Journal of Epidemiology and Public Health

Moodsense: A Browser-Based Ensemble Sentiment Analysis System For Real Time Mood Tracking With Clinical Phq-9 Validation

Abstract

Yeshwanth Raghav Anarajula Venkata Sai

We present MoodSense, an open-source mood-tracking system that combines three complementary sentiment analysis models DistilBERT, VADER, and TextBlob into a unified Wellbeing Score validated against clinical PHQ-9 depression labels. Unlike prior tools requiring server infrastructure, MoodSense runs entirely in the browser as a Progressive Web App. The system achieves 90.6% accuracy on SST-2 (a 23.7 percentage point improvement over VADER alone), a Spearman correlation of 0.566 against emotion-valence proxy labels, and a statistically significant correlation of r = 0.61 (p < 0.001, 95% CI [0.54, 0.67]) with PHQ-9 depression scores in a cross-sectional user study (n = 312). An ablation study with confidence intervals, McNemar’s tests for pairwise model comparison, and 5-fold cross-validation confirm robustness. A built-in crisis keyword detector and an Anthropic Claude-powered conversational companion make MoodSense one of the first open-source tools integrating ensemble NLP with conversational AI support in a deployable browser application.

PDF