Multimodal Personality-aware Depression Detection

Call for Participation

Depression is a pervasive mental health disorder with a profound impact on global wellbeing, affecting millions and significantly hindering quality of life, productivity, and overall health. Many datasets primarily focus on establishing direct correlations between behavioral signals and depression levels. This approach, while valuable, often fails to account for the nuanced individual variations in depressive symptoms. Moreover, the demographic scope of existing datasets is predominantly confined to young adults, thereby overlooking other critical age groups, such as the elderly, who may present distinct patterns of depression.

The Multimodal Personality-aware Depression Detection (MPDD) aims to address these gaps by integrating multiple data types (audio, video, text) with individual difference information such as personality traits, age, gender, and health conditions. This challenge seeks to develop models that can accurately detect depression across a broader demographic, fostering more inclusive and effective mental health detection systems.

The advent of specialized datasets has been instrumental in advancing depression research by providing multimodal data that can enhance the detection and understanding of depressive symptoms. For instance, the AVEC challenge has contributed rich audiovisual interview data annotated with depression-related information, facilitating more nuanced assessments. The DAIC-WoZ dataset has furthered our understanding by incorporating clinical interviews and PHQ-9 scores. Other datasets, including the Pitts burgh dataset, the D-Vlog dataset, the EATD-Corpus, the CMDC, and the MODMA, have expanded the scope by integrating diverse data types like EEG and utilizing standard depression scales for evaluation. Additionally, datasets such as AMI GOS and DEAP have incorporated physiological data, offering insights into the physiological underpinnings of depression, while the SEED dataset provides EEG and eye movement data annotated with personality traits, highlighting the role of individual differences in depressive symptoms.

To advance the development of personalized and accurate depression detection systems, we are launching the MPDD 2025: Multimodal Personality-aware Depression Detection. This challenge aims to push the boundaries of depression detection by integrating multimodal data (audio, video, text) with individual difference information such as personality traits, age, gender, and health conditions.

We have designed two tracks to address the unique challenges of depression detection in different populations. We warmly invite researchers from both academia and industry to participate and collaborate in finding innovative solutions to these important and real-world challenges.

Track Setting and Evaluation

Track Setting

Track 1: Elderly Depression Detection(MPDD-Elderly). focuses on detecting depression in the elderly population, leveraging data from participants with audiovisual interviews, personality labels based on the BigFive-10 scale, health condition labels, and detailed demographic information. This track aims to explore how factors such as chronic illnesses, living condi- tions, and personality traits influence depression manifestation in older adults.

Track 2: Young Adult Depression Detection(MPDD-Young). depression detection among young adults, a high-risk group often understudied. The dataset includes participants with audiovisual recordings, personality labels, and demographic information. This track seeks to understand how factors like geography environment, age, gender, and personality traits contribute to depression in young adults.

Evaluation Metrics

The MPDD Challenge evaluates models using Accuracy (Acc) and F1 score (F1). For each task, final Acc and F1 are computed as the average of their weighted and unweighted versions. Track-level scores are then obtained by averaging the results across all tasks. The Colab leaderboard follows this same scoring rule:

Acc_track = Average(Acc^task1 + Acc^task2 + …)

F_track = Average(F1^task1 + F1^task2 + …)

Score = (Acc_track + F_track) / 2

Dataset

As compared to the existing datasets, the MPDD dataset significantly enhances the breadth of contextual diversity and the depth of annotation comprehensiveness, as underscored in Table 1. By offering a more inclusive and granular depiction of depression, the MPDD dataset is poised to facilitate the evolution of sophisticated multimodal models. These models are better equipped to navigate the intricacies of mental health conditions, particularly depression. The dataset’s thorough annotation scheme is instrumental in discerning individual differences. This capability is essential for fostering timely interventions and conducting precise assessments of depression, thereby contributing to the broader objective of improved mental health management.

Track 1: MPDD-Elderly Dataset: The MPDD-Elderly dataset is a comprehensive resource tailored for the detection and analysis of depression among the elderly population. Comprising data from elderly participants, this dataset is designed to support a variety of classification and regression tasks related to depression identification. The annotations within the MPDD-Elderly dataset are meticulously crafted to capture a broad spectrum of factors influencing depression in the elderly. They are grounded in the PHQ-9 Scale Scores, which offer a reliable quantitative measure of depressive symptoms, allowing for standardized assessment across the participant cohort. Personality traits are evaluated using the BigFive-10 scale, providing insights into the individual differences that may contribute to the manifestation of depression. The dataset also incorporates comprehensive demographic information, encompassing age, disease, family situation, and economic status, which paints a detailed picture of each participant’s background and potential stressors. Furthermore, disease labels are included to account for various health conditions, such as those related to the endocrine, circulatory system, and nervous system, recognizing the impact of physical health on mental health. This multifaceted approach to annotation ensures a nuanced understanding of depression in the elderly, facilitating the development of more precise and personalized assessment tools.

