AI-Driven Adaptive Learning System Predicts Student Engagement and Achievement in STEM Education

Artificial intelligence is moving from the margins of educational research into the core of how universities might one day teach science, technology, engineering, and mathematics. A new study published in Discover Artificial Intelligence presents an AI-driven adaptive learning system designed to predict student engagement, forecast academic achievement, recommend personalized STEM learning strategies, and explain its own decisions to educators. The research, led by Lanlan Zhang of Yunnan Minzu University and Guangxi Normal University together with Zhenping Zhang of the Kunming Institute of Precision Machinery, addresses one of the most persistent problems in modern higher education: students who begin STEM programs but quietly disengage long before they formally drop out.

The scale of the challenge is considerable. Previous research cited in the study indicates that roughly 40 to 50 percent of students who start a STEM degree eventually abandon it or switch to a non-STEM major. Student engagement, defined as the behavioral, cognitive, and emotional commitment a learner brings to their studies, has long been identified as a key factor in achievement and retention. Yet most STEM courses still operate on a one-size-fits-all model, delivering identical content, pacing, and assessment to every student regardless of prior knowledge or learning trajectory. In large-enrollment courses, instructors face practical impossibilities when it comes to providing individualized feedback and timely intervention to each learner.

The new framework tackles this problem by integrating four functions into a single architecture: behavioral engagement classification, achievement forecasting, adaptive recommendation generation, and explainable decision support. Rather than treating these as separate research problems, as most prior studies have done, the system processes them jointly. The researchers built and tested their model using two publicly available educational datasets. The first, the Open University Learning Analytics Dataset, known as OULAD, contains anonymized records for 32,593 learners across seven course presentations, including demographic details, assessment scores, and detailed clickstream logs from a virtual learning environment. The second, the ASSISTments 2009-2010 dataset, captures roughly 525,534 interactions from 4,217 students solving 26,688 mathematics problems in an intelligent tutoring system developed at Worcester Polytechnic Institute.

One of the study’s methodological innovations lies in how engagement itself was measured. Because OULAD does not include validated psychological engagement surveys, the team constructed a behavioral engagement proxy from observable activity. Each learner received a composite score combining four normalized indicators: the number of virtual learning environment interactions weighted at 0.35, login regularity at 0.25, assessment submission rate at 0.20, and forum participation at 0.20. Students in the top 30 percent of this distribution were labeled high engagement, the middle 40 percent moderate, and the bottom 30 percent low. A sensitivity analysis comparing this baseline against four alternative weighting schemes, including equal weighting and click-dominant formulations, showed that classification performance remained stable, with accuracy ranging only from 0.938 to 0.956 across schemes. This robustness lends credibility to the proxy-based labeling strategy, though the authors caution that these labels capture behavioral engagement rather than validated psychological constructs.

The predictive engine at the heart of the system is a multi-task deep learning architecture with six modules. Heterogeneous input features pass through an embedding and normalization layer that converts categorical variables and continuous measures into a common representation. A temporal learning behavior encoder, built as a bidirectional Long Short-Term Memory network followed by a one-dimensional convolutional layer, learns patterns in how student activity unfolds over time while detecting local behavioral motifs. An attention-based fusion module then combines these temporal signals with structured learner features. From this fused representation, two task-specific branches diverge: a softmax-based engagement classifier producing probabilities across the three engagement classes, and a regression head outputting a continuous achievement score mapped from final outcomes, where distinction equals 1.00, pass equals 0.75, fail equals 0.35, and withdrawal equals 0.00.

Training the network required balancing two competing objectives, which the researchers achieved through a weighted multi-task loss function. A parameter lambda controls the relative contribution of engagement classification loss, computed as categorical cross-entropy, and achievement regression loss, computed as mean squared error. An ablation study across lambda values from 0.30 to 0.90 revealed a clear trade-off: the highest validation F1-score of 0.953 occurred at lambda equals 0.85, while the lowest RMSE of 0.163 and highest R-squared of 0.794 occurred at lambda equals 0.30. Because engagement classification directly drives the recommendation component, the team retained lambda equals 0.85 as the engagement-prioritized configuration, while acknowledging that no single value is universally optimal for both tasks.

