💬 NLP

Assessing Algorithmic Bias in Language-Based Depression Detection: A Comparison of DNN and LLM Approaches

This study compares Deep Neural Networks and Large Language Models for detecting depression in clinical interviews, finding that while LLMs generally outperform DNNs and exhibit reduced gender bias, both approaches still struggle with racial disparities, with specific fairness-aware techniques like worst-group loss for DNNs and ethical prompting for LLMs offering partial mitigation.

Original authors: Obed Junias, Prajakta Kini, Theodora Chaspari

Published 2026-01-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Obed Junias, Prajakta Kini, Theodora Chaspari

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have two different types of detectives trying to solve a mystery: "Is this person feeling depressed?" based only on what they say during an interview.

Detective A (The DNN) is like a highly trained, specialized junior officer. They have studied thousands of mental health books and know the specific "rules" of language that often signal sadness.
Detective B (The LLM) is like a super-smart, general-purpose genius who has read almost everything on the internet. They haven't been specifically trained just for this one case, but they are incredibly good at understanding nuance, tone, and context.

The researchers in this paper wanted to see which detective is better at solving the case, and more importantly, does either detective treat people fairly? Do they make mistakes more often for certain groups of people (like men vs. women, or different racial backgrounds) than others?

Here is the breakdown of their investigation using simple analogies:

1. The Setup: The Interview Room

The detectives were given transcripts from the DAIC-WOZ, a collection of real clinical interviews where people talked about their lives. The researchers looked at two main "suspects" for bias:

Gender: Men vs. Women.
Race/Ethnicity: White/Caucasian vs. African American vs. Hispanic.

2. The Showdown: Who is Better?

The Result: The LLM (Detective B) was generally better at spotting depression than the DNN (Detective A).

The Analogy: Think of the DNN as a student who memorized a textbook. They know the definitions but might miss the "vibe" of a conversation. The LLM is like a seasoned counselor who can read between the lines.
The Catch: Even though the LLM was smarter overall, it still had blind spots. It was particularly good at helping the Hispanic group (who the DNN struggled with almost completely), but it still struggled to treat African American and White participants exactly the same way.

3. Fixing the Biases: The "Training"

The researchers tried to "fix" the detectives to make them fairer.

Fixing Detective A (The DNN)

Since the DNN is a machine learning model, you can change its "math" to force it to be fair. They tried two methods:

Method 1: "The Worst-Group Rule" (Worst-Group Loss). Imagine a teacher who says, "I don't care how well you do for the easy students; I only care if you can pass the hardest student." The model was forced to focus on the group it was failing the most.
- Result: This worked well! It balanced the scores without ruining the detective's overall accuracy.
Method 2: "The Average Rule" (Fairness-Regularized Loss). This method tried to make the scores equal for everyone at the same time.
- Result: This made the detective too confused. It tried so hard to be fair to everyone that it actually got worse at detecting depression overall.

Fixing Detective B (The LLM)

You can't retrain the LLM easily, so the researchers tried Prompting (giving the detective a specific set of instructions before they start).

The "Ethical Framing" Prompt: They gave the LLM a note that said, "Hey, treat everyone equally regardless of their gender or race. Don't use stereotypes. Just look at the words."
- Result: This helped a lot with gender bias. When the LLM was given just one example (1-shot) and this ethical note, it treated men and women much more fairly.
- The Limit: Giving the LLM more examples (3-shot or 5-shot) didn't help fix the bias any further. Also, this "Ethical Note" didn't really fix the bias regarding race. The LLM still had trouble treating different racial groups equally, no matter how many instructions or examples were given.

4. The Cost of Justice

There was a trade-off in time and money.

Detective A (DNN) was incredibly fast, solving a case in 0.002 seconds.
Detective B (LLM) took about 0.68 seconds per case.
The Takeaway: While the LLM was slower and more "expensive" to run, it was generally more accurate and easier to fix for gender bias than the faster, cheaper DNN.

Summary of Findings

The LLM is the stronger detective overall, especially for groups that the DNN completely missed (like Hispanic participants).
Bias is stubborn. Even the smartest AI (LLM) still struggles to treat different racial groups exactly the same, even when you give it ethical instructions.
The "Worst-Group" math trick was the best way to fix the DNN's bias without breaking its accuracy.
Ethical instructions helped the LLM treat men and women fairly, but only when the detective was looking at just one example at a time.

In short: The paper shows that while AI is getting better at detecting depression, we still have a long way to go to make sure it doesn't accidentally discriminate against people based on who they are. The "fixes" we try depend heavily on whether we are using a specialized model or a general-purpose giant.

