Abstract
Background and aims
Gaming disorder [GD] risk has been associated with the way gamers bond with their visual representation (i.e., avatar) in the game-world. More specifically, a gamer's relationship with their avatar has been shown to provide reliable mental health information about the user in their offline life, such as their current and prospective GD risk, if appropriately decoded.
Methods
To contribute to the paucity of knowledge in this area, 565 gamers (Mage = 29.3 years; SD =10.6) were assessed twice, six months apart, using the User-Avatar-Bond Scale (UABS) and the Gaming Disorder Test. A series of tuned and untuned artificial intelligence [AI] classifiers analysed concurrently and prospectively their responses.
Results
Findings showed that AI models learned to accurately and automatically identify GD risk cases, based on gamers' reported UABS score, age, and length of gaming involvement, both concurrently and longitudinally (i.e., six months later). Random forests outperformed all other AIs, while avatar immersion was shown to be the strongest training predictor.
Conclusion
Study outcomes demonstrated that the user-avatar bond can be translated into accurate, concurrent and future GD risk predictions using trained AI classifiers. Assessment, prevention, and practice implications are discussed in the light of these findings.
Introduction
Since their commercial conception in the 1970s, videogames have become integrated into modern popular culture (Will, 2019). Alongside a boom in technological advancements and improved internet capabilities, the gaming industry has developed into a global community allowing millions around the world, and in Australia (where the present study was carried out), to enjoy gaming as a shared activity (Statista, 2023; Stavropoulos, Motti-Stefanidi, & Griffiths, 2022).
In the past two decades, gaming has greatly proliferated, with recent nationwide data suggesting that approximately 70% of all Australians (i.e., 17 million) play videogames in some form or frequency, while the vast majority of households (i.e., 8.6 million), including those with children, have access to digital game devices (Brand, Todhunter, & Jervis, 2017). Alongside the growth of gaming, gaming pathologies have begun to emerge (King et al., 2020). Literature highlights that while most gamers enjoy positive outcomes such as psychomotor/dexterity, cognitive, health, and educational benefits (Granic, Lobel, & Engels, 2014; Koulouris, Jeffery, Best, O'Neill, & Lutteroth, 2020; Nuyens, Kuss, Lopez-Fernandez, & Griffiths, 2017; Raith et al., 2021; Watson et al., 2019), a minority of gamers may experience harmful effects associated with excessive and/or disordered gaming (e.g., reduced educational/work performance, distress, loneliness; Burleigh, Griffiths, Sumich, Stavropoulos, & Kuss, 2019; Nuyens, Kuss, Lopez-Fernandez, & Griffiths, 2019; Şalvarlı & Griffiths, 2022; Stavropoulos et al., 2019; Colder Carras, Stavropoulos, Motti-Stefanidi, Labrique, & Griffiths, 2021; Van Looy, 2015; Šporčić & Glavak-Tkalić, 2018).
There is consensus that disordered gaming occurs as a consequence of the interplay between factors related to the individual players (e.g., personality, psychopathology), their immediate and more distant environmental surroundings (e.g., adverse family/peer interactions), as well as the game applications themselves (e.g., reinforcement schedules; King et al., 2019; Starcevic & Khazaal, 2020; Stavropoulos, Rennie, Morcos, Gomez, & Griffiths, 2021). For instance, in relation to individual factors, Király, Koncz, Griffiths, and Demetrovics (2023) highlighted disordered gaming risk factors including gender (being male), age (being younger), personality traits (higher neuroticism, higher impulsivity, low self-esteem), comorbidities (e.g., anxiety, autistic behaviours), motivation factors (e.g., escapism), and neurobiological predispositions (e.g., reduced grey-matter volume in the ventromedial and dorsolateral prefrontal brain areas). In relation to environmental factors, disordered gaming risk factors include poor quality of family relationships and parental monitoring, childhood maltreatment and easy access to gaming equipment, as well as pro-gaming peers and broader cultural influences (Király et al., 2023). Finally, in relation to specific structural characteristics of the game itself, disordered gaming risk factors include rewarding and reinforcing gaming experiences through operant conditioning processes, online game delivery, monetization aspects (e.g., buying/selling game winning equipment using offline currencies), and distinct game genres (e.g., Massively Multiplayer Online Role-Playing Games; MMORPGs; involving character development, socialization, competition and achievement elements; Király et al., 2023).
It should be noted that although higher gaming time has been related to higher disordered gaming risk, scholars have contended that it may not necessarily indicate disordered gaming, unless it compromises functionality in the gamer's everyday life (e.g., employment, education, and family life; Billieux, Flayelle, Rumpf, & Stein, 2019; Griffiths, 2010). Consequently, it is emphasized that high gaming involvement should be distinguished from disordered gaming (Billieux et al., 2019; Griffiths, 2010). Such literature has led to further calls for research examining the potentially harmful consequences of excessive gaming, as well as better identifying risk factors for developing problematic gaming patterns (Király, Potenza, & Demetrovics, 2022).
Disordered gaming
The World Health Organization (WHO) officially included gaming disorder (GD) in the 11th revision of the International Classification of Diseases (ICD-11; WHO, 2019). The ICD-11 defines GD as a pattern of gaming behaviour characterized by impaired control over gaming, increasing priority given to gaming over other activities to the extent that it takes precedence in daily life, and continuation/escalation of gaming despite the occurrence of negative consequences. The ICD-11 further states that a diagnosis of GD must have a significant impairment to an individual's personal, family, social, educational, occupational and/or other important areas of functioning (typically evident over a period of at least 12 months). Given the increased recognition of disordered gaming as a legitimate psychiatric condition, research into more specific risk factors and potential influencers of addictive gaming has greatly increased (Bäcklund, Elbe, Gavelin, Sörman, & Ljungberg, 2022; Liao, Chen, Huang, & Shen, 2022).
The WHO's (2019) diagnostic classification of GD followed the inclusion of the provisional diagnosis of internet gaming disorder (IGD) in the fifth edition of the Diagnostic and Statistical Manual for Mental Disorders (DSM-5; American Psychiatric Association, 2013). According to the DSM-5 (2013), and similar to WHO (2019) the criteria for diagnosing IGD includes: preoccupation with gaming, withdrawal symptoms when gaming is not possible, tolerance (i.e., needing to spend increasing amounts of time gaming), unsuccessful attempts to control or reduce gaming, loss of interest in other activities, continued excessive gaming despite negative consequences, and significant impairment in personal, social, educational, or occupational areas of functioning (with at least five of these criteria being met for more than a year to be considered as having a gaming disorder).
In the present study, the ICD-11 criteria for GD (WHO, 2019) were employed for three compelling reasons: (i) it is the only official (and not provisional) disordered gaming diagnosis currently employed worldwide; (ii) it has been supported that the ICD-11 diagnostic framework emphasizes more serious/pivotal (and a succinct number of) GD symptoms, without compromising diagnostic validity (Jo et al., 2019); and (iii) it provides consistency and comparability in relation to empirical evidence internationally (Pontes & Griffiths, 2019).
User-avatar bond
A number of scholars in the gaming studies field have reiterated that greater emphasis should be given to game-related features. This includes the user-avatar bond (UAB), as a potential GD risk factor in role-playing games (RPGs; Green, Delfabbro, & King, 2021; Lemenager, Neissner, Sabo, Mann, & Kiefer, 2020). RPGs have been consistently demonstrated to be a genre of videogames that have a higher risk of GD among individuals (Stavropoulos, Gomez, Mueller, Yucel, & Griffiths, 2020; Stavropoulos, Pontes, Gomez, Schivinski, & Griffiths, 2020; Szolin, Kuss, Nuyens, & Griffiths, 2022). An avatar is a visual in-game representation of the player, with the term originating from the Sanskrit word ‘avatāra’, referring to the embodiment of a deity in a human form (Lochtefeld et al., 2002; Szolin et al., 2022).
Within the gaming context, the avatar facilitates a process whereby the gamer may, to an extent, experience embodiment with their gaming persona/figure, while they are able to portray themselves in ways that align more with their desired self-expressions (Šporčić & Glavak-Tkalić, 2018; Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020). Consequently, a complex psychological attachment is facilitated between gamers and their avatars. This increases game engagement and can also influence some gamers' online and offline behaviours through subconscious processes (e.g., altered perceptions, automatic thoughts, and non-deliberate actions corresponding with their avatar features; Burleigh, Stavropoulos, Liew, Adams, & Griffiths, 2018; Liew, Stavropoulos, Adams, Burleigh, & Griffiths, 2018; Ortiz de Gortari, Pontes, & Griffiths, 2015; Ratan, Beyea, Li, & Graciano, 2020). Considering the UAB's particular strength/intensity, empirical research indicates that factors such as age, and the duration of engagement with the game world, may play a critical role in the how an individual connects with their avatar (e.g., younger gamers, with lengthier game involvement, could be more UAB receptive/susceptible, due to more dynamic/fluid personality features and time/emotional game investment; Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020; Stavropoulos, Ratan, & Lee, 2022; Rehbein, 2016).
Moreover, Blinka et al. (2008) noted that the UAB encompasses critical aspects and subdimensions. These entail identification (e.g., the gamer becomes more like their avatar, and they feel the same or alike), immersion (e.g., the avatar's needs in the world of the game [such as participating in a competition/task] are experienced as offline needs by the gamer, and can even be prioritised to their needs outside of the game [such as sleeping and/or eating] in the case of disordered gaming), and compensation/idealization (e.g., the avatar is who/how the gamer would like to have been in their offline life, but they may not be in a position to; the avatar may express an individual's ideal self).
