summaries transmitted. Future platforms serving adolescent populations must integrate these protections by design, not as an afterthought.

8.4. Algorithmic Fairness and Interpretability

Machine learning models risk perpetuating biases when certain demographic groups — including specific age bands, genders, or cultural contexts — are underrepresented in training data [253,254]. Age itself can be a source of bias: models trained primarily on adult data may misclassify normative adolescent behaviors as pathological [255]. Furthermore, complex deep learning models often function as opaque “black boxes” that provide limited insight into their predictions, a significant barrier to clinical adoption [256,257]. Addressing these challenges requires diverse training datasets, rigorous subgroup validation, and models that are both interpretable and actionable. LLMs offer particular promise here, as they can generate human-readable explanations of their inferences and engage in conversational assessment, enhancing both transparency and trust in clinical settings [13,58,105].

8.5. Multidimensional Integration

No single data source captures the full complexity of mental states, and comprehensive assessment requires integrating across modalities — linguistic, behavioral, temporal, and physiological signals — as well as across the broader CPST dimensions [14,17,18,156]. For adolescents, this means combining digital behavior data with developmental context: social network analysis reveals how peer influence shapes mental health, physiological sensing captures stress responses during school hours, and temporal patterns track sleep and circadian changes that are particularly pronounced during adolescence. The pathway from isolated markers to integrated, real-world assessment tools will require interdisciplinary collaboration among clinicians, data scientists, and developmental researchers, alongside prospective validation in adolescent populations [257].

9. Conclusions

This scoping review maps and synthesizes current knowledge on the computational assessment of the cognitive and affective dimensions of Cyber-Syndrome. It shows that computational approaches support objective, continuous, and ecologically valid inference of mental states from digital behavior, offering a powerful paradigm for assessing the thinking-space dimension of Cyber-Syndrome within the broader CPST framework. For adolescent populations, the reviewed evidence underscores that age-stratified calibration and personalized modeling are essential, because developmental changes shape the associations between digital markers and mental health outcomes. Nevertheless, major challenges must be resolved before clinical translation can succeed. Future work should prioritize theory-driven mechanistic modeling, privacy-preserving technologies, personalized affective digital twins, and multidimensional integration across the CPST framework. Such advances would move computational mental health from prediction toward actionable, ethical, and personalized care.