PaperlayerView exact revision on arXiv

Source-linked AI summary

Closing the Loop in Affect-Driven Game Adaptation: A Systematic Review

Phil Lopes, Nuno Fachada, Maria Fonseca

arXiv:2505.01351v2cs.HC

TL;DR

The review addresses the limited evidence on integrating sensing, player-experience modeling, and content adaptation into complete empirically evaluated game systems. Using a PRISMA-guided synthesis of 23 studies, it finds that complete loops were uncommon and usually used affective information to support objectives such as difficulty, engagement, rehabilitation, or performance rather than to target affective state itself. Within the review scope, stress, anxiety, horror, and related states were rarely explicit adaptation targets.

  • Problem

    Few studies examine how sensing, modeling, and adaptation are integrated into complete, empirically evaluated gameplay systems.

    Source: §I · p. 1, B. Contribution and Research Focus · p. 2

  • Method

    The paper conducts a PRISMA-guided systematic review of empirical complete-loop systems linking player data, experience modeling, and adaptive game content.

    Source: B. Contribution and Research Focus · p. 2, 3) Content Adaptation · p. 3, §IV · p. 6

  • Results

    Complete-loop systems were relatively uncommon and predominantly used affective information for difficulty, engagement, rehabilitation, or performance objectives rather than explicit affective adaptation.

    Source: §IV · p. 6, §V · p. 10, §VI · p. 13

  • Takeaways & Limitations

    Affective information entering an adaptive loop should not be equated with affective adaptation, especially for stress, anxiety, horror, and related states.

    Source: §VI · p. 13

    The targeted keyword strategy may underrepresent studies using alternative terminology such as biofeedback, affective interaction, or physiological self-regulation.

    Source: E. Limitations of this Review · p. 12

Abstract

from arXiv · show

Recognizing player state is only one component of affective game adaptation; inferred experience must also be translated into adaptive interventions that modify gameplay or game content. Although player experience modeling and content adaptation are established research areas, fewer studies examine how sensing, modeling, and adaptation are integrated into complete, empirically evaluated gameplay systems. This PRISMA-guided systematic review analyzes 23 empirical studies published from January 1, 2015, to December 31, 2025, that implement a complete experience-driven loop defined here as the combination of player data acquisition, player experience modeling, and adaptive game content. Complete-loop systems were relatively uncommon in the retrieved corpus, and the selected systems were predominantly oriented toward dynamic difficulty adjustment, engagement, rehabilitation, or performance-related goals. Game telemetry was the dominant input modality, while non-invasive sources with affective relevance, such as facial expression analysis and peripheral interaction data, were less common. Knowledge-based methods, including rule-based systems and heuristics, dominated both modeling and adaptation because of their interpretability and low deployment requirements, whereas machine learning approaches were less frequent and remained constrained by data availability, transparency, and runtime integration challenges. Most importantly, affective information was often used to support challenge calibration or related adaptation objectives, while stress, anxiety, horror, and related affective states were rarely addressed as explicit adaptation targets. These findings identify a gap within this review scope: affective information may enter an adaptive loop without making affective state the objective of adaptation.

I. INTRODUCTION

The review asks whether affective information is connected to complete adaptive gameplay loops rather than used only for experience recognition. It distinguishes affect-informed adaptation from adaptation that explicitly targets affective experience.

I. INTRODUCTION

  • Affective computing must translate inferred states into interventions that shape subsequent experience, not merely recognize affect.Source: §I · p. 1

A. Motivation and Scope

  • Complete-loop adaptation can pursue difficulty, engagement, rehabilitation, performance, or affective experience, so affective input alone does not establish affective adaptation.Source: §I · p. 1, A. Motivation and Scope · p. 2

I. INTRODUCTION

  • The review examines whether sensing, experience modeling, and content adaptation operate together in empirically evaluated systems.Source: §I · pp. 1–2

B. Contribution and Research Focus

  • The review distinguishes affect-informed adaptation, which supports another objective, from affective adaptation, which makes affective state the explicit objective.Source: B. Contribution and Research Focus · p. 2

II. CONCEPTUAL FRAMEWORK

  • The conceptual framework defines the loop as player data feeding a player experience model whose inferences drive game-content adaptation.Source: Figure 1 · p. 2, §II · p. 2

A. Defining the Experience-Driven Loop

The review defines a complete loop through linked player data, experience modeling, and content adaptation. It classifies input modalities and contrasts interpretable knowledge-based models with more data-intensive statistical and machine-learning approaches.