The subdataset is structured to support both 3 types of classification tasks:

Binary classification: normal / depressed
Ternary classification: normal / mildly depressed / severely depressed
Quinary classification: normal / mildly depressed / moderately depressed / moderately severe / severe

Track 2: MPDD-Young Dataset: The MPDD-Young dataset is designed to advance the field of multimodal depression detection, particularly four Young Adult Depression Detection. Comprising data from young adults, this dataset offers a comprehensive view of depressive symptoms through various modalities and tasks. The annotations within the MPDD-Young dataset are multifaceted, providing a rich set of data for analysis. They are based on the PHQ-9 Scale Scores, which offer a quantitative measure of depressive symptoms, allowing for a standardized assessment of depression severity. Additionally, personality traits are evaluated using the BigFive-10 scale, which helps in understanding the individual differences that may influence the manifestation of depression. The dataset also includes demographic information such as gender, age, and birth region. This comprehensive profile of the participants enables a more nuanced analysis of depression, taking into account the diverse backgrounds and personal characteristics of the individuals involved.

The dataset is structured to support both 3 types of classification tasks:

Binary classification: normal / depressed
Ternary classification: normal / mildly depressed / severely depressed

For both tracks, the subject samples are divided into a training set, a validation set, and a test set. 80% of the data set is used for training and validation, while the remaining 20% is used for testing.

Rules

Access to Data Participants may freely use pre-trained weights obtained from open-source datasets or models to initialize their target models. This practice is allowed to facilitate faster development cycles and to encourage innovation in model architecture and algorithm design.
Use of Pre-trained Models: Participants may freely use pre-trained weights obtained from open-source datasets or models to initialize their target models. This practice is allowed to facilitate faster development cycles and to encourage innovation in model architecture and algorithm design.
Restrictions on Data Usage: To maintain the integrity of the challenge, participants are explicitly prohibited from using additional private or in-house datasets to pre-train their models. Compliance with this rule ensures that all participants are evaluated on an equal footing.
Submission Limitations: Participants are restricted to a maximum of five submission attempts per sub-challenge. Each submission will consist of the model’s predictions on the test partition, along with the corresponding model and checkpoints/weights.
Evaluation Metrics: The organizers will evaluate the submitted models using two predefined metrics: Accuracy and F1 Score. These metrics will be computed by comparing the participants’ predictions against the ground-truth labels.
Final Ranking: The final test results of each team will be used for the final ranking in the challenge. This ensures that the models are thoroughly tested and their performance is accurately assessed.
Reproducibility: All materials, including data, must be accessible to all participating teams to ensure that the models submitted by participants are fully reproducible.
Organizer’s Role: The organizers will not actively participate in the challenge but will undertake a re-evaluation of the findings from the best-performing systems in each sub-challenge to validate the results.
Eligibility: The challenge is open to any researcher from any organization based anywhere in the world, fostering a global collaborative environment in addressing the complex issue of depression detection.

Registration

The rankings of our challenge are based on the CodaLab Leaderboard, so entrants will need to register on Codalab using the GROUP NAME provided on the EULA or the email that sends the EULA in order for entrants to upload their results and view the rankings. We will provide the CodaLab link when the challenge is released.

To further safeguard the security and compliance of the data, please complete the following steps before contacting us to request access to the challenge:

1. Download the MPDD Dataset License Agreement PDF.

2. Carefully review the agreement: The agreement outlines in detail the usage specifications, restrictions, and the responsibilities and obligations of the licensee. Please read the document thoroughly to ensure complete understanding of the terms and conditions.

3. Manually sign the agreement: After confirming your full understanding and agreement with the terms, fill in the required fields and sign the agreement by hand as formal acknowledgment of your acceptance.

Once you have completed the above steps, please submit the required materials to us through the following channels: Primary contact email: sstcneu@163.com CC email: fuchangzeng@qhd.neu.edu.cn

User License Agreement Link (should be signed with a full-time faculty or researcher): Agreementt.pdf
(Applications not following the rule might be ignored)