In online education, identifying at-risk students early remains an unsolved problem. It is not for lack of accurate models: the limit is trust. Deep neural systems have produced predictions that are hard to justify, and this black-box issue has slowed adoption in real pedagogical settings. EduRiskX addresses the problem with a neuro-symbolic strategy: a temporal Transformer-based predictor works alongside an F-Logic rule engine. The neural component models longitudinal student activity sequences using temporal attention, class-weighted loss, and dynamic weekly truncation. The symbolic component acts as a data-driven expert system: the rule base is built exclusively from the training data, but it is grounded in established educational theories — Engagement Theory and the Student Integration Model — to mimic the diagnostic logic of human educators. The two signals, neural risk probability and symbolic confidence score, are combined through a logistic regression-based fusion mechanism that learns the relative contribution of each component.

The results on the Open University Learning Analytics Dataset are solid: with a strict 80/10/10 student-level split, the system reaches an accuracy of 0.900 and an F1-score of 0.894 at the end of the semester (Week 38). More relevant for early intervention, the average detection week is 9.32 and the detection rate is 94.30 percent. Compared with time-series models such as PatchTST and iTransformer, and with LSTM and CNN baselines, EduRiskX improves recall and identifies risk earlier under identical conditions.

The point is not just performance. The presence of F-Logic rules shifts the focus from prediction to explanation. In education, where decisions affect people and academic paths, a risk score without reasons is difficult to use. Symbolic rules link predictions to observable behavioral patterns and theoretical constructs, offering an auditable trail. That is a structural advantage over purely neural models.

There is a constraint to note: the rule base is built exclusively from training data. This ensures consistency with available evidence, but it raises an evergreen issue for neuro-symbolic systems: the explanation is only as good as the data it comes from. If the data reflect existing inequalities or behaviors of specific cohorts, rules can formalize them rather than dismantle them. This is not unique to EduRiskX, but a common trade-off for all systems that promise interpretability without external data governance.

From a deployment perspective, the source does not specify whether the framework is intended for cloud or on-premise environments. Still, for those evaluating self-hosted architectures in educational contexts, the issue is relevant: student activity data and risk signals are typically subject to privacy constraints and may require local control. A symbolic component does not remove infrastructure requirements, but it can make the system more verifiable during audits, because rules can be inspected unlike a purely neural network. That is not a compliance guarantee, but it changes the comparison between opaque and explainable models.

The real test for approaches like EduRiskX will not be in papers, but in institutional decision-making. If an online university has to intervene with a student mid-semester, it will first ask why, not just how likely. Neuro-symbolic frameworks try to answer that question without sacrificing the predictive power of Transformers. Whether that will be enough to overcome the trust crisis that has so far held back black-box models remains open.