A. Defining the Experience-Driven Loop

1) Player Data:

  • Player data include telemetry, physiology, questionnaires, bodily expressions, peripheral input, and third-person annotations.Source: 1) Player Data · p. 3

2) Player Experience Model:

  • The player experience model is any process that uses input data to infer affect, engagement, challenge, performance, preference, or another adaptation-relevant construct.Source: 2) Player Experience Model · p. 3
  • Knowledge-based methods are interpretable and data-light, whereas statistical and machine-learning methods can identify harder-to-encode patterns but require more data and reduce transparency.Source: 2) Player Experience Model · p. 3
  • Modeling remains distinct from adaptation: modeling assesses player state, while adaptation governs the system’s response to that assessment.Source: 3) Content Adaptation · p. 3, 2) Player Experience Model · p. 3

3) Content Adaptation:

Content adaptation is the autonomous transformation of game content or mechanics in response to modeled player experience. The review operationalizes affective objectives separately and uses a targeted PRISMA-referenced search focused on complete adaptive systems.

3) Content Adaptation:

  • Content adaptation transforms game mechanics or digital content to optimize a player experience or emotional response through system-driven responses.Source: 3) Content Adaptation · p. 3
  • Adaptation objectives motivate computational methods, which produce changes to layouts, enemies, pacing, narrative, interfaces, or audio-visual effects.Source: 3) Content Adaptation · p. 3
  • Affect-informed adaptation uses affective information for objectives such as difficulty or engagement, whereas affective adaptation explicitly targets affective experience.Source: 3) Content Adaptation · p. 3

B. Identifying the Experience-Driven Loop

C. Focusing on Horror, Stress, and Anxiety

III. METHODOLOGY

B. Inclusion and Exclusion Criteria

The review included peer-reviewed English studies from 2015–2025 that empirically evaluated human-player interaction leading to game adaptation within a complete loop. Searches covered five databases selected across computing, games, affective computing, and applied health.

B. Inclusion and Exclusion Criteria

C. Selected Databases

  • Searches covered IEEE Xplore, ACM Digital Library, SpringerLink, Scopus, and b-on to span computing, game research, affective computing, and applied health domains.Source: C. Selected Databases · pp. 4–5
  • IEEE Xplore required an adapted full-text-and-metadata query preserving the game, PCG, affective-state, and adaptation dimensions.Source: C. Selected Databases · p. 5

D. Reviewing Procedures

The review screened studies using predefined eligibility criteria and classified selected complete-loop systems across methodological and application dimensions. An evidence-maturity indicator then distinguished feasibility demonstrations from stronger evaluations of adaptive effects.

D. Reviewing Procedures

  • Three reviewers independently screened titles and abstracts after duplicate removal, applied the same criteria during full-text screening, and resolved disagreements by consensus.Studies outside scope, reviews, conceptual work, and studies failing inclusion criteria were excluded.Source: D. Reviewing Procedures · p. 5

E. Dimensions of Analysis

  • The analysis classified studies by sample size, input data, adaptation objectives, modeling techniques, adaptation methods, game type, adapted content, and relation to stress, anxiety, or horror.The relation dimension used a five-point scale ranging from minimal explicit relation to direct use as an adaptation target.Source: E. Dimensions of Analysis · p. 5

F. Empirical Evidence Maturity

  • Evidence maturity was coded by consensus using evaluation scale, comparison conditions, outcome measures, quantitative analysis, and direct testing of the adaptive mechanism.Limited evidence covered exploratory or small-scale evaluations, moderate evidence included structured evaluations with constraints, and stronger evidence involved controlled comparisons or deployment-scale testing with direct adaptation analysis.Source: F. Empirical Evidence Maturity · p. 5
  • The PRISMA process yielded 23 studies for in-depth analysis after the defined selection and screening process.The flow diagram summarizes records identified, screened, excluded, and included.Source: F. Empirical Evidence Maturity · p. 5, Figure 2 · p. 6

IV. RESULTS

The review retained 23 empirically evaluated complete-loop studies after screening and found that most systems pursued difficulty, performance, engagement, rehabilitation, or related goals. Affect-informed systems were more common than systems making affective experience itself the explicit adaptation objective.