Additionally, it has been argued that the need of some gamers, who might be experiencing low-self-esteem and/or may be dissatisfied by their offline self, could lead them to escape their discomfort through their idealized avatars within the game world (Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020; Stavropoulos, Ratan et al., 2022). Such avatar-mediated mood modification tendencies may cause some gamers to immerse/over-engage with (and emotionally depend on) their in-game character, fuelling their GD risk (Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020; Stavropoulos, Ratan et al., 2022). These findings are reinforced by other notable studies (e.g., those examining wishful avatar identification; Burleigh et al., 2018; Green et al., 2021; Liew et al., 2018; Yee, Bailenson, & Ducheneaut, 2009).
It has also been proposed that the UAB could operate as a form of ‘digital phenotype’, meaning a digital/gamified footprint of an individual's mental health, that, if analysed, can be translated into information not only concerning the gamer's risk of GD, but also for other psychopathological conditions (e.g., depression, anxiety [Loi, 2019; Stavropoulos et al., 2021; Zarate, Stavropoulos, Ball, de Sena Collier, & Jacobson, 2022]). Despite the consistent associations between GD and the UAB in the extant literature, the translation of the UAB into GD risk has never to date, to the best of the authors' knowledge, been investigated (Burleigh et al., 2018; Liew et al., 2018; Ortiz de Gortari et al., 2015; Ratan et al., 2020).
The present study
Analytical advancements in the field of machine learning (ML) can support artificial intelligence (AI) applications, which allow the automatic prediction/translation of one form of information/data into another (e.g., a gamer's UAB into GD risk; Horton & Kleinman, 2015; Kuhn & Wickham, 2020). To achieve such predictions, ML/AI procedures require training on related data, where predictors (e.g., UAB sub-dimensions) and outcomes (e.g., GD risk) are known, such that they can learn how to interpret/use the first variable to identify the latter (in the form of supervised algorithms; Horton & Kleinman, 2015; Kuhn & Silge, 2022; Kuhn & Wickham, 2020). After this stage is completed, a new set of data is examined by the trained AI/ML, where the accuracy of its predictions is validated (i.e., while in the first stage of the process AI learns to detect GD risk based on the UAB, in the second stage it makes predictions to demonstrate their learning quality; Kuhn & Silge, 2022).
Indeed, recent research examples have aimed to use ML/AI to diagnose GD via Resting Brain State, MRI, PET and EEG data with encouraging findings (Han et al., 2021; Song et al., 2021). Taking these into consideration, the present study innovatively examined a recently collected longitudinal dataset using AI/ML classifiers, aiming to translate gamers reported UAB identification, immersion, and compensation/idealization into their present and prospective (i.e., six months later) GD risk, while also taking into consideration their age and years of videogame engagement. In particular the choice of a longitudinal design was chosen over cross-sectional data collection because it allows the examination of the direction of causality between the behaviours examined, while additionally enabling the potential translation of the user-avatar bond into prospective GD risk (Zarate, Dorman, Prokofieva, Morda, & Stavropoulos, 2023). Consequently, the following research questions (RQs) were formulated:
RQ1: How can, if at all, ML/AI applications be trained to identify whether a gamer presents with current GD risk, based on their UAB reported identification, immersion, compensation, age, and length of gaming involvement (i.e., concurrent GD phenotype)?
RQ2: How can, if at all, ML/AI applications be trained to identify whether a gamer presents with future GD risk (i.e., six months later), based on their UAB reported identification, immersion, compensation, age, and length of gaming involvement (i.e., prospective GD phenotype)?
Methods
Participants
A sample of 627 gamers were initially recruited. Of these, seven were excluded as preview-only responses, 19 as spam, one as a bot, 12 due to lack of consent, eight for failing validity questions (e.g., claimed they played non-existing games; e.g., Risk of Phantom), and 15 for insufficient responses. Therefore, the final sample comprised 565 role-playing-gamers (Mage = 29.3 years SD = 10.6, Minage = 12, Maxage = 68; Malescisgender = 283, 50.1%), who were longitudinally assessed in the community, six months apart (two time-points, T1 and T2). With regards to demographics at T1, 271 (55.3%) reported being full-time employed, 176 (36%) had an undergraduate degree, 359 (73.6%) stated heterosexual orientation, 410 (72.5%) identified as of Australian/English ancestry, 142 (25.1%) resided with their family of origin, and 148 (30.2%) were single.
With regards to gaming patterns at T1, they reported having been a gamer for on average for 5.62 years (Min=<1 year, Max = 30 years; SD = 4.49), for an average of 2.23 h daily during weekdays (Min=<1 h, Max = 15 h; SD = 1.82) and 3.39 h during the weekend (Min=<1 h, Max = 18; SD = 2.40). Considering social media use patterns at T1, they reported having been a social media user for an average of 7.06 years (Min=<1 year, Max=17; SD = 7.06), spending an average time of 2.55 h during weekdays (Min=<1 h, Max = 15 h; SD = 2.16), and 3.01 h during the weekend (Min=<1 h, Max = 16 h; SD = 2.48) with 145 (26%) reporting Facebook as their preferred platform. The maximum random sampling error for a sample of 565 at the 95% confidence interval (z = 1.96) equalled ± 4.12% satisfying Hill's (1998) recommendations. Missing values of the analysed variables at T1 ranged between 3 (0.5% not stating their age) to 16 (2.83% not answering Item 9 on the User-Avatar Bond Scale), and were missing completely at random in the broader dataset (MCARtest = 38.4, p = 0.14 [9 missing patterns]; Little (1988).
Attrition between waves was 276 participants (48.8%). Therefore, retention/attrition were studied in relation to participants' sociodemographic information considering statistical significance and effect size (Cohen's d, very small∼0.01, small∼0.20, medium∼0.50, large, 0.80, very large∼1.20; Sawilowsky, 2009); Cramer's V > 0.25 = very strong, >0.15 = strong, >0.10 = moderate, >0.05 = weak, >0 no or very weak). Low to moderate effect-sizes were found regarding the associations between attrition and gender (χ2 = 4.26, df = 6, p = 0.642, Cramer's V = 0.087), sexual orientation (χ2 = 7.75, df = 4, p = 0.101, Cramer's V = 0.126), ancestry (χ2 = 8.94, df = 4, p = 0.063, Cramer's V = 0.126), romantic relationship engagement (χ2 = 3.76, df = 4, p = 0.440, Cramer's V = 0.088), educational status (χ2 = 11.2, df = 7, p = 0.129, Cramer's V = 0.152), employment status (χ2 = 7.58, df = 6, p = 0.271, Cramer's V = 0.124), number of years spent gaming (tWelch's = 3.509, df = 526, p < 0.001, Cohen's d = 0.296), average daily gaming time during the week (tStudent = 0.873, df = 555, p = 0.383, Cohen's d = −0.0741), average daily gaming time during the weekend (tStudent = 0.159, df = 553, p = 0.874, Cohen's d = 0.0135), number of years spent using social media (tStudent = 2.501, df = 556, p = 0.013, Cohen's d = 0.2118), average daily social media use time during the week (tStudent = −2.313, df = 543, p = 0.021, Cohen's d = −0.1983), average daily social media use time during the weekend (tWelch = −2.447, df = 501, p = 0.015, Cohen's d = −0.2111), and age (tStudent = 4.967, df = 560, p < 0.001, Cohen's d = 0.4192). Tables 1 and 2 provide detailed description of the sample at T1.
