Search

PurposeAccidents and unintentional injuries account for the greatest number of adolescent deaths, often involving use of alcohol and other substances. This article describes the iterative design and development of Interactive Narrative System for Patient-Individualized Reflective Exploration (INSPIRE), a narrative-centered behavior change environment for adolescents focused on reducing alcohol use. INSPIRE is designed to serve as an extension to clinical preventive care, engaging adolescents in a theoretically grounded intervention for health behavior change by leveraging 3D game engine and interactive narrative technologies.MethodsAdolescents were engaged in all aspects of the iterative, multiyear development process of INSPIRE through over 20 focus groups and iterative pilot testing involving more than 145 adolescents. Qualitative findings from focus groups are reported, as well as quantitative findings from small-scale pilot sessions investigating adolescent engagement with a prototype version of INSPIRE using a combination of questionnaire and interaction trace log data.ResultsAdolescents reported that they found INSPIRE to be engaging, believable, and relevant to their lives. The majority of participants indicated that the narrative's protagonist character was like them (84%) and that the narrative featured virtual characters that they could relate to (79%). In the interactive narrative, the goals most frequently chosen by adolescents were "stay in control" (60%) and "do not get in trouble" (55%).ConclusionsWith a strong theoretical framework (social-cognitive behavior change theory) and technology advances (narrative-centered learning environments), the field is well positioned to design health behavior change systems that can realize significant impacts on behavior change for adolescent preventive health.

Game-based learning environments are designed to provide effective and engaging learning experiences for students. Predictive student models use trace data extracted from students' in-game learning behaviors to unobtrusively generate early assessments of student knowledge and skills, equipping game-based learning environments with the capacity to anticipate student outcomes and proactively deliver adaptive scaffolding or notify instructors. Reflection is a key component of self-regulated learning, and it is critical in effective learning. However, there is currently limited work exploring the utility of reflection for inducing accurate predictive student models. This article presents a predictive student modeling framework that leverages natural language responses to in-game reflection prompts to predict student learning outcomes in a game-based learning environment for middle school microbiology, CRYSTAL ISLAND. With data from a pair of classroom studies involving 118 middle school students, we investigate the accuracy of early prediction models that utilize features extracted from student trace data combined with word embedding-based representations (i.e., GloVe, ELMo) of student reflection responses. We evaluate the accuracy of the predictive models over time using data from incremental segments of each student's interaction with the game-based learning environment, and we compare against models that omit student reflection features. Results reveal that models encoding students' natural language reflections with ELMo word embeddings yield significantly improved accuracy compared to other representations, with the greatest accuracy demonstrated by an ensemble of predictive models. We discuss the implications of these results for the design of game-based learning environments.

A distinctive feature of game-based learning environments is their capacity for enabling stealth assessment. Stealth assessment analyzes a stream of fine-grained student interaction data from a game-based learning environment to dynamically draw inferences about students' competencies through evidence-centered design. In evidence-centered design, evidence models have been traditionally designed using statistical rules authored by domain experts that are encoded using Bayesian networks. This article presents DEEPSTEALTH, a deep learning-based stealth assessment framework, that yields significant reductions in the feature engineering labor that has previously been required to create stealth assessments. DEEPSTEALTH utilizes end-to-end trainable deep neural network-based evidence models. Using this framework, evidence models are devised using a set of predictive features captured from raw, low-level interaction data to infer evidence for competencies. We investigate two deep learning-based evidence models, long short-term memory networks (LSTMs) and n-gram encoded feedforward neural networks (FFNNs). We compare these models' predictive performance for inferring students' knowledge to linear-chain conditional random fields (CRFs) and naïve Bayes models. We perform feature set-level analyses of game trace logs and external pre-learning measures, and we examine the models' early prediction capacity. The framework is evaluated using data collected from 182 middle school students interacting with a game-based learning environment for middle grade computational thinking. Results indicate that LSTM-based stealth assessors outperform competitive baseline approaches with respect to predictive accuracy and early prediction capacity. We find that LSTMs, FFNNs, and CRFs all benefit from combined feature sets derived from both game trace logs and external pre-learning measures.

Over the past decade, there has been growing interest in real-time assessment of student engagement and motivation during interactions with educational software. Detecting symptoms of disengagement, such as off-task behavior, has shown considerable promise for understanding students' motivational characteristics during learning. In this paper, we investigate the affective role of off-task behavior by analyzing data from student interactions with CRYSTAL ISLAND, a narrative-centered learning environment for middle school microbiology. We observe that off-task behavior is associated with reduced student learning, but preliminary analyses of students' affective transitions suggest that off-task behavior may also serve a productive role for some students coping with negative affective states such as frustration. Empirical findings imply that some students may use off-task behavior as a strategy for self-regulating negative emotional states during learning. Based on these observations, we introduce a supervised machine learning procedure for detecting whether students' off-task behaviors are cases of emotion self-regulation. The method proceeds in three stages. During the first stage, a dynamic Bayesian network (DBN) is trained to model the valence of students' emotion self-reports using collected data from interactions with the learning environment. In the second stage, a novel simulation process uses the DBN to generate "alternate futures" by modeling students' affective trajectories as if they had engaged in fewer off-task behaviors than they did during their actual learning interactions. The alternate futures are compared to students' actual traces to produce labels denoting whether students' off-task behaviors are cases of emotion self-regulation. In the final stage, the generated emotion self-regulation labels are predicted using off-the-shelf classifiers and features that can be computed in run-time settings. Results suggest that this approach shows promise for identifying cases of off-task behavior that are emotion self-regulation. Analyses of the first two phases suggest that trained DBN models are capable of accurately modeling relationships between students' off-task behaviors and self-reported emotional valence in CRYSTAL ISLAND. Additionally, the proposed simulation process produces emotion self-regulation labels with high levels of reliability. Preliminary analyses indicate that support vector machines, bagged trees, and random forests show promise for predicting the generated emotion self-regulation labels, but room for improvement remains. The findings underscore the methodological potential of considering alternate futures when modeling students' emotion self-regulation processes in narrative-centered learning environments.