MentalCare-AI: Mental Health AI Detector Using Explainable AI

Abstract

In the modern digital healthcare ecosystem, mental health support systems face persistent challenges related to diagnostic accuracy, algorithmic transparency, and user data privacy. Existing digital tools often rely on unimodal data inputs and opaque decision-making models, resulting in limited clinical trust and poor user engagement. This paper presents MentalCare-AI, a privacy-aware multimodal mental health companion designed to address these limitations through a principled integration of Natural Language Processing (NLP), speech-based acoustic analysis, and Explainable Artificial Intelligence (XAI). The proposed system combines contextual text embeddings generated by DistilBERT with acoustic representations extracted by wav2vec 2.0, fusing the two modalities through a cross-attention mechanism to produce holistic mental health risk assessments. User privacy is preserved through automated Personally Identifiable Information (PII) detection and removal, strict data minimization policies, and ephemeral audio processing. Model interpretability is achieved through SHAP (SHapley Additive exPlanations)-based visual explanations that identify the top contributing linguistic and paralinguistic features for each prediction. The proposed MentalCare-AI framework, trained and evaluated on the DAIC-WOZ dataset, achieves an F1-score of 0.87, representing a statistically significant improvement over unimodal text-only (F1: 0.72) and audio-only (F1: 0.70) baselines. These results demonstrate the viability of building AI-assisted mental health screening tools that are simultaneously accurate, explainable, and privacy-preserving, offering a responsible blueprint for next-generation digital mental healthcare.