Participant's age, gaming/social media use years and daily week and weekend consumed time at T1
| Age | Number of years spent gaming | Mean daily gaming time in the week | Mean daily gaming time at the weekend | Number of years spent using social media | Mean daily social media use time in the week | Mean daily social media use time at the weekend | |
| N | 562 | 556 | 557 | 555 | 558 | 545 | 543 |
| Mean | 29.3 | 5.62 | 2.23 | 3.39 | 7.06 | 2.55 | 3.01 |
| SD | 10.6 | 4.49 | 1.82 | 2.40 | 4.41 | 2.16 | 2.48 |
| Min | 12.0 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 | 0.00 |
| Max | 68.0 | 30.0 | 15.0 | 18.0 | 17.0 | 15.0 | 16.0 |
Participants' sociodemographic, gaming and social media use information at T1
| N | Total N | Proportion | p | ||
| Gender | Man (cisgender) | 283 | 565 | 0.501 | 1.000 |
| Woman (cisgender) | 259 | 565 | 0.458 | 0.053 | |
| Man (transgender) | 4 | 565 | 0.007 | <0.001 | |
| Woman (transgender) | 1 | 565 | 0.002 | <0.001 | |
| Nonbinary | 12 | 565 | 0.021 | <0.001 | |
| Not Listed | 3 | 565 | 0.005 | <0.001 | |
| Prefer not to say | 3 | 565 | 0.005 | <0.001 | |
| Sexual Orientation | Heterosexual-Straight | 359 | 488 | 0.736 | <0.001 |
| Homosexual | 36 | 488 | 0.074 | <0.001 | |
| Bisexual | 75 | 488 | 0.154 | <0.001 | |
| Asexual | 5 | 488 | 0.010 | <0.001 | |
| Other | 13 | 488 | 0.027 | <0.001 | |
| Ancestry | Aus./Engl. | 412 | 565 | 0.552 | 0.015 |
| Chinese | 20 | 565 | 0.035 | <0.001 | |
| German | 7 | 565 | 0.012 | <0.001 | |
| Indian | 10 | 565 | 0.018 | <0.001 | |
| Other | 118 | 565 | 0.209 | <0.001 | |
| Occupational Status | Full-time employed | 271 | 490 | 0.553 | 0.021 |
| Part-time employed | 77 | 490 | 0.157 | <0.001 | |
| Student | 64 | 490 | 0.131 | <0.001 | |
| Trainee | 2 | 490 | 0.004 | <0.001 | |
| Not currently working | 32 | 490 | 0.065 | <0.001 | |
| On temporary leave (education leave, public service leave, training, maternity leave) | 5 | 490 | 0.010 | <0.001 | |
| Other | 39 | 490 | 0.080 | <0.001 | |
| Educational Status | Professional degree (i.e., MD, JD, etc. completed) | 10 | 489 | 0.020 | <0.001 |
| PhD degree (completed) | 17 | 489 | 0.035 | <0.001 | |
| Postgraduate studies (MSc completed) | 67 | 489 | 0.137 | <0.001 | |
| Undergraduate university course (completed) | 176 | 489 | 0.360 | <0.001 | |
| Intermediate between secondary level and university (e.g., technical training) | 97 | 489 | 0.198 | <0.001 | |
| Senior secondary school (Years 11–12) | 101 | 489 | 0.207 | <0.001 | |
| Secondary school (Years 7–10) | 9 | 489 | 0.018 | <0.001 | |
| Other | 12 | 489 | 0.025 | <0.001 | |
| Livingwith_w1 | Family of origin (two parents/partners, only child) | 34 | 564 | 0.060 | <0.001 |
| Family of origin (two parents/partners and siblings) | 108 | 564 | 0.191 | <0.001 | |
| Mother (only child, parent divorced-separated-widowed) | 19 | 564 | 0.034 | <0.001 | |
| Mother and sibling(s) (parent divorced-separated-widowed) | 17 | 564 | 0.030 | <0.001 | |
| Father (only child, parent divorced-separated-widowed) | 6 | 564 | 0.011 | <0.001 | |
| Father and sibling(s) (parent divorced-separated-widowed) | 5 | 564 | 0.009 | <0.001 | |
| With partner | 149 | 564 | 0.264 | <0.001 | |
| Alone | 61 | 564 | 0.108 | <0.001 | |
| With friend(s) | 28 | 564 | 0.050 | <0.001 | |
| Temporary accommodation | 4 | 564 | 0.007 | <0.001 | |
| Other | 18 | 564 | 0.032 | <0.001 | |
| With partner and children | 115 | 564 | 0.204 | <0.001 | |
| Relationship status | Single | 148 | 490 | 0.302 | <0.001 |
| In a romantic relationship (A romantic relationship is defined as a romantic commitment of particular intensity between two individuals of the same or the opposite sex (When you like a guy [girl] and he [she] likes you back). | 157 | 490 | 0.320 | <0.001 | |
| Engaged | 24 | 490 | 0.049 | <0.001 | |
| Married | 145 | 490 | 0.296 | <0.001 | |
| De facto | 16 | 490 | 0.033 | <0.001 | |
| Partner games together | Yes | 99 | 344 | 0.288 | <0.001 |
| No | 245 | 344 | 0.712 | <0.001 | |
| Partner uses social media together | Yes | 227 | 340 | 0.677 | <0.001 |
| No | 113 | 340 | 0.333 | <0.001 | |
| Social media users | Yes | 550 | 565 | 0.973 | <0.001 |
| No | 15 | 565 | 0.027 | <0.001 | |
| Facebook users | No | 168 | 565 | 0.297 | <0.001 |
| 397 | 565 | 0.703 | <0.001 | ||
| Twitter users | No | 320 | 565 | 0.566 | 0.002 |
| 245 | 565 | 0.434 | 0.002 | ||
| Instagram users | No | 195 | 565 | 0.345 | <0.001 |
| 370 | 565 | 0.655 | <0.001 | ||
| Pinterest users | No | 469 | 565 | 0.830 | <0.001 |
| 96 | 565 | 0.170 | <0.001 | ||
| TikTok users | No | 368 | 565 | 0.651 | <0.001 |
| Tik Tok | 197 | 565 | 0.349 | <0.001 | |
| Most preferred social media | 145 | 557 | 0.260 | <0.001 | |
| 66 | 557 | 0.118 | <0.001 | ||
| 135 | 557 | 0.242 | <0.001 | ||
| 5 | 557 | 0.009 | <0.001 | ||
| Tik Tok | 99 | 557 | 0.178 | <0.001 | |
| Other, please define which | 107 | 557 | 0.192 | <0.001 | |
| Gaming with best friend | No | 336 | 565 | 0.595 | <0.001 |
| Yes | 229 | 565 | 0.405 | <0.001 | |
| Using social media with best friend | No | 189 | 565 | 0.335 | <0.001 |
| Yes | 376 | 565 | 0.665 | <0.001 | |
| Gaming with other friends | No | 312 | 565 | 0.552 | 0.015 |
| Yes | 253 | 565 | 0.448 | 0.015 | |
| Using social media with offline friends | No | 154 | 565 | 0.273 | <0.001 |
| Yes | 411 | 565 | 0.727 | <0.001 | |
| Gaming with family members | No | 406 | 565 | 0.719 | <0.001 |
| Yes | 159 | 565 | 0.281 | <0.001 | |
| Using social media with family members | Yes | 472 | 564 | 0.837 | <0.001 |
| No | 92 | 564 | 0.163 | <0.001 |
Note. Hₐ is proportion ≠ 0.5.
Measures
In addition to data concerning demographics, gaming use, and social media use, the following data were collected.
Gaming Disorder Test (GDT-4; Pontes et al., 2021)
The GDT-4 assesses the diagnostic features/severity of disordered gaming with a design directly modelled on the WHO (2019) conceptualisation. There are four items each addressing a particular symptom (e.g., “I have had difficulties controlling my gaming activity”) using a five-point Likert-type scale from 1 (Never) to 5 (Very often). Total scores range from 4 to 20 with higher scores indicating greater GD severity. Participants at GD risk were classified those with more than 3/5 (Often) in ¾ of the GDT-4 items (Pontes et al., 2021). The internal consistency coefficients were sufficient across both study waves (Cronbach's α GDT wave 1 = 0.808, McDonald's ω GDT wave 1 = 0.812, Cronbach's α GDT wave 2 = 0.854, McDonald's ω GDT wave 2 = 0.862).
User-Avatar-Bond Questionnaire (UAB-Q; Blinka, 2008)
The UAB-Q was used to assess different gamer-avatar bond dimensions. The 12 UAB-Q items are answered on a 5-point Likert scale from 1 (strongly disagree) to 5 (strongly agree) comprising three factors: identification (four items; “Both me and my character are the same”), immersion (five items: “Sometimes I think just about my character while not gaming”), and compensation (three items: “I would rather be like my character”). The scores range from 12 to 60 with higher scores indicating stronger UAB experiences both overall and on the respective subscales. The internal consistency coefficients were sufficient across both study waves (Cronbach's α UAB-Q wave 1 = 0.804; McDonald's ω UAB-Q wave 1 = 0.813, Cronbach's α UAB-Q wave 2 = 0.849; McDonald's ω UAB-Q wave 2 = 0.867, Cronbach's α Ident. wave 1 = 0.701; McDonald's ω Ident. wave 1 = 0.729, Cronbach's α Ident. wave 2 = 0.770; McDonald's ω Ident. wave 2 = 0.789 Cronbach's α Immers. wave 1 = 0.717; McDonald's ω Immers. wave 1 = 0.727, Cronbach's α Immers. wave 2 = 0.764; McDonald's ω Immers. wave 2 = 0.775, Cronbach's α Comp. wave 1 = 0.604; McDonald's ω Comp. wave 1 = 0.656, Cronbach's α Comp. wave 2 = 0.660; McDonald's ω Comp. wave 2 = 0.709).
Procedure
Approvals were granted by the Victorian University Human Research Ethics Committee [HRE21-044], the Department of Education and Training of The Victorian State Government, Australia [2022_004542], and the Melbourne Archdiocese of Catholic Schools [1179]. Participants were sampled from the community (e.g., RMIT, Victoria, Melbourne and Deakin Universities), Victorian public and catholic schools, Australian gamers' groups (e.g., Aus Gaymers Network), venues (e.g., Fortress Melbourne), and online forums (e.g., AusGamers), as well as advertising via YouTube videos. Gamers older than 12 years were eligible to voluntarily/anonymously participate and were provided with the plain language information statement describing the study aims, risks and their participation rights (e.g., withdrawal without any penalties and/or repercussions at any point) and provided their informed consent. For adolescents (i.e., 12–18 years), these were firstly addressed by their responsible parent/guardian and secondly by the adolescents themselves. Data collection involved three data-streams, paired via a non-identifiable code, unique for each participant: (i) a battery of demographic, internet/gaming/social media use questions, and psychometric questionnaires/scales available via an online Qualtrics link; (ii) wearing an actigraphy tracker (Fitbit) for seven days to monitor physical activity/sleep (e.g., daily steps and sleep duration), that was electronically paired with the other data-streams via a unique code (i.e., records were automatically collected via the Fitbit portal based on the participant's code and those not owning a Fitbit were provided with a device during a mutually arranged/agreed meeting with the research team) and; (iii) carrying a mobile monitoring application, called Aware Light (Van Berkel, D’Alfonso, Susanto, Ferreira, & Kostakos, 2023) recording screen on/off time, number and length of calls (i.e., duration) and texts (i.e., length in characters) for seven days (i.e., Light Aware data were also matched with the other data-streams through the unique participant code). The procedure was repeated four times, once every six months, with the present study being based on the first two completed collection waves (for detailed information see Supplementary Materials 1.