Technical Summary: Assessing Algorithmic Bias in Language-Based Depression Detection

Problem Statement
While Natural Language Processing (NLP) offers promising avenues for automated depression detection through linguistic cues, significant concerns remain regarding algorithmic bias across socio-demographic groups. Existing literature indicates that linguistic markers of depression vary by gender, race, and ethnicity, potentially leading to disparate performance in automated systems. Although prior work has addressed bias in audio-based or multimodal depression detection, there is a notable gap in evaluating text-only models trained on clinical interview transcripts. Furthermore, the impact of emerging Large Language Models (LLMs) on these disparities, and the efficacy of bias mitigation strategies within clinical NLP contexts, remains underexplored. This study investigates these issues by comparing Deep Neural Network (DNN) embeddings with few-shot LLM approaches on the Distress Analysis Interview Corpus/Wizard-of-Oz (DAIC-WOZ) dataset, focusing on gender and race/ethnicity as primary sources of bias.

Methodology
The study utilizes the DAIC-WOZ dataset, comprising clinical interviews scored by the Patient Health Questionnaire-8 (PHQ-8). After excluding participants with missing demographic data, the analysis covers 186 participants for gender and 184 for race/ethnicity.

Model Architectures:
- DNN-Based Approach: The study employs Mental-RoBERTa, a domain-adapted model pre-trained on mental health corpora. To handle long transcripts, a chunking strategy (300-word windows with 80-word overlap) was used, with final predictions determined by majority voting.
- LLM Approach: GPT-4o-mini was utilized for few-shot learning. The study compared Unguided Prompting (UP) against Guided Prompting (GP), which includes specific linguistic cues for depression, and Guided Prompting with Ethical Framing (GP-EF), which adds explicit instructions to avoid stereotypes and ensure equitable treatment. Experiments were conducted in 1-shot, 3-shot, and 5-shot settings.
Bias Mitigation Strategies:
- For DNNs: Two fairness-aware loss functions were implemented:
  - Worst-Group Loss: Penalizes the model based on the subgroup with the highest cross-entropy loss to improve performance on the worst-performing group.
  - Fairness-Regularized Loss: Adds a penalty term proportional to the maximum disparity in subgroup-specific losses.
- For LLMs: Bias mitigation was explored solely through in-context learning via prompt engineering (GP and GP-EF) and varying shot counts.
Evaluation Metrics:
Performance was measured using Balanced Accuracy (BA), True Positive Rate (TPR), Precision, and F1 score. Fairness was quantified using Equalized Odds (EO), defined as $1 - |TPR_{non-sensitive} - TPR_{sensitive}|$ . Statistical significance was assessed via t-tests and ANOVA.

Key Contributions
This paper contributes to the field by:

Providing a comparative analysis of DNN-based embeddings and LLMs for depression detection, evaluating both performance and algorithmic fairness.
Examining fairness-aware modeling constraints (loss functions) for text-based depression detection, a task previously dominated by research on tabular or multimodal data.
Investigating in-context learning techniques with varied prompt framings to mitigate bias in LLMs, an approach not previously studied for depression detection.

Results

Performance Comparison: LLMs generally outperformed DNN-based embeddings in depression classification, particularly for underrepresented groups such as Hispanic participants.
Gender Bias:
- DNNs: Showed significant gender disparities (EO = 0.85) before mitigation. The Worst-Group Loss achieved a better balance between performance and fairness compared to the Fairness-Regularized Loss, which improved fairness but at the cost of reduced accuracy.
- LLMs: Exhibited reduced gender bias compared to DNNs (EO = 0.93 in 1-shot). The GP-EF strategy further improved EO in the 1-shot setting, making gender differences in TPR statistically insignificant. However, increasing the number of shots (3 and 5) did not consistently reduce disparities further.
Race/Ethnicity Bias:
- DNNs: Significant disparities existed, with Hispanic participants performing near chance levels before debiasing. Both loss functions improved performance for Hispanic and African American groups, with Worst-Group Loss offering a superior balance.
- LLMs: While LLMs improved overall detection for Hispanic participants compared to DNNs, racial disparities persisted. Neither prompt framing (GP, GP-EF) nor increasing the number of shots effectively reduced disparities between African American/Hispanic and White/Caucasian groups.
Efficiency: DNN-based embeddings were significantly faster (0.0025s per sample) compared to LLMs (0.68s per sample), though LLMs demonstrated superior performance and fairness outcomes.

Significance and Claims
The paper concludes that while LLMs offer superior performance and generally reduced gender bias compared to traditional DNN embeddings, socio-demographic disparities—particularly across racial groups—persist in both model types. The study highlights that bias mitigation strategies are not one-size-fits-all; Worst-Group Loss proved effective for DNNs, while Ethical Framing in prompts showed promise for LLMs in specific settings (1-shot gender analysis).

The authors modestly claim that these findings underscore the complexity of balancing fairness and accuracy in clinical NLP. They emphasize the necessity of tailoring debiasing strategies to specific model types and demographic contexts. The study acknowledges limitations, including the underrepresentation of Hispanic participants with depression in the dataset and the evaluation of only two representative models per category, suggesting that future work should explore broader model ranges and multilingual datasets to validate generalizability.