# Improving the Generalizability of Depression Detection by Leveraging Clinical Questionnaires

**Authors:** Thong Nguyen, Andrew Yates, Ayah Zirikly, Bart Desmet, Arman Cohan  
**arXiv:** [2204.10432](https://arxiv.org/abs/2204.10432)  
**Date:** 2022-04-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2204.10432v1/

## Key points

- Automated methods are used to identify and analyze mental health conditions from various sources of information, including social media.
- Challenges in deploying these models in real-world healthcare applications include poor out-of-domain generalization and a lack of trust in black box models.
- A team of researchers proposed an approach for depression detection that leverages the presence of symptoms described in PHQ9, a questionnaire used by clinicians during the depression screening process.
- The model was tested on three social media datasets using dataset-transfer experiments and compared to a standard BERT-based approach.
- Grounding the model in PHQ9's symptoms substantially improved its ability to generalize to out-of-distribution data while still performing competitively on in-domain data.
- Qualitative analyses suggested that this approach produced a model that was easier to inspect and more clinically relevant.
- These findings have important implications for improving the generalizability and trustworthiness of automated methods for detecting mental health conditions in real-world healthcare settings.

## Summary

The use of automated methods to identify and analyze mental health conditions, such as depression, from various sources of information, including social media, has become increasingly popular. However, deploying these models in real-world healthcare applications presents challenges such as poor out-of-domain generalization and a lack of trust in black box models. In response to these challenges, a team of researchers led by Thong Nguyen proposed an approach for depression detection that leverages the presence of symptoms described in PHQ9, a questionnaire used by clinicians during the depression screening process. The model was tested on three social media datasets using dataset-transfer experiments and compared to a standard BERT-based approach. The results showed that grounding the model in PHQ9's symptoms substantially improved its ability to generalize to out-of-distribution data while still performing competitively on in-domain data. Additionally, qualitative analyses suggested that this approach produced a model that was easier to inspect and more clinically relevant. These findings have important implications for improving the generalizability and trustworthiness of automated methods for detecting mental health conditions in real-world healthcare settings.