Simulation-Based Hybrid LSTM–XGBoost Framework for Early Prediction of Simulated Learning-Loss Risk Using LMS Log Features

Agunawan (Instutut Teknologi dan Bisnis Nobel Indonesia)
Ruslan (Universitas Negeri Makassar)
Dandi Darmadi (Universitas Andi Djemma)

Journal of Embedded Systems, Security and Intelligent Systems
Vol 7 No 3 (2026): September 2026

Purpose – Learning loss remains a concern in Indonesian higher education after the pandemic, while LMS-based early warning systems remain limited for timely intervention. This study designs and evaluates a hybrid LSTM–XGBoost framework for early prediction of simulated LMS-based learning-loss risk, prioritizing architectural innovation over metric superiority. Methods – Using 6,440 synthetic LMS records structured from publicly documented Sevima EdLink log fields, early risk was predicted from weeks 1–4 features and labeled at week 6 with behavioral proxy rules. An 80:20 group-aware train–test split by student identifier used leakage-safe, training-only scaling and one-hot encoding. Models were compared as an untuned architectural benchmark with matched-feature ablations, approximate confidence intervals, calibration/threshold analysis, and feature importance inspection. Findings – The temporal-only LSTM model produced the strongest overall predictive performance, while the LSTM with static-feature fusion showed comparable results. The hybrid LSTM–XGBoost decision stage remained competitive but did not outperform the matched LSTM configurations or demonstrate a clear advantage over simpler baseline models. Ablation analysis further showed that neither static-feature integration nor the use of XGBoost as the final decision engine provided a meaningful performance improvement. Changes in engagement, feedback activity, interaction patterns, and early time-on-task emerged as the most influential simulated indicators of learning-loss risk. Research Implications – The hybrid architecture offers a replicable blueprint for LMS early-warning pipelines that separate temporal extraction (LSTM) from risk classification (XGBoost). Institutional use requires real LMS validation, recall optimization, and ethical compliance. Originality – This simulation-based late-fusion LSTM–XGBoost blueprint separates the prediction window (weeks 1–4) from the outcome window (week 6) and evaluates architectural contribution.