A. PRISMA Filtering Results

  • 23 studies remained from 672 identified records after duplicate removal, title-and-abstract screening, and full-text eligibility assessment.Forty full texts were assessed, and the main later-stage exclusion reason was the absence of a complete empirically evaluated loop.Source: A. PRISMA Filtering Results · p. 6, §IV · p. 6

B. Overview of the Selected Corpus

  • The selected corpus was dominated by difficulty-, performance-, and task-calibration systems spanning entertainment, rehabilitation, education, and exergames.These systems adapted tasks, enemies, pacing, level structure, or player parameters to maintain challenge, engagement, training, or performance.Source: B. Overview of the Selected Corpus · pp. 6–7
  • Affective or physiological information was often integrated into systems that still adapted difficulty or engagement rather than targeting affective state itself.Examples included EEG-, facial-expression-, EDA-, and biometric-informed adaptations.Source: B. Overview of the Selected Corpus · p. 6
  • Explicit affective adaptation was comparatively rare, although some studies adapted content using arousal, valence, emotion regulation, or emotional-state objectives.Nogueira et al. combined physiological arousal and valence modeling with adaptations in a procedural horror game.Source: B. Overview of the Selected Corpus · pp. 6–7
  • Applied and serious-game studies appeared frequently, but they commonly used small samples, constrained interactions, or domain-specific criteria, while large-scale deployments were rare.Sifa et al. was identified as an exception because of its commercial online-game setting and unusually large participant base.Source: B. Overview of the Selected Corpus · p. 7
  • Within the review scope, stress, anxiety, horror, and related affective states were seldom explicit adaptation targets.The review treats this scarcity as a finding about the selected corpus rather than as a failure of its scope.Source: B. Overview of the Selected Corpus · p. 7

C. Comparative Analysis by Dimension

Across 23 complete-loop studies, telemetry and dynamic difficulty adjustment were most common, while knowledge-based methods dominated modeling and adaptation. Affective information was often incorporated without making affective state the explicit adaptation target.

Input data

  • Game telemetry was the predominant input source, while physiology and questionnaires appeared in several studies and bodily expressions or peripheral input appeared in only five studies combined.Thirteen of 23 studies used more than one input type.Source: C. Comparative Analysis by Dimension · pp. 7–8

Adaptation objectives

  • 20 studies used dynamic difficulty adjustment, making it the dominant adaptation objective, while engagement appeared in eight and emotion in seven.GT input was used in 19 of the 20 DDA-focused studies.Source: C. Comparative Analysis by Dimension · p. 9
  • Affective information often supported difficulty, challenge, engagement, rehabilitation, or performance regulation rather than direct adaptation of affective state.Five of the seven emotion-oriented studies also included DDA.Source: C. Comparative Analysis by Dimension · pp. 9–10

Modeling and adaptation methods

  • Knowledge-based approaches dominated both stages, with 14 modeling instances and 21 adaptation use cases, compared with nine statistical and seven ML modeling instances and five ML inference use cases.Modeling and adaptation techniques were coded separately on a best-effort basis because some studies did not clearly distinguish them.Source: C. Comparative Analysis by Dimension · pp. 9–10

Game domains and adapted content

Affective targets

  • Complete-loop systems explicitly targeting stress, anxiety, horror, or closely related affective states were rare within the selected corpus.This scarcity was treated as a finding within the review scope rather than as a failure of the review.Source: C. Comparative Analysis by Dimension · p. 10

V. DISCUSSION

The review finds that complete-loop adaptation is implemented, but most systems use affective or experiential information for performance-oriented objectives rather than directly targeting affective state. Knowledge-based methods and telemetry predominate, while evidence about adaptive mechanisms producing intended affective effects remains limited.

V. DISCUSSION

  • Most complete-loop systems used affective or experiential information to support difficulty, engagement, rehabilitation, or performance objectives rather than directly target affective state.This pattern is the review’s central distinction between affect-informed and affective adaptation.Source: §V · p. 10, A. Main Technical Findings · p. 10

