PhD Thesis Defense by Eleftherios Kalafatis titled "Content generalization and automated testing through intelligent agents in dynamic serious games for health"

In May 2026 Eleftherios Kalafatis completed his PhD Thesis titled "Content generalization and automated testing through intelligent agents in dynamic serious games for health" at the National Technical University of Athens (NTUA).

Abstract: The present thesis focuses on the design and implementation of a comprehensive framework for managing, generalizing, and automatically testing content in Serious Games for Health (SGHs), incorporating artificial intelligence techniques to support personalization and improve the effectiveness of digital health interventions. The development of SGHs is typically a time-consuming and costly process, while the lack of standardized approaches for content reuse and objective evaluation limits their scalability and broader application. The proposed framework addresses these challenges through three main research axes: the generalization and reuse of SGH content, the development of Procedural Content Generation (PCG) techniques for personalized interventions, and the automated evaluation of dynamic game content and PCG mechanisms using intelligent agents. The framework is demonstrated through five original Serious Games for Health addressing Obstructive Sleep Apnea (OSA), nutritional education, stress management, type 1 diabetes mellitus (T1DM), and obesity. The first axis focuses on separating health-related content from the underlying game mechanics. Game objectives, mechanisms, graphics, and other domain-dependent elements are systematically identified and transformed into generalized representations. A color-coding methodology is introduced to replace domain-specific references, while a dedicated "Content Manager" provides programming interfaces for switching and adapting content. This allows the same game mechanics to be reused across different health domains without redesigning the entire application. The effectiveness of this approach is demonstrated by adapting an SGH originally developed for Obstructive Sleep Apnea to the field of depression while preserving the core gameplay mechanisms. The second axis concerns the development and generalization of Procedural Content Generation techniques. A PCG methodology based on a genetic algorithm is initially developed for an OSA Serious Game to generate personalized game-character profiles and is subsequently extended to applications involving type 1 diabetes and childhood obesity. The genetic algorithm encodes game characteristics as genes within chromosomes and uses fitness functions based on player behavior and performance to dynamically generate personalized recommendations, missions, and difficulty levels. In applications targeting T1DM and childhood obesity, physical activity sensor data are additionally incorporated to adapt game recommendations according to the child's daily behavior. The third research axis introduces automated testing methodologies based on machine learning and deep reinforcement learning agents. Player modeling is initially explored through biofeedback sensors, screen recordings, and subjective player reports to estimate indicators such as engagement and stress. In the Serious Game developed for stress management, physiological measurements obtained through photoplethysmography sensors demonstrate strong correlations between measured and model-estimated stress indicators. To reduce dependence on extensive human testing, the thesis proposes an automated evaluation framework in which deep reinforcement learning agents interact directly with Serious Games. The framework consists of three main components: the SGH environment containing the game logic, an Interaction Interpreter that translates game mechanics into representations accessible to intelligent agents, and a Deep Reinforcement Learning environment responsible for agent training and evaluation. The Game Description Language and the Gymnasium environment are used to standardize interactions, while Stable Baselines 3 is employed for implementing the reinforcement learning algorithms.

Two complementary validation approaches are introduced: exhaustive validation, in which trained agents interact with a broad range of procedurally generated content, and scenario-based validation, in which agents encounter realistic content configurations controlled exclusively by the PCG mechanism. Evaluation of the framework in the Serious Game for OSA demonstrates a 5% improvement in performance for genetic-algorithm-based content generation compared with random content generation, while agents reach the 50% win-rate threshold earlier during training. Changes in dominant genetic-algorithm genes are also found to correlate directly with variations in agent success rates, supporting the effectiveness of the adaptation mechanism.

In addition to automated evaluation, experimental studies with real users demonstrate improvements in both user experience and clinical outcomes. In the OSA Serious Game, the personalized approach significantly increases users' feeling of competence while reducing negative experience. In the intervention targeting type 1 diabetes and childhood obesity, a statistically significant reduction in BMI z-score is observed, providing evidence of the potential clinical benefits of personalized Serious Games for Health.

Overall, the thesis contributes a unified methodological framework for the development, personalization, reuse, and automated evaluation of Serious Games for Health. By combining content generalization, Procedural Content Generation, artificial intelligence, and deep reinforcement learning agents, the proposed approach aims to reduce development time and cost while enabling more adaptive, personalized, and objectively evaluated digital health interventions. The results demonstrate the framework's potential to support chronic disease self-management and facilitate the broader adoption of Serious Games as complementary tools in modern healthcare.