JMIR Publications

ABSTRACT

Background:

The increasing prevalence of mental health problems among populations prompts the necessity of a mental health self-monitoring platform for the general public as an intermediary to professional treatment, which enables self-diagnosis of mental disorders based on natural lifelog data from daily living environments.

Objective:

In this study, we proposed a novel approach to self-predicting depression states with their severity levels based on naturally collected emotional narrative text, referred to here as “sentiment description.”

Methods:

Unlike conventional text-based approaches, which are mainly lexicon-driven, we extracted holistic mental health clues from the contextual connotations embedded in sentiment descriptions. By integrating them with lexicon-driven features, we aimed to enhance the reliability and explainability of depression diagnosis, supported by rich evidence from various angles. In particular, the holistic mental health clues extracted are designed to have a direct link to the standard mental health indices for diagnosing depression, the level of depression-indicative sentiments associated with chosen aspects, and the personal psychological tendencies affecting the sentiment description-based self-diagnosis network (SD2Net). In SD2Net, 1) holistic clues are extracted based on a large linguistic foundation model, GPT, with optimally designed prompts, 2) the holistic clues and raw texts are transformed into the connotation-driven and lexicon-driven feature embeddings are fused into multi-level depression state prediction, together two learned embeddings are fused into multi-level depression state prediction, together with the regression of their severity-levels, based on a hierarchical decision-making framework. For the experiment, we used both a benchmark dataset and a dataset custom-collected from real subjects on a long-term basis. Notably, we annotated the custom-collected data by fusing multiple self-questionnaires and psychiatrists’ inputs to ensure the trustworthiness of ground truth labels.

Results:

The experiments demonstrate the state-of-the-art performance of the proposed approach for holistic clue-reinforced multi-level depression state prediction, achieving an accuracy of approximately 80% using both a custom-collected dataset and the DepSign benchmark dataset. The ablation study confirms the effectiveness of the proposed holistic clue model, which yields an average improvement of over 10% in accuracy compared to models without it.

Conclusions:

The proposed approach, SD2Net, shows the effectiveness of incorporating holistic mental health clues from sentiment descriptions for multi-level depression state prediction, highlighting its potential as a reliable self-monitoring tool for the general public.