Data analysis
To address RQ1 (i.e., concurrent GD digital phenotype; identifying present GD risk based on an individual's age, number of years spent gaming, and reported avatar identification, immersion and compensation/idealization) machine learning (ML) procedures using the Tidymodels package were conducted in R-Studio (Horton & Kleinman, 2015; Kuhn & Wickham, 2020). Firstly, data were balanced considering Yes/No GD risk cases to improve learning/ML-prediction using the synthetic minority oversampling technique (SMOTE; DMwR package; Torgo & Torgo, 2013). This algorithm introduces additional cases of the minority group by taking into consideration a potential number (k) of their nearest neighbours based on Euclidean distance (Chawla, Bowyer, Hall, & Kegelmeyer, 2002).
Practically, k-NN operates by identifying the distance between a suggested case and all other data cases considered. Firstly, it chooses a number (k) of cases nearest to the point of interest. Then, it attaches the most frequent class to that point (e.g., Yes/No GD risk; Chawla et al., 2002). Secondly, data were split into 4/5 training and 1/5 testing, stratifying Yes/No GD risk proportions to be equal across the splits, while adopting a conservative bell-shaped Bayesian prior distribution. It should be noted that when adopting a Bayesian perspective, a potential distribution/variability is required for every model parameter before proceeding to data analysis. The range of these values was carefully/modestly/conservatively suggested here to follow a Cauchy shape (i.e., t-shape with seven degrees of freedom; Muth, Oravecz, & Gabry, 2018).
Finalized training and testing datasets were similar regarding Yes/No GD risk proportions (χ2 = 0, df = 1, p = 1). For cross-validation and ML hyperparameters' tuning, training data were additionally divided 10 times (i.e., folds) and training data bootstrapped versions were also created. Thirdly, the ML recipe (i.e., predictive equation) was introduced, such that: (i) the binary Yes/No GD risk at T1 was the outcome and age, number of years spent gaming, avatar-identification, avatar-immersion and avatar-identification were the independent predictors; (ii) a minimum ratio of 50% GD risk cases was maintained across all samples tested, including the cross-validation and bootstrapped training data versions; and (iii) zero variance, strongly sparse/skewed, and potentially highly intercorrelated predictors were excluded, to solidify findings. It should also be highlighted that the latter did not effectively exclude any predictor in the current recipe.
Predictors were also scaled and centred prior to the recipe to accommodate classification (i.e., 0 = mean and 1 = Standard Deviation [SD]; Kuhn & Wickham, 2020). Fourthly, a series of supervised ML models (i.e., models where the outcome is known in the training step/stage) recommended for binary classification (see Table 3) were introduced, alongside the null model (i.e., no ML prediction) in their tuned and their untuned versions, where hyperparameters were appropriately adjusted (Kuhn & Wickham, 2020). A hyper-parameter constitutes an ML parameter, the value of which needs to have been specified prior to the learning ML being trained, in contrast to simple parameters which are “learned” during the training of the model. Therefore, hyperparameters pose external model configurations (i.e., not based on the data) employed for the estimation of model parameters. Fine-tuned hyperparameters increase the capacity of a learning model to perform with higher accuracy, and are achieved through a “grid” process in tidymodels (Kuhn & Wickham, 2020).
ML models trained, tuned and tested
| Type | Operation | Hyperparameters tuned | R-package/engine employed |
| Least Absolute Shrinkage Selection Operator (LASSO) | LASSO constitutes a regression analysis based, supervised ML classifier, that applies variable selection and regularization to increase prediction accuracy. It achieves that via reducing noises and selecting certain features to regularize the model. From a calculation perspective lasso considers the magnitude rate of the coefficient, as a penalty to the loss function. Therefore, the loss function is amended to reduce model complexity via restraining the sum of predictors' coefficients [Loss function = OLS + A (penalty) X summation (addition of s size[s] of coefficients)]. | penalty = To perform regularization (i.e., L1), LASSO considers/adds a penalty to the size of regression coefficients (i.e., predictor effects), aiming to minimize them. The optimum penalty value is obtained via the tuning process. | glmnet |
| K Nearest Neighbours (k-NN) | Th k-NN algorithm entails a supervised, non-parametric classification/prediction, that relies on estimating proximity/relevance/distance of one case with “k” others, as per their Euclidean distance. Alternatively, k-NN classifies/categorizes a case taking into consideration its neighbouring cases (i.e., similarity of a case with previously identified cases). | neighbors = The number (k) of neighbouring points to be considered in order to optimize the learning/prediction performance of the algorithm, as defined via the tuning process. | knn |
| Support Vector Machine Kernel (SVM-K) | Kernel ML is based on pattern examination/analysis and is mostly known via its popular support-vector machine (SVM) version. The kernel function refers to a mathematic procedure, which enables SVM to pursue deep learning via conducting bidimensional classifications of uni-dimensional data through the projection of a lower-dimension to a higher one. Subsequently, a kernelized SVM employs a linear computation to address non-linear/classification problems. | cost = In SVM, cost resembles/postulates the logistic function via a piecewise linear. In practice, the cost hyperparameter programs/guides the algorithm's optimization regarding the rate/size of misclassification allowed in the training sample. Higher cost values indicate tighter margins and the opposite. degree = The degree hyperparameter dictates the flexibility/boundaries of prediction(s), such that higher values allow higher flexibility. scale_factor = The scaling hyper-parameter of categorical/classification kernel(s) reflects the optimum normalization patterns/process (i.e., kernel width) required to avoid any data modification. | kernlab |
| X Gradient Boosting (XGB) | XGBoost is recommended for structured/tabular data. It implements gradient boosted decision trees to optimize prediction. XGBoost does so via providing a parallel tree boosting that integrates/considers weak prediction/learner models/decision trees. However, and in contrast to random forest bagging of generated trees, XG-Boosting operates in a sequential manner, with any subsequent tree being influenced by the previous/last tree outcome. | mtry = The number of independent variables to be randomly assessed at each decision tree split. min_n = An integer/value/number for the least data points in a node (i.e., tree branch) that enables further split. tree_depth = The value defining the highest tree depth (i.e., subsequent splits) suggested to optimize prediction. Learn rate (i.e., shrinkage) = The value/rate required for the boosting adaptation to occur over successive iterations. loss_reduction = The reduction rate of the loss function suggested to progress with tree splits. sample_size = The amount/proportion of data required to be utilized in the algorithm's fitting process over each iteration. | xgboost |
| Random Forests | Random Forest is a flexible and broadly employed supervised, ensemble (i.e., composite) ML model, that integrates/considers the results of numerous decision trees (i.e., bagging), while being trained/learning to address a prediction/classification task. Practically, random forests conduct a meta-estimation that averages/considers the outcomes of multiple decision tree classifiers, implemented on different data sub-samples, to improve accuracy and deter over-fitting. | mtry = The number of independent variables to be randomly assessed at each decision tree split. min_n = An integer/value/number for the least data points in a node (i.e., tree branch) that enables further split. | ranger |
| Naïve Bayes | Naïve Bayes operates as a probabilistic, supervised, ML classifier, which functions generatively. This suggests that it aims to model the data class distribution, while assuming conditional independence probability (i.e., data characteristics/measures are independent) to predict the way a specific class would generate input data. | smoothness = This refers to the Kernel component Smoothness, which defines the density value required for the algorithm to converge quicker, to the real density of random numeric predictors. Laplace = Laplace transformation/smoothing refers to a technique/strategy/method that addresses the problem/risk of zero probability in the algorithm. | naivebayes |
| Logistic Regression | Logistic Regression is also considered a supervised ML classifier that employs a logistic function to predict/model binary/dichotomous dependent outcomes. | penalty = In logistic regression, as with LASSO, the regularization penalty hyperparameter aims to address generalization error and therefore reduce overfitting risks. As such, it enhances the probability of simpler concluded models. mixture = A regularization parameter value ranging between 0 and 1 to enhance model accuracy [mixture 1 corresponds with LASSO; 0 with ridge regression and in the interim with elastic modelling in between LASSO and ridge]. | glm |
Note: Glmnet is derived from “Friedman et al. (2010). Package ‘glmnet’. CRAN R Repositary.”; Ranger is derived from “Wright and Ziegler (2017). Package ‘ranger’.” Kernlab is derived from “Karatzoglou, Smola, and Hornik (2023). Package ‘kernlab’. CRAN R Project”. Xgboost is derived from “Chen et al. (2023). Package ‘xgboost’. R version, 90, 1–66.”. All other engines ae derived from “ Kuhn, M., & Silge, J. (2022). Tidy Modeling with R. " O'Reilly Media, Inc.".