A. Main Technical Findings

  • Knowledge-based methods predominated because they are interpretable, require less training data, and can be deployed in constrained settings.ML methods were less frequent and faced data, transparency, and runtime integration challenges; RL was rare in the corpus.Source: A. Main Technical Findings · p. 10
  • Game telemetry was the dominant input modality, whereas facial-expression analysis and peripheral input were less common than telemetry and physiology.The discussion attributes telemetry’s prevalence to ease of collection, non-intrusiveness, and suitability for large-scale studies.Source: A. Main Technical Findings · p. 10
  • Complete-loop systems were often developed in applied contexts such as health, rehabilitation, education, and training, but adapted content primarily served non-affective objectives.The applied distribution is secondary to the review’s central objective distinction.Source: A. Main Technical Findings · p. 10
  • Many studies demonstrated feasibility, usability, descriptive feedback, or small-scale implementation without establishing that the adaptive mechanism produced the intended experiential or affective effect.The evidence-maturity coding therefore limits conclusions about adaptive efficacy.Source: A. Main Technical Findings · p. 10

B. Affective Information Does Not Necessarily Imply Affective Adaptation

The review distinguishes affective information entering an adaptive loop from affective adaptation itself, finding that complete-loop systems remain uncommon and usually target difficulty, engagement, rehabilitation, or performance. Explicit affective objectives such as stress, anxiety, horror, or tension were rare, while integrating sensing, modeling, and coordinated adaptation remained difficult.

B. Affective Information Does Not Necessarily Imply Affective Adaptation

C. Interpretation and Implementation of the Experience-Driven Loop

D. Answering the Research Questions

  • Complete-loop systems predominantly addressed dynamic difficulty adjustment, engagement, rehabilitation, or performance, with telemetry and knowledge-based methods more common than machine learning.Machine learning appeared in several studies but was less frequent.Source: D. Answering the Research Questions · p. 11
  • Affective, physiological, or experiential cues more often supported challenge calibration, engagement, rehabilitation, or performance regulation than affective state as an independent adaptation objective.Stress, anxiety, and related states were rarely treated as explicit adaptation targets.Source: D. Answering the Research Questions · pp. 11–12, B. Affective Information Does Not… · p. 11

E. Limitations of this Review

  • The review’s targeted search strategy may underrepresent studies using alternative terms such as biofeedback, affective interaction, or physiological self-regulation, especially in clinical contexts.Evidence-maturity coding was comparative rather than a formal assessment of methodological quality, risk of bias, or construct validity.Source: E. Limitations of this Review · p. 12
  • Ambiguous boundaries between player experience modeling and content adaptation complicated classification when the same measure both represented player ability and directly determined adaptation.Examples included score-based timer changes and heuristic time or distance adjustments.Source: E. Limitations of this Review · p. 12

F. Recommendations and Future Directions

The review recommends making adaptation objectives explicit, expanding affectively relevant input modalities, and reporting content-generation mechanisms as components of complete adaptive loops. It also emphasizes ethical safeguards and cautions that generative methods require modeling, objectives, evaluation, controllability, interpretability, and user-safety provisions.

F. Recommendations and Future Directions

  • Future systems should distinguish affective information used for DDA, engagement, rehabilitation, or performance calibration from systems treating affective state itself as the adaptation objective.This distinction is especially important because stress, anxiety, horror, and related states were rarely explicit adaptation targets.Source: F. Recommendations and Future Directions · p. 12
  • Facial-expression analysis and peripheral interaction data should be explored as less represented affectively relevant input modalities, with transparent interpretation and evaluation against the intended objective.These modalities may support less intrusive affect-informed or affective adaptation.Source: F. Recommendations and Future Directions · p. 12
  • Procedural content generation and other content-modification methods should be explicitly connected to the player experience model and adaptation objective within the reported loop.This framing separates content variation mechanisms from complete experience-driven adaptation.Source: F. Recommendations and Future Directions · p. 12
  • Large Language Models may generate adaptive dialogue, narratives, or scenarios, but they are not complete adaptive systems without experience modeling, explicit objectives, empirical evaluation, and ethical safeguards.Practical challenges include computational cost, response consistency, controllability, interpretability, and user safety.Source: F. Recommendations and Future Directions · p. 12

G. Ethical Considerations

  • Affect-adaptive systems should address transparency, informed consent, and handling of affective or physiological data, especially for vulnerable users and deliberately targeted states.Designs should also avoid manipulative difficulty tuning or reinforcing negative affective states.Source: G. Ethical Considerations · pp. 12–13

VI. CONCLUSION

  • The review’s conclusion is that affective information entering an adaptive loop should not be equated with affective adaptation.Far fewer systems treated affective state itself as the primary adaptation objective.Source: §VI · p. 13

APPENDIX A GLOSSARY

The glossary defines methods and concepts used in affective game adaptation, including affective modeling, dynamic difficulty adjustment, procedural content generation, and common machine-learning and optimization approaches.

