A transparent four-feature speech model for depression screening applicable across clinical and community settings, including assisted-living environments

TL;DR

This paper introduces a simple AI model that detects depression using speech, suitable for use in various settings including assisted-living environments.

Contribution

A transparent, four-feature AI model for depression screening that is lightweight and applicable across diverse populations.

Findings

The model achieved 92% sensitivity and moderate discriminative performance (AUC = 0.760) using four linguistic features from speech.

A semantic feature (emb_1) captures emotional or cognitive tension not directly expressed through negative words.

The model outperforms more complex models in the literature while being suitable for real-time, on-device use.

Abstract

Depression in older adults, often underrecognized and frequently conflated with cognitive symptoms, remains a major challenge in settings such as assisted-living communities. However, the need for scalable, speech-based screening tools extends across diverse populations and is not restricted to older adults or residential care. Depression in older adults is both common and frequently underdiagnosed, and while assisted-living environments represent a high-need deployment context, the present model is population-agnostic and can be validated across multiple real-world settings. Depression often co-occurs with mild cognitive impairment, creating a complex and vulnerable clinical landscape. Despite this urgency, scalable, interpretable, and easy-to-administer tools for early screening remain scarce. In this study, we introduce a transparent and lightweight AI-driven screening model that uses only four linguistic features extracted from brief conversational speech to detect depression with high sensitivity. Trained on the DAIC-WOZ dataset and optimized for deployment in resource-constrained settings, our model achieved moderate discriminative performance (AUC = 0.760) with a clinically calibrated sensitivity of 92%. Beyond raw accuracy, the model offers insights into how affective language, syntactic complexity, and latent semantic content relate to psychological states. Notably, one semantic feature derived from transformer embeddings, emb_1, appears to capture deeper emotional or cognitive tension not directly expressed through lexical negativity. Although the dataset does not contain explicit cognitive-status labels, these findings motivate future research to test whether similar semantic patterns may overlap with linguistic indicators of cognitive-affective strain observed in prior work. Our approach outperforms many more complex models in the literature, yet remains simple enough for real-time, on-device use, marking a step forward in making mental health AI both interpretable and clinically actionable. The resulting framework is population-agnostic and can be validated in assisted-living environments as one of several high-need deployment settings.

Figures