Fifthly, model and recipes were combined to create different workflows, which were: (i) trained in the default versions on the training data; (ii) tuned considering their hyperparameters via the bootstrapped versions the training data, and; (iii) tested across their default/tuned versions on the testing data. To address RQ2 (i.e., prospective GD digital phenotype in six months), the same procedure was repeated with GD T2 being the outcome/dependent variable. Findings were compared based on their confusion matrices, accuracy, precision, the area under the curve, recall, and f-measures (see yardstick r package; Kuhn, Vaughan, & Vaughan, 2020).1
Preceding the analysis, estimation for the sample size was also considered from the overfitting perspective of the developed models. In machine learning, overfitting refers to the modelling error occurring, when a function used in a model is too closely aligned to a limited set of data points. This indicates insufficiency of the data and results in a model generating accurate predictions for training data but not for new/testing data (Chawla et al., 2002). The present study addressed overfitting by considering the imbalance in the dataset, and using the Synthetic Minority Over-Sampling TEchnique (SMOTE; Chawla et al., 2002; Torgo & Torgo, 2013), applying early stopping in Random Forest application, as well as use of regularization technique LASSO. Further measures included cross-validation and hyperparameter tuning of the developed models (see Table 3).
Ethics
All procedures performed in the study involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards. The paper does not contain any studies with animals performed by any of the authors. Informed consent was obtained from all individual participants included in the study.
Results
Before addressing RQ1 and RQ2, Yes/No GD riskwave_1 participants were identified with Nno_GD_Risk = 430 (80.22%) and NYes_GD_Risk = 106 (19.78%). For RQ1, to accommodate ML learning, oversampling of the minority class was conducted using k-NN SMOTE (Chawla et al., 2002; Torgo & Torgo, 2013) resulting in a balanced dataset (i.e., NYes_GD_Risk = 530; 50%). Data were then split into 80% training and 20% testing and the proportions of Yes/No GD risk were compared across the two parts showing non-significant differences (χ2 = 0, df = 1, p = 1; Cramer's V = 0.00; 50% Yes GD risk across both training and testing). The prediction recipe was introduced, scaling of predictors was conducted, descriptives of the training, testing and whole dataset were estimated (see bake recipe section; Supplementary Material 2), while 10 sub-divisions and bootstrapped versions of the training data were produced for cross-validation and hyperparameter tuning (see folds & train_boot section, Supplementary Material 2). Models and workflows of the Null, LASSO, SVM-Kernel, Random Forests, Naïve Bayes, and Logistic Regression (see Table 3) in their default hyperparameter versions (i.e., untuned) were then introduced, trained on the training data, and tested on the testing data. Table 4 summarizes their performance suggesting that, while all classifiers performed/learned acceptably and better than the null model, except LASSO, Random Forests learning outperformed other classifiers with excellent indicators across all criteria (see Fig. 1). Immersion was the most significant predictor for Random Forests (i.e., >25 points) with all other predictors exceeding 10 points (see VIP section, Supplementary Material 2).
Null model and untuned algorithms performance on testing data (GD Wave 1)
| Null model | Random forests | Logistic regression | LASSO | Naïve Bayes | SVM Kernel | |
| ROC_AUC | 0.5 | 0.975 | 0.701 | 0.5 | 0.788 | 0.741 |
| PPV | 0.5 | 0.942 | 0.641 | 0.5 | 0.8 | 0.664 |
| F_meas | 0.667 | 0.933 | 0.673 | 0.667 | 0.712 | 0.685 |
| Recall | 1 | 0.925 | 0.708 | 1 | 0.642 | 0.708 |
| Accuracy | 0.5 | 0.934 | 0.656 | 0.5 | 0.741 | 0.675 |
Untuned classifiers performance across the criteria (GD Wave 1)
Citation: Journal of Behavioral Addictions 12, 4; 10.1556/2006.2023.00062
To optimize learning and modelling capacity, the versions of LASSO, SVM-Kernel, Random Forests, Naïve Bayes and Logistic Regression, as well as XGB and k-NN were later tuned (see Table 3 regarding their respective hyperparameters' functions), trained on the training data and tested on the testing data. Table 5 summarizes the tuned hyperparameters' values per classifier and Table 6 their performance. Results suggest that, while all classifiers performed/learned acceptably and better than the null model, including LASSO, Random Forests learning outperformed other classifiers comparatively with excellent indicators across all criteria, followed by XGB, SVM-Kernel, and k NN (see Fig. 2).
Hyperparameter tuning summary across classifiers (GD Wave 1)
| Type | Hyperparameters tuned | Tuning results |
| Least Absolute Shrinkage Selection Operator (LASSO) | penalty | 0.00139 |
| K Nearest Neighbours (k-NN) | neighbors | 10 |
| Support Vector Machine Kernel (SVM-K) | cost | 32 |
| scale_factor | 1 | |
| X Gradient Boosting (XGB) | mtry | 1 |
| min_n | 6 | |
| tree_depth | 15 | |
| Learn rate (i.e., shrinkage) | 11 | |
| loss_reduction | 0.0425 | |
| sample_size | 0.171 | |
| Random Forests | mtry | 1 |
| min_n | 6 | |
| Naïve Bayes | smoothness | 0.5 |
| Laplace | 0 | |
| Logistic Regression | penalty | 0.00234 |
| mixture | 0.55 |
See Table 3 for detailed information regarding the classifiers applied.
Tuned algorithms performance on testing data (GD Wave 1)
| Null model | Random forests | Logistic regression | LASSO | Naïve Bayes | SVM Kernel | XGB | k-NN | |
| ROC_AUC | 0.5 | 0.981 | 0.704 | 0.704 | 0.811 | 0.96 | 0.955 | 0.939 |
| PPV | 0.5 | 0.951 | 0.647 | 0.647 | 0.755 | 0.959 | 0.873 | 0.966 |
| F_meas | 0.667 | 0.938 | 0.676 | 0.676 | 0.725 | 0.916 | 0.889 | 0.876 |
| Recall | 1 | 0.925 | 0.708 | 0.708 | 0.698 | 0.877 | 0.906 | 0.802 |
| Accuracy | 0.5 | 0.939 | 0.66 | 0.66 | 0.736 | 0.92 | 0.887 | 0.887 |
Tuned classifiers performance across the criteria (GD Wave 1)
Citation: Journal of Behavioral Addictions 12, 4; 10.1556/2006.2023.00062
The same process was repeated for RQ2 with Random Forests again outperforming other classifiers in both their tuned and untuned versions. Tables 7–9 summarize the performance of the untuned versions, the tuned hyperparameters' values, and the performance of the tuned classifiers respectively. Figures 3 and 4 visualize the performance of the tuned and untuned models (see Supplementary Material 3 and 4).
Null model and untuned algorithms performance on testing data (GD Wave 2)
| Null model | Random forests | Logistic regression | LASSO | Naïve Bayes | SVM Kernel | |
| ROC_AUC | 0.5 | 0.959 | 0.718 | 0.724 | 0.744 | 0.708 |
| PPV | 0.5 | 0.897 | 0.667 | 0.629 | 0.735 | 0.68 |
| F_meas | 0.667 | 0.897 | 0.643 | 0.65 | 0.673 | 0.63 |
| Recall | 1 | 0.897 | 0.621 | 0.672 | 0.621 | 0.586 |
| Accuracy | 0.5 | 0.897 | 0.655 | 0.638 | 0.698 | 0.655 |
Hyperparameter tuning summary across classifiers (GD Wave 2)
| Type | Hyperparameters tuned | Tuning results |
| Least Absolute Shrinkage Selection Operator (LASSO) | penalty | 0.00569 |
| K Nearest Neighbours (k-NN) | neighbors | 10 |
| Support Vector Machine Kernel (SVM-K) | cost | 32 |
| scale_factor | 1 | |
| X Gradient Boosting (XGB) | mtry | 1 |
| min_n | 3 | |
| tree_depth | 11 | |
| Learn rate (i.e., shrinkage) | 0.00268 | |
| loss_reduction | 0.495 | |
| sample_size | 0.336 | |
| Random Forests | mtry | 1 |
| min_n | 6 | |
| Naïve Bayes | smoothness | 0.5 |
| Laplace | 0 | |
| Logistic Regression | penalty | 0.0264 |
| mixture | 0.35 |
See Table 3 for detailed information regarding the classifiers applied.
Tuned algorithms performance on testing data (GD Wave 2)
| Null model | Random forests | Logistic regression | LASSO | Naïve Bayes | SVM Kernel | XGB | k-NN | |
| ROC_AUC | 0.5 | 0.959 | 0.72 | 0.721 | 0.773 | 0.95 | 0.85 | 0.904 |
| PPV | 0.5 | 0.883 | 0.673 | 0.685 | 0.792 | 0.981 | 0.8 | 0.947 |
| F_meas | 0.667 | 0.898 | 0.655 | 0.661 | 0.717 | 0.946 | 0.741 | 0.75 |
| Recall | 1 | 0.914 | 0.638 | 0.638 | 0.655 | 0.914 | 0.69 | 0.621 |
| Accuracy | 0.5 | 0.897 | 0.664 | 0.672 | 0.741 | 0.948 | 0.759 | 0.793 |
Untuned classifiers performance across the criteria (GD Wave 2)
Citation: Journal of Behavioral Addictions 12, 4; 10.1556/2006.2023.00062
Tuned classifiers performance across the criteria (GD Wave 2)
Citation: Journal of Behavioral Addictions 12, 4; 10.1556/2006.2023.00062
Discussion
The present longitudinal study employed a relatively large, normative sample of gamers to train AI/ML automated procedures to identify an individual's concurrent and prospective (i.e., six months later) GD risk, based on their age, number of years spent gaming, and reported avatar identification, immersion, and compensation/idealization. Five untuned (i.e., in their default versions) and seven tuned (i.e., ML/AI hyper-parameters/calculation features specifically adjusted to improve learning) recommended, and widely employed ML classifiers, were comparatively examined twice (i.e., current and prospective GD risk; Blinka, 2008; Kuhn & Silge, 2022).
