Abstract: Smartphone habits among students often signal their academic trajectory, yet capturing these patterns for reliable forecasting remains challenging. This study aimed to create an efficient deep learning system that interprets everyday mobile device activity to anticipate student grades and identify intervention points.

Anonymized usage records from 1,200 undergraduates captured key metrics, including total screen hours, app types (social, study, gaming), peak usage times, and basic demographics such as age and program year. After cleaning sparse data through simple averaging and normalizing sequences, we built a hybrid neural network: convolutional layers first detected local patterns in app switches and session lengths, followed by LSTM units to model daily rhythms across weeks, with an attention layer spotlighting critical periods like late-night scrolling. The system trained on an 80/20 split using standard optimization, benchmarked against simpler classifiers.

Results showed our model achieving 93.2% prediction accuracy for final GPA categories, with 92% F1-score—outpacing random forests (85%) and gradient boosters (87%) by wide margins. Key insights revealed that keeping recreational apps below 4 hours daily correlated with 15-20% higher outcomes, while fragmented study sessions hurt more than total time.
These discoveries offer educators a practical, non-intrusive tool to spot at-risk students early through existing device data, promoting balanced tech use without overhauling routines. By turning passive logs into proactive guidance, this approach bridges digital behavior and learning success in modern campuses.

Keywords: Smartphone analytics, Academic forecasting, Hybrid neural networks, Behavioral thresholds, Educational AI.

Download: | DOI: 10.17148/IMRJR.2026.030206