  • Affective Modeling estimates a person’s affective state from observable data.Source: Appendix A · p. 13
  • Dynamic Difficulty Adjustment automatically modifies difficulty in response to a user’s performance, state, or experience.Source: Appendix A · p. 13
  • A Partially Ordered Set Master is an online DDA algorithm that updates beliefs over ordered difficulty settings from feedback.Source: Appendix A · p. 13
  • Procedural Content Generation algorithmically creates content such as levels, tasks, or game scenarios.Source: Appendix A · p. 13
  • Reinforcement Learning learns actions through rewards or penalties received from interaction, while Deep Reinforcement Learning uses neural networks for complex spaces.Source: Appendix A · p. 13
  • Russell’s Circumplex Model of Affect represents affective states along valence and arousal dimensions.Source: Appendix A · p. 14

APPENDIX B DETAILED SUMMARIES OF INCLUDED STUDIES

The included studies implement experience-driven loops across rehabilitation, education, entertainment, immersion, and affect regulation, using telemetry, physiological signals, behavioral modeling, and adaptive content. Most systems target difficulty, engagement, performance, or training outcomes, while fewer directly adapt affective experience such as horror, stress, or anxiety.

Difficulty and performance adaptation

  • Player-modeling loops commonly used telemetry or performance signals to adjust difficulty, including rehabilitation tasks, RPG quests, maze complexity, platform geometry, and exergame timing.Examples include MCTS-based rehabilitation adaptation, TF-based quest recommendation, DRL maze adaptation, hybrid EDA-performance DDA, and score-based timer changes.Source: Appendix B · pp. 14–16

Personalized content modeling

  • Several systems modeled experience or preferences to personalize content, including TF-ASP adaptation to player ability, regression-based enemy-speed adaptation, weighted performance indices, and personality-based difficulty selection.Player-based enemy-speed adaptation increased self-reported valence, while the prototype and personality studies reported broadly positive or engagement-related outcomes.Source: Appendix B · pp. 14, 16, 17

Affect-focused and immersive adaptation

  • Nogueira et al. directly adapted a procedural horror game using real-time arousal and valence estimates, changing level generation, encounters, audiovisual effects, and player attributes.Biofeedback significantly affected immersion, tension, positive affect, and negative affect across symbiotic, equilibria, and control conditions.Source: Appendix B · p. 14

Adaptation objectives

  • Across the broader set, adaptation typically pursued flow, engagement, learning, rehabilitation, or difficulty calibration rather than explicitly regulating horror, stress, or anxiety.This pattern includes engagement-oriented DDA, adapted challenge preferences, EEG-triggered excitement optimization, and emotion-driven programming-game adaptation aimed at reducing frustration or boredom.Source: Appendix B · pp. 14–15

Knowledge-based adaptation

  • Rule-based and heuristic systems remained prominent, including emotion-state rules with Markov transitions, fuzzy affect regulation, semantic expert rules, and physiological difficulty mappings.These systems adapted educational tasks, pacing, sound, arithmetic difficulty, serious-game feedback, and shooter parameters from modeled player states.Source: Appendix B · pp. 15–16

Machine-learning adaptation

  • Machine-learning approaches supported several complete loops, with RFC-based facial-behavior adaptation converging faster than heuristics and neural-network enemies producing stronger long-term engagement than heuristic enemies.The neural-network condition also slightly increased stress-related tension-pressure, while facial-behavior frustration effects were mixed.Source: Appendix B · p. 15

Physiological and affective sensing

  • Affectively relevant sensing was used for immersion and emotional regulation, with fNIRS-based visual adaptation slightly increasing immersion and EEG-adaptive VR producing more stable emotion regulation and higher satisfaction.The fNIRS system significantly improved temporal dissociation, while the EEG system adapted emotionally valenced scene elements such as animals and imagery.Source: Appendix B · pp. 15–16

Rehabilitation and serious games

  • Rehabilitation and serious-game systems adapted challenge and feedback through haptic, biometric, motion, telemetry, and expert-defined inputs, generally emphasizing motivation, usability, safe training, or challenge calibration.The haptic rehabilitation loop was preferred over non-DDA trials by 36% versus 7%, while the speech-therapy framework adapted task complexity, hints, feedback, and timing.Source: Appendix B · pp. 14, 16, 17

Related papers

Loading 2505.01351v2…