This systematic literature review examined how computer vision, wearable sensing, and multimodal artificial intelligence (AI) are used to infer psychological, affective, and behavioural states in tennis and related racket sports. Target constructs included the zone or flow-like states, emotion, stress, attention-related behaviour, and psychologically relevant performance states. Behavioural and perceptual studies were treated as indirect evidence when they supplied a plausible observable proxy rather than a validated psychological measure. Reporting followed PRISMA 2020. A structured Scopus search identified 309 records published from 2015 to 2025; 41 full texts were assessed and 10 studies met the archived eligibility criteria. Eligible studies were English-language peer-reviewed articles involving racket-sport contexts and digital sensing or computational methods with a behavioural, affective, or psychological target. Comparator conditions were extracted when reported but were not required for inclusion. The synthesis used descriptive and thematic analysis because constructs, labels, samples, settings, and performance metrics were heterogeneous. Multimodal approaches showed advantages in the few studies with direct modality comparisons, while naturalistic tennis affect recognition reached 68.9% accuracy in one match-footage study. However, no common external-validation benchmark was identified. Reviewer-level screening logs were not retained, so Cohen’s kappa cannot be reconstructed retrospectively; FICO was used as an internal appraisal rubric, not a validated risk-of-bias tool. Future work should prioritize naturalistic datasets, psychometric anchoring, subject-independent validation, transparent models, and privacy-preserving governance.