The data were split into training and testing parts for the AIs to be trained and assessed respectively, while a prediction recipe was introduced. The models were trained, tuned, and tested, such that their capacity to learn whether an individual presents or not to be at GD risk at present and six months later, could be confirmed. Findings demonstrated that while all AI classifiers tested in the present study, were able to learn and performed better than the null model (i.e., random prediction), Random Forests had the strongest learning potential. Of the UAB aspects identified, immersion was the most important predictor of GD risk.
Gaming disorder and user-avatar bond
The present study's findings align with previous studies suggesting that stronger/higher UAB experiences are more likely to associate with excessive/disordered/problematic gaming, when/if there is a tendency for the individual to ‘escape from reality’, as a result of identity-related issues including poor self-concept, psychological vulnerability, and ‘wishful identification’ (i.e., compensation for negative self-perceptions; Green et al., 2021; Lemenager et al., 2020; Šporčić & Glavak-Tkalić, 2018; Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020; Van Looy, 2015). Moreover, scholars have supported that one of the most important indicators of GD is the process of transporting the players' psyche into the gaming environment, that is, the facilitation of a true “detachment from reality and the actual self” (Šporčić & Glavak-Tkalić, 2018, p. 8).
Relatedly, the player-avatar connection/interaction is maintained by identification and idealisation, and subsequently strengthened through both the immersive qualities of the game itself, and the ‘escape motives’ of players (Green et al., 2021; Lemenager et al., 2020; Šporčić & Glavak-Tkalić, 2018; Stavropoulos, Gomez et al., 2020; Stavropoulos, Pontes et al., 2020; Stavropoulos, Ratan et al., 2022). Therefore, the immersion factor, expressing the experience of the avatar's needs as offline needs of the gamer, can be seen as advancing UAB understanding, while sharpening the explanatory framework for players vulnerable to GD (Stavropoulos, Ratan et al., 2022). It is perhaps unsurprising that of all the UAB aspects considered within the present study, immersion was found to be the strongest predictor of GD risk. In other words, whether a gamer resembles their avatar (i.e., identification) or wishes to be like their avatar (i.e., compensation/idealization) appears to induce lower GD risk, compared to the extent that a gamer fuses with their avatar's needs, experiencing them as theirs (Ratan et al., 2020). The latter increases more their GD likelihood and presents an opportunity for AI to better learn to detect those at risk of GD.
Furthermore, the methods employed in the present study expand and advocate for the ML/AI translation of the UAB into GD risk, while considering the age of the gamer and the number of years they have spent gaming. Findings suggest that the UAB could operate as a diagnostic indicator of GD risk both at present and prospectively (six months later), when addressed using trained ML/AI procedures. This aligns with past literature recommending the careful decoding/interpretation of the health/mental health information likely embedded in the UAB (Stavropoulos et al., 2021). Indeed, the avatar's customization by the gamer, allows conscious and less conscious projections of the gamer's wishes and characteristics into the avatar, such that avatars and the way the gamers bond with them may prove to be a valuable source of information (Stavropoulos, Ratan et al., 2022).
These interpretations reinforce (and align with) the proposed notion of ‘digital phenotype’, suggesting that an individual's cyber-behaviour and choices, such as their user-avatar customization and bond, may operate as a unique ‘footprint’ of what they are experiencing offline, if/when appropriately translated (Loi, 2019; Stavropoulos et al., 2021; Zarate et al., 2022). This possibility is additionally strengthened by the work of Lemenager et al. (2020), who reported: (i) a consistent association between disordered gaming and bonding with the avatar, and; (ii) enhanced activation of brain regions during times an individual is consumed by thoughts regarding their avatar. Interestingly, the notion of game transfer phenomena, described as the tendency of gamers to experience altered/involuntary cognitions/thoughts, behaviours and perceptions outside of their gaming sessions, has also been associated with suffering from a medical condition and/or drug abuse, indirectly advocating for the health phenotyping/footprint potential of gaming behaviours including UAB (Ortiz de Gortari & Griffiths, 2015).
Implications, limitations, and further research
The automation of the decoding of such information using trained AI/ML procedures demonstrated here, likely revolutionizes the potential use of the UAB as a cyber-phenotype, meaning a source of information about the health/mental health of the user outside the game. More specifically, findings of the present study may: (i) pave the way for large-scale, avatar-mediated (and therefore, more gamer-friendly), low-cost, ML/AI-facilitated GD risk diagnostic procedures; (ii) help in the development of more effective GD prevention strategies, through the targeting of AI-detected GD risk gamer groups based on the way they bond with their avatars and; (iii) encourage the implementation of AIs for evaluation of user information potentially embedded within the UAB. In particular, from a conceptual perspective, and in relation to the notion of digital phenotype, the use of ML/AI to show the GD diagnostic potential of the UAB, expands past studies in the field, suggesting the need for exploration of further health and mental information likely embedded within the UAB, independent of GD risk (e.g. depression, anxiety; Lemenager et al., 2020; Loi, 2019; Ortiz de Gortari & Griffiths, 2015).
Overall, the present study suggests that GD risk can be predicted using ML/AI algorithms, that are capable of combining different variables on a large scale with reduced rates of misdiagnosis, providing more accurate diagnostic and/or risk indicators. In turn, these techniques may provide clinically relevant insights into assessment and save significant time for clinicians. Furthermore, from a GD treatment perspective, the present findings argue in favour of the utilization of the user-avatar bond when addressing GD symptoms. As Tisseron (2009) suggested, the UAB can provide the map for more accurate case formulation that can in turn drive more effective GD treatment plans, when and where avatars are involved. For instance, by observing avatar characteristics, possessions, and needs/commitments in the virtual world (e.g. using the empty chair technique to invite the ‘avatar’ to talk in a disordered gamer's session), clinicians may be able to work collaboratively with the treatment seeker/receiver to understand what they could be missing in their offline lives and plan how to pursue it to reduce their game-dependency (Tisseron, 2009). However, the findings of the present study should be interpreted taking into account the limitations of the present study, which utilised a rather small, community-sourced sample and relied exclusively on self-reported data, that might invite potential biases and confounding variables effects.
Conclusion
Despite such limitations, the present study innovatively aimed to unlock the mental health diagnostic potential, likely embedded within the UAB, through the pioneering use of a sequence of different ML classifiers and emphasizing an individual's disordered gaming risk. It did so while abiding with open science principles (i.e., accessible code and findings), such that research teams in the field can employ ML/AI to other already collected datasets related to the UAB to corroborate or negate the present findings. Furthermore, and in the context of the present study, ML/AI is converted from a game mechanic employed by industry to increase game engagement, and thus likely GD risk (Millington, 2009) into a GD protective factor.
Funding sources
VS received funding by the RMIT University, Early Career Researcher Fund ECR 2020, number 68761601, and the Australian Research Council, Discovery Early Career Researcher Award, 2021, number DE210101107.
Authors' contributions
VS contributed to the paper's conceptualization, data curation, formal analysis, methodology, project administration, and writing of the original draft. DZ, MP, NVDB, LK, AGA, AG, SB and MDG contributed to writing, reviewing, and editing the final draft.
Conflict of interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. MDG has received research funding from Norsk Tipping (the gambling operator owned by the Norwegian government). MDG has received funding for a number of research projects in the area of gambling education for young people, social responsibility in gambling and gambling treatment from Gamble Aware (formerly the Responsibility in Gambling Trust), a charitable body which funds its research program based on donations from the gambling industry. MDG undertakes consultancy for various gambling companies in the area of player protection and social responsibility in gambling. MR currently works for Mighty Serious, a gaming company focused on videogames aiming to drive positive behavioral change.
Supplementary material
Supplementary data to this article can be found online at https://doi.org/10.1556/2006.2023.00062.
References
Allwright, S. (2022, May 14). What is a good accuracy score in machine learning? Retrieved July 12, 2023, from: https://stephenallwright.com/good-accuracy-score/.
American Psychiatric Association. (2013). Diagnostic and statistical manual of mental disorders (5th ed.). American Psychiatric Association.
Bäcklund, C., Elbe, P., Gavelin, H. M., Sörman, D. E., & Ljungberg, J. K. (2022). Gaming motivations and gaming disorder symptoms: A systematic review and meta-analysis. Journal of Behavioral Addictions, 11(3), 667–688. https://doi.org/10.1556/2006.2022.00053.
Billieux, J., Flayelle, M., Rumpf, H. J., & Stein, D. J. (2019). High involvement versus pathological involvement in video games: A crucial distinction for ensuring the validity and utility of gaming disorder. Current Addiction Reports, 6, 323–330. https://doi.org/10.1007/s40429-019-00259-x.
