Initial public release of code accompanying the manuscript: "Lightweight Multimodal AI Framework for Early Depression Screening in Young Adults."

This repository contains the official implementation accompanying the research manuscript: “Lightweight Multimodal AI Framework for Early Depression Screening in Young Adults.” Depression in young adults remains underdiagnosed due to social stigma, limited access to professional care, and reliance on subjective self-report measures. This work proposes a computationally efficient multimodal framework for early depression screening using conversational data. The system integrates: Textual representations extracted using a pre-trained DistilBERT model Acoustic features derived from Mel-Frequency Cepstral Coefficients (MFCCs) with statistical aggregation Feature-level fusion through concatenation A lightweight feed-forward neural classifier for efficient inference The model is evaluated on the DAIC-WOZ (Distress Analysis Interview Corpus) dataset using an internal stratified train–validation split. Results demonstrate stable validation performance under limited multimodal sample availability. Threshold-based calibration significantly improves recall, making the system suitable for recall-oriented early screening scenarios. The framework emphasizes: Computational efficiency Deployability on resource-constrained devices Privacy-conscious feature extraction Modular design for reproducibility ⚠️ Note: The DAIC-WOZ dataset is not included due to licensing restrictions. This release ensures reproducibility and long-term preservation of the implementation through Zenodo DOI assignment.