The experimental results on OULAD were strong, though with an instructive nuance. The proposed model achieved an accuracy of 0.958 plus or minus 0.004, an F1-score of 0.959 plus or minus 0.004, and a macro-averaged one-vs-rest AUC of 0.997 plus or minus 0.001 for engagement classification, outperforming logistic regression, support vector machines, random forests, LSTM, and Transformer baselines, with paired t-tests confirming significance at p below 0.0001. However, XGBoost achieved the best overall predictive performance, with an engagement F1-score of 0.971 and an achievement prediction R-squared of 0.846, significantly outperforming the proposed model on the regression task. The authors are candid about this: the contribution of their framework lies not in beating every standalone algorithm on every metric, but in integrating engagement classification, achievement prediction, temporal modeling, adaptive recommendation, and explainability within one deployable decision-support system.

Validation on the ASSISTments mathematics data extended the framework into a genuine STEM domain. In a simulated knowledge-tracing scenario using a temporal split, where prior interactions trained the model and subsequent interactions tested it, the proposed model achieved an accuracy of 0.824, an AUC of 0.843, and an F1-score of 0.821, exceeding established knowledge-tracing baselines including Bayesian Knowledge Tracing, Deep Knowledge Tracing, SAKT, and AKT. The model also produced the highest simulated learning gain of 0.301 and a mastery progression score of 0.732. Mastery states were defined conservatively: a learner was considered to have mastered a skill only after achieving at least 80 percent correctness over a minimum of five attempts, a threshold designed to filter out instability from small sample sizes.

Translating predictions into classroom action is handled by a rule-based recommendation layer rather than a learned recommender, a deliberate design choice favoring transparency and reproducibility. Learners predicted to have low engagement or achievement scores below a defined threshold receive high-priority support, including content review, concept scaffolding, guided examples, and engagement prompts. Those with moderate engagement and declining achievement trends receive reinforcement through targeted practice and feedback, while learners with high engagement and strong predicted achievement are routed to enrichment activities such as advanced STEM extension problems. Crucially, the system explains itself: SHAP analysis, based on Shapley values from cooperative game theory, quantifies each feature’s contribution to every prediction, both globally across the test set and locally for individual learners. The analysis identified virtual learning environment activity, login frequency, study time, and assessment submissions as the most influential predictors, consistent with prior learning analytics literature.

The authors are careful to delineate what their results do and do not demonstrate. All findings derive from retrospective data, and the adaptive recommendations were evaluated under simulated conditions rather than in live classrooms, so causal effects on learning outcomes remain unproven and would require randomized controlled trials. Engagement labels, being behavioral proxies, may not capture cognitive or emotional dimensions of engagement. Generalization beyond the Open University and ASSISTments contexts, computational demands on institutions with limited infrastructure, and questions of fairness, bias, privacy, and transparency all remain open challenges. Future work outlined by the team includes prospective classroom deployment, multimodal engagement indicators such as eye-tracking and physiological sensors, collaborative learning interventions, and the use of large language models to generate natural-language explanations of predictions. If those steps succeed, the framework could offer a scalable template for identifying struggling STEM students early, tailoring support to individual needs, and giving educators not just a list of at-risk names but an interpretable account of why each student needs help.

Subject of Research: An AI-driven adaptive learning system for predicting student engagement and achievement in STEM education

Article Title: An artificial intelligence driven adaptive learning system for enhancing student engagement and academic achievement in STEM education

Article References: Zhang, L., & Zhang, Z. (2026). An artificial intelligence driven adaptive learning system for enhancing student engagement and academic achievement in STEM education. Discover Artificial Intelligence, 6(1), Article 1238. https://doi.org/10.1007/s44163-026-02279-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02279-9

Keywords: artificial intelligence, adaptive learning, STEM education, student engagement, academic achievement, machine learning, learning analytics, explainable AI, knowledge tracing, multi-task learning, OULAD, ASSISTments