Blinka, L. (2008). The relationship of players to their avatars in MMORPGs: Differences between adolescents, emerging adults and adults. Cyberpsychology: Journal of Psychosocial Research on Cyberspace, 2(1), 5. Full text: http://cyberpsychology.eu/view.php?cisloclanku=2008060901&article=5.
Brand, J. E., Todhunter, S. & Jervis, J. (2017). Digital Australia 2018. Eveleigh, NSW: IGEA. Retrieved July 12, 2023, from: https://igea.net/wp-content/uploads/2021/10/DA22-Report-SUMMARY-FINAL-17-10-21.pdf.
Burleigh, T. L., Griffiths, M. D., Sumich, A., Stavropoulos, V., & Kuss, D. J. (2019). A systematic review of the co-occurrence of gaming disorder and other potentially addictive behaviors. Current Addiction Reports, 6, 383–401. https://doi.org/10.1007/s40429-019-00279-7.
Burleigh, T. L., Stavropoulos, V., Liew, L. W., Adams, B. L., & Griffiths, M. D. (2018). Depression, internet gaming disorder, and the moderating effect of the gamer-avatar relationship: An exploratory longitudinal study. International Journal of Mental Health and Addiction, 16, 102–124. https://doi.org/10.1007/s11469-017-9806-3.
Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16, 321–357. https://doi.org/10.1613/jair.953.
Chen, T., He, T., Benesty, M., Khotilovic, V., Tang, Y., Cho, H., … Yuan, J. (2023). XGBoost package. CRAN.
Colder Carras, M., Stavropoulos, V., Motti-Stefanidi, F., Labrique, A., and Griffiths, M. D. (2021). Draconian policy measures are unlikely to prevent disordered gaming. Journal of Behavioral Addictions, 10, 849–853. https://doi.org/10.1556/2006.2021.00075.
Friedman, J., Hastie, T., Tibshirani, R., Narasimhan, B., Tay, K., Simon, N., … Yang, J. (2010). Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software, 33(1), 1–22. https://doi.org/10.18637/jss.v033.i01.
Granic, I., Lobel, A., & Engels, R. C. M. E. (2014). The benefits of playing video games. American Psychologist, 69(1), 66–78. https://doi.org/10.1037/a0034857.
Green, R., Delfabbro, P. H., & King, D. L. (2021). Avatar identification and problematic gaming: The role of self-concept clarity. Addictive Behaviors, 113, 106694. https://doi.org/10.1016/j.addbeh.2020.106694.
Griffiths, M. D. (2010). The role of context in online gaming excess and addiction: Some case study evidence. International Journal of Mental Health and Addiction, 8, 119–125. https://doi.org/10.1007/s11469-009-9229-x.
Han, X., Wei, L., Sun, Y., Hu, Y., Wang, Y., Ding, W., … Zhou, Y. (2021). MRI-based radiomic machine-learning model may accurately distinguish between subjects with internet gaming disorder and healthy controls. Brain Sciences, 12(1), 44. https://doi.org/10.3390/brainsci12010044.
Hill, R. (1998). What sample size is “enough” in internet survey research? Interpersonal Computing and Technology: An Electronic Journal for the 21st Century, 6(3–4).
Horton, N. J., & Kleinman, K. (2015). Using R and RStudio for data management, statistical analysis, and graphics .CRC Press.
Jiao, Y., & Du, P. (2016). Performance measures in evaluating machine learning based bioinformatics predictors for classifications. Quantitative Biology, 4, 320–330. https://doi.org/10.1007/s40484-016-0081-2.
Jo, Y. S., Bhang, S. Y., Choi, J. S., Lee, H. K., Lee, S. Y., & Kweon, Y. S. (2019). Clinical characteristics of diagnosis for internet gaming disorder: Comparison of DSM-5 IGD and ICD-11 GD diagnosis. Journal of Clinical Medicine, 8(7), 945. https://doi.org/10.3390/jcm8070945.
Karatzoglou, A., Smola, A., & Hornik, K. (2023). Kernlab: Kernel-based machine learning lab. R Package. https://cran.r-project.org/web/packages/kernlab/index.html.
King, D. L., Chamberlain, S. R., Carragher, N., Billieux, J., Stein, D., Mueller, K., … Delfabbro, P. H. (2020). Screening and assessment tools for gaming disorder: A comprehensive systematic review. Clinical Psychology Review, 77, 101831. https://doi.org/10.1016/j.cpr.2020.101831.
King, D. L., Delfabbro, P. H., Perales, J. C., Deleuze, J., Király, O., Krossbakken, E., & Billieux, J. (2019). Maladaptive player-game relationships in problematic gaming and gaming disorder: A systematic review. Clinical Psychology Review, 73, 101777. https://doi.org/10.1016/j.cpr.2019.101777.
Király, O., Koncz, P., Griffiths, M. D., & Demetrovics, Z. (2023). Gaming disorder: A summary of its characteristics and aetiology. Comprehensive Psychiatry, 122, 152376. https://doi.org/10.1016/j.comppsych.2023.152376.
Király, O., Potenza, M. N., & Demetrovics, Z. (2022). Gaming disorder: Current research directions. Current Opinion in Behavioral Sciences, 47, 101204. https://doi.org/10.1016/j.cobeha.2022.101204.
Koulouris, J., Jeffery, Z., Best, J., O'Neill, E. & Lutteroth, C. 2020, Me vs. Super(wo)man: Effects of customization and identification in a VR exergame. In: Proceedings of the 2020 CHI conference on human factors in computing systems. Vol. 2020-April (pp. 1–17), CHI conference on human factors and computing systems, Association for Computing Machinery, New York, USA. https://doi.org/10.1145/3313831.3376661.
Kuhn, M., & Silge, J. (2022). Tidy modeling with R .O'Reilly Media, Inc.
Kuhn, M., Vaughan, D., & Vaughan, M. D. (2020). Package ‘yardstick’ (version 1.2.0). Publisher of the software.
Lemenager, T., Neissner, M., Sabo, T., Mann, K., & Kiefer, F. (2020). “Who am i” and “how should i be”: A systematic review on self-concept and avatar identification in gaming disorder. Current Addiction Reports, 7, 166–193. https://doi.org/10.1007/s40429-020-00307-x.
Liao, Z., Chen, X., Huang, Q., & Shen, H. (2022). Prevalence of gaming disorder in East Asia: A comprehensive meta-analysis. Journal of Behavioral Addictions, 11(3), 727–738. https://doi.org/10.1556/2006.2022.00050.
Liew, L. W., Stavropoulos, V., Adams, B. L., Burleigh, T. L., & Griffiths, M. D. (2018). Internet gaming disorder: The interplay between physical activity and user–avatar relationship. Behaviour & Information Technology, 37(6), 558–574. https://doi.org/10.1080/0144929X.2018.1464599.
Little, R. J. A. (1988). A test of missing completely at random for multivariate data with missing values. Journal of the American Statistical Association, 83(404), 1198–1202. https://doi.org/10.1080/01621459.1988.10478722.
Lochtefeld, J. (2002). Avatar. In The illustrated Encyclopedia of Hinduism (Vol. 1, pp. 72–73). Rosen Publishing.
Loi, M. (2019). The digital phenotype: A philosophical and ethical exploration. Philosophy & Technology, 32(1), 155–171. https://doi.org/10.1007/s13347-018-0319-1.
Millington, I., & Funge, J. (2009). Artificial intelligence for games. CRC Press.
Muth, C., Oravecz, Z., & Gabry, J. (2018). User-friendly Bayesian regression modeling: A tutorial with rstanarm and shinystan. The Quantitative Methods for Psychology, 14(2), 99–119. https://doi.org/10.20982/tqmp.14.2.p099.
Nuyens, F., Kuss, D. J., Lopez-Fernandez, O., & Griffiths, M. D. (2017). The experimental analysis of problematic video gaming and cognitive skills: A systematic review. Journal de Thérapie Comportementale et Cognitive, 27(3), 110–117. https://doi.org/10.1016/j.jtcc.2017.05.001.
Nuyens, F. M., Kuss, D. J., Lopez-Fernandez, O. & Griffiths, M. D. (2019). The empirical analysis of non-problematic video gaming and cognitive skills: A systematic review. International Journal of Mental Health and Addiction, 17, 389–414. https://doi.org/10.1007/s11469-018-9946-0.
Ortiz de Gortari, A. B., & Griffiths, M. D. (2015). Game transfer phenomena and its associated factors: An exploratory empirical online survey study. Computers in Human Behavior, 51, 195–202 .https://doi.org/10.1016/j.chb.2015.04.060.
Ortiz de Gortari, A. B., Pontes, H. M., & Griffiths, M. D. (2015). The Game Transfer Phenomena Scale: An instrument for investigating the nonvolitional effects of video game playing. Cyberpsychology, Behavior, and Social Networking, 18(10), 588–594. https://doi.org/10.1089/cyber.2015.0221.
Pontes, H., Schivinski, B., Sindermann, C., Li, M., Becker, B., Zhou, M., & Montag, C. (2021). Measurement and conceptualization of gaming disorder according to the World Health Organization Framework: The development of the Gaming Disorder Test. International Journal of Mental Health and Addiction, 19, 508–528. https://doi.org/10.1007/s11469-019-00088-z.
