Abstract:
Depression is a prevalent mental health disorder in the world, but diagnosis is
constrained by stigma, cost and lack of clinical services. This study describes
an AI-Based Real-Time Multi-Agent Depression Detection and Therapy System
which aims at early detection and provides therapy suggestions. NLP and speech
emotion analysis are incorporated to assess the user input and calculate the level
of depression using the Patient Health Questionnaire-9 (PHQ-9). Five specialized
agents work collaboratively to process text and voice data, retrieve linguistic and
emotional features, and categorize mental states from minimal to severe. A fusion
model of PHQ-9 scores, classifier confidence, and speech-based emotion indicators
produces a factor-calibrated depression index for precise severity differentiation.
Depending on the identified level, the system can offer AI-assisted therapeutic
care for minimal or moderate cases, while severe cases are referred to professional
care. For voice-based emotion detection, wav2vec 2.0 was trained and reached
81% testing accuracy when classifying emotions into happy, sad, fearful, angry,
and neutral categories. Among the 10 participants diagnosed with depression, the
system correctly matched the actual severity level for 9 individuals, resulting in an
accuracy of 90%. Among these matched cases, a comparison based only on PHQ-9
severity revealed one mismatch: PHQ-9 indicated minimal severity, while both the
actual diagnosis and the system identified a moderate level. This demonstrates
that the proposed system provides a more accurate severity classification than
reliance on PHQ-9 scores alone. A comparison of Large Language Models showed
that GPT-3.5 and Mistral produced the most fluent outputs, with BERTScores of
71% and 69.2% respectively, outperforming Gemini at 61% and Claude at 59.8%.
Among them, GPT-3.5 was identified as the most suitable model to integrate into
the system due to its overall performance.
