MindPhone AI: Intelligent Prediction of Smartphone Addiction

Abstract

Excessive smartphone use is increasingly framed as a behavioral addiction, traditionally assessed through validated self-report instruments such as the Smartphone Addiction Scale. Machine-learning approaches promise a more scalable, passively-collected alternative to survey-based screening, but most published systems still rely on self-reported psychometric ground truth for model training. Methods. We describe a Flask-based system that extracts Android Digital Wellbeing usage counters like usage minutes, app opens, notification counts, screen-time percentage, either from a connected device via ADB or from a generated fallback dataset and classifies each usage profile into one of five heuristically defined addiction-risk tiers such as MINIMAL, LOW, MODERATE, HIGH, SEVERE using a Random Forest classifier where n_estimators=100 with Standard Scaler feature normalization and LabelEncoder target encoding. The risk labels themselves are produced by a deterministic weighted formula over the same usage counters rather than an independently measured outcome. Results is on the project's static 14-app sample, the classifier reached 100% training accuracy with no held-out test set; illustrative single predictions for example an Instagram usage pattern were classified SEVERE with 86% predicted-class probability. These figures reflect fit to a formula-derived label on training data, not validated predictive performance, and must not be reported as generalization accuracy.