Pontes, H. M., & Griffiths, M. D. (2019). A new era for gaming disorder research: Time to shift from consensus to consistency. Addictive Behaviors, 11 ,100242. https://doi.org/10.1016/j.addbeh.2019.106059.
Raith, L., Bignill, J., Stavropoulos, V., Millear, P., Allen, A., Stallman, H. M., … Kannis-Dymand, L. (2021). Massively multiplayer online games and well-being: A systematic literature review. Frontiers in Psychology, 12, 698799. https://doi.org/10.3389/fpsyg.2021.698799.
Ratan, R., Beyea, D., Li, B. J., & Graciano, L. (2020). Avatar characteristics induce users’ behavioral conformity with small-to-medium effect sizes: A meta-analysis of the Proteus effect. Media Psychology, 23(5), 651–675. https://doi.org/10.1080/15213269.2019.1623698.
Rehbein, F., Staudt, A., Hanslmaier, M., & Kliem, S. (2016). Video game playing in the general adult population of Germany: Can higher gaming time of males be explained by gender specific genre preferences? Computers in Human Behavior, 55, 729–735. https://doi.org/10.1016/j.chb.2015.10.016.
Şalvarlı, Ş. İ., & Griffiths, M. D. (2022). The association between internet gaming disorder and impulsivity: A systematic review of literature. International Journal of Mental Health and Addiction, 20, 92–118. https://doi.org/10.1007/s11469-019-00126-w.
Sawilowsky, S. (2009). New effect size rules of thumb. Journal of Modern Applied Statistical Medhods, 8(2), 597–599. https://doi.org/10.22237/jmasm/1257035100.
Song, K. R., Potenza, M. N., Fang, X. Y., Gong, G. L., Yao, Y. W., Wang, Z. L., … Zhang, J. T. (2021). Resting‐state connectome‐based support‐vector‐machine predictive modeling of internet gaming disorder. Addiction Biology, 26(4), e12969. https://doi.org/10.1111/adb.12969.
Šporčić, B., & Glavak-Tkalić, R. (2018). The relationship between online gaming motivation, self-concept clarity and tendency toward problematic gaming. Cyberpsychology: Journal of Psychosocial Research on Cyberspace, 12(1), 4. https://doi.org/10.5817/CP2018-1-4.
Starcevic, V., & Khazaal, Y. (2020). Problematic gaming, personality, and psychiatric disorders. Frontiers in Psychiatry, 10, 1004. https://doi.org/10.3389/fpsyt.2019.01004.
Statista (2023). Online games – worldwide. Retrieved July 12, 2023, from: https://www.statista.com/outlook/dmo/digital-media/video-games/online-games/worldwide.
Statology. (2021, September 9), What is considered a good AUC score? Retrieved July 12, 2023, from: https://www.statology.org/what-is-a-good-auc-score/.
Stavropoulos, V., Adams, B. L., Beard, C. L., Dumble, E., Trawley, S., Gomez, R., & Pontes, H. M. (2019). Associations between attention deficit hyperactivity and internet gaming disorder symptoms: Is there consistency across types of symptoms, gender and countries? Addictive Behaviors Reports, 9, 100158. https://doi.org/10.1016/j.abrep.2018.100158.
Stavropoulos, V., Gomez, R., Mueller, A., Yucel, M., & Griffiths, M. (2020). User-avatar bond profiles: How do they associate with disordered gaming? Addictive Behaviors, 103, 106245. https://doi.org/10.1016/j.addbeh.2019.106245.
Stavropoulos, V., Motti-Stefanidi F. & Griffiths M. D. (2022). Risks and opportunities for youth in the digital era. European Psychologist, 27(2), 86–101. https://doi.org/10.1027/1016-9040/a000451.
Stavropoulos, V., Pontes, H. M., Gomez, R., Schivinski, B., & Griffiths, M. (2020). Proteus effect profiles: How do they relate with disordered gaming behaviours? Psychiatric Quarterly, 91, 615–628. https://doi.org/10.1007/s11126-020-09727-4.
Stavropoulos, V., Ratan, R., & Lee, K. M. (2022). User-avatar bond: Risk and opportunities in gaming and beyond. Frontiers in Psychology, 13, 923146. https://doi.org/10.3389/fpsyg.2022.923146.
Stavropoulos, V., Rennie, J., Morcos, M., Gomez, R., & Griffiths, M. D. (2021). Understanding the relationship between the Proteus effect, immersion, and gender among World of Warcraft players: An empirical survey study. Behaviour and Information Technology, 40(8), 821–836. https://doi.org/10.1080/0144929X.2020.1729240.
Szolin, K., Kuss, D., Nuyens, F., & Griffiths, M. (2022). Gaming disorder: A systematic review exploring the user-avatar relationship in videogames. Computers in Human Behavior, 128, 107124. https://doi.org/10.1016/j.chb.2021.107124.
Tisseron, S. (2009). L’ado et ses avatars. Adolescence, 27(3), 591–600.
Torgo, L., & Torgo, M. L. (2013). DMwR: Functions and data for data mining with R. R package version 0.4, 1. CRC Press.
Van Berkel, N., D’Alfonso, S., Susanto, R., Ferreira, D., & Kostakos, V. (2023). AWARE-Light: A smartphone tool for experience sampling and digital phenotyping. Personal and Ubiquitous Computing, 27, 435–445. https://doi.org/10.1007/s00779-022-01697-7.
Van Looy, J. (2015). Online games characters, avatars, and identity. In: Mansell R. & Hwa Ang, P. (Eds.), The international encyclopedia of digital communication and society (pp. 748–758). Hoboken, NJ: Wiley-Blackwell. https://doi.org/10.1002/9781118767771.wbiedcs106.
Watson, D. S., Krutzinna, J., Bruce, I. N., Griffiths, C. E., McInnes, I. B., Barnes, M. R., & Floridi, L. (2019). Clinical applications of machine learning algorithms: Beyond the black box. BMJ, 364, l886. https://doi.org/10.1136/bmj.l886.
Wills, J. (2019). Gamer nation: Video games and American culture. Johns Hopkins University Press.
World Health Organization (2019). International classification of Diseases, Eleventh Revision. Retrieved September 1, 2023, from: https://icd.who.int/browse11.
Wright, M., & Ziegler, A. (2017). Ranger: A fast implementation of Random Forests for high dimensional data in C++ and R. Journal of Statistical Software, 77(1), 1–17. https://doi.org/10.18637/jss.v077.i01.
Yee, N., Bailenson, J. N., & Ducheneaut, N. (2009). The Proteus effect: Implications of transformed digital self-representation on online and offline behavior. Communication Research, 36(2), 285–312. https://doi.org/10.1177/0093650208330254.
Zarate, D., Dorman, G., Prokofieva, M., Morda, R., & Stavropoulos, V. (2023). Online behavioral addictions: Longitudinal network analysis and invariance across men and women. Technology Mind and Behavor, 4(2), 1–12. https://doi.org/10.1037/tmb0000105.supp.
Zarate, D., Stavropoulos, V., Ball, M., de Sena Collier, G., & Jacobson, N. C. (2022). Exploring the digital footprint of depression: A PRISMA systematic literature review of the empirical evidence. BMC Psychiatry, 22(1), 421. https://doi.org/10.1186/s12888-022-04013-y.
Accuracy reflects the ratio of correctly predicted cases, across the total number of cases. It is produced through the accumulation of the true positive and the true negative cases divided by the sum of all true positive, true negative, false positive and false negative cases. Accuracy values closer to 1 are considered desirable. Accuracy >0.90 = Excellent; 70%<Accuracy<90% = Very good; 60%<Accuracy<70% = Good; Accuracy<60% is poor (Allwright, 2022).
Area under the curve (AUC) refers to the area under the receiver operating characteristic (ROC) curve, as the latter is visualized in an orthogonal axis system/graph, where the horizontal line captures the false positive rate (FPR; 1 – specificity) and the vertical axis the sensitivity (True positive rate [TPR]; values closer to 1 are considered better/improved). AUC <0.5 = No discrimination; 0.5<AUC<0.7 = Poor discrimination; 0.7<AUC<0.8 = Acceptable discrimination; 0.8<AUC<0.9 = Excellent discrimination; AUC>0.9 = Outstanding discrimination (Statology, 2021).
Positive Predictive Value [PPV] or Precision is irrespective of the prevalence of a condition, and reflects the proportion/ratio of all the true positive classified cases divided by the addition of the true positive and the false positive cases (i.e., how many of those classified as positive were actually positive? Values closer to 1 are considered better/improved).
Recall or sensitivity is associated to the prevalence of a condition and reflects the proportion/ratio of all the true positive classified cases divided by the sum of all the true positive and the false negative classified cases (i.e., how many of the true positive cases have been recalled? Values closer to 1 are considered better/improved).
Specificity reflects the proportion/ratio of all the true negative classified cases divided by the sum of all the true negative and the false positive classified cases (i.e., how many of the true negative cases have been correctly classified? Values closer to 1 are considered better/improved).
F-Measure or F1-score/F-Score reflects the ratio of the multiplication of recall and precision, multiplied by two and then divided by the accumulation of recall and precision, such that the balance between precision and recall achieved by the model is captured. Higher values are considered better/improved (Jiao & Du, 2016).
