InterMind: Doctor-Patient-Family Interactive Depression Assessment Empowered by Large Language Models

This paper presents InterMind, a large language model-powered system that enhances depression assessment by integrating multi-role interactions among doctors, patients, and families while employing retrieval-augmented generation and chain-of-thought techniques to improve diagnostic accuracy, interpretability, and reduce hallucinations.

The Gist

Imagine you are trying to solve a complex puzzle, but you only have one piece of the picture. That is often how depression is diagnosed today: a doctor asks a patient questions, and the patient tries to describe how they feel. But sometimes, people hide their pain, forget details, or simply don't have the words to explain what's wrong.

InterMind is a new digital tool designed to solve this puzzle by bringing in more pieces of the picture. It's like upgrading from a solo detective to a full investigative team.

Here is a simple breakdown of how it works, using some everyday analogies:

1. The Problem: The "Blind Spot"

Currently, most depression check-ups are a two-person conversation (Doctor + Patient).

• The Issue: Patients might be too shy to admit they are struggling, or they might forget to mention that they haven't slept in three days.
• The Missing Piece: Family members often see things the patient doesn't (like a sudden change in appetite or mood swings) but aren't usually part of the official medical interview.

2. The Solution: The "Three-Way Team"

InterMind changes the game by creating a Doctor-Patient-Family triangle. It uses a special type of AI (called a Large Language Model) to act as a bridge between these three groups.

Think of the system as having two distinct AI characters:

Character A: The "Empathetic Chatbot" (The Listener)

• What it does: This AI acts like a friendly, non-judgmental counselor. It chats with the patient to hear their side of the story. Crucially, it also chats with the family members to hear what they observe.
• The Analogy: Imagine a translator who speaks both "Patient" and "Family." It gathers the patient's internal feelings and the family's external observations, weaving them together into one complete story so nothing gets lost in translation.
• How it learned: Since real therapy sessions are private, the researchers taught this AI by taking real stories people shared on social media and rewriting them into perfect, supportive conversations. This gave the AI a massive library of "how to listen" without needing to invade real people's privacy.

Character B: The "AI Psychiatrist" (The Analyst)

• What it does: Once the Chatbot has gathered all the stories, this second AI steps in. It doesn't just say "Depressed" or "Not Depressed." Instead, it acts like a detective with a rulebook.
• The Analogy: Think of the medical rulebook (DSM-V) as a massive library of thousands of pages. A human doctor can't read the whole library in 5 minutes. The AI, however, has a super-fast librarian (called RAG) that instantly finds the exact rules that match the symptoms the patient described.
• The "Chain of Thought": Instead of guessing, the AI is forced to "show its work." It's like a student solving a math problem on a whiteboard:
  1. Step 1: "The patient said they cry often."
  2. Step 2: "The family said they stopped eating."
  3. Step 3: "The rulebook says these two things together match 'Moderate Depression'."
  4. Step 4: "Therefore, the report should suggest X treatment."
     This step-by-step thinking prevents the AI from making up facts (hallucinations) and makes the diagnosis explainable.

3. The Result: A Better Report

Instead of a simple "Yes/No" answer, the system generates a comprehensive report for the human doctor.

• It includes the severity of the depression.
• It lists specific symptoms matched to medical rules.
• It offers tailored advice for both the patient and the family on how to care for each other.

Why This Matters

• For the Patient: They feel heard, and their family gets involved, making the diagnosis more accurate.
• For the Doctor: They get a "pre-drafted" report that highlights the most important facts, saving them time and reducing the chance of missing a clue.
• For the System: It turns a subjective, guess-heavy process into a structured, evidence-based one.

The Catch (Limitations)

The authors are honest that this is still a prototype. Right now, the "Family" part of the conversation is simulated by the AI based on existing data, not real-time family members. It's like a flight simulator: it's incredibly good at training and testing, but it still needs real pilots (doctors) and real passengers (patients) to fly the actual plane in the real world.

In short: InterMind is a smart assistant that helps doctors see the whole picture of a patient's mental health by listening to both the patient and their family, using a "rulebook" to ensure the diagnosis is accurate and easy to understand.

Technical Summary

Here is a detailed technical summary of the paper "InterMind: Doctor-Patient-Family Interactive Depression Assessment Empowered by Large Language Models."

1. Problem Statement

Depression affects over 300 million people globally, yet current assessment methods face three critical limitations:

• Single-Role Bias: Existing paradigms rely heavily on patient self-reports or patient-doctor dyads. Patients often conceal symptoms or lack the ability to articulate feelings accurately, while family members (who observe daily behaviors) are excluded from the evaluation loop.
• Lack of Interpretability: Current Automatic Depression Detection (ADD) methods typically treat the problem as a binary classification or regression task. They output a label (e.g., "depressed") without providing the reasoning, symptom analysis, or alignment with clinical standards, making them unusable for assistive diagnosis.
• LLM Limitations: While Large Language Models (LLMs) show promise, they suffer from hallucinations, lack of domain-specific expertise (e.g., DSM-V criteria), and an inability to generate structured, professional diagnostic reports without specific guidance.

2. Methodology: The InterMind System

The authors propose InterMind, a system integrating two specialized LLM components: an AI Psychological Chatbot and an AI Psychiatrist. The system facilitates interaction between doctors, patients, and families.

A. AI Psychological Chatbot (Data Generation & Counseling)

To address the scarcity of high-quality psychological counseling data:

• Prompt Engineering for Data Augmentation: The authors developed a prompt engineering strategy to rewrite social media posts (rich in authentic emotional experiences) into multi-turn psychological counseling dialogues.
• Dataset Construction: These generated dialogues ($D_1$) are combined with existing clinical interview data ($D_2$) to create a comprehensive dataset $D$.
• Instruction Fine-Tuning: The LLM is fine-tuned on this dataset to simulate an empathetic counselor capable of engaging patients and family members to gather multi-perspective information (e.g., mood swings, insomnia).

B. AI Psychiatrist (Diagnostic Report Generation)

This component transforms collected dialogues into standardized assistive diagnostic reports. It employs three key techniques to ensure accuracy and interpretability:

1. Role-Playing Dialogue Generation: Symptoms are split into overlapping subsets for the "patient" and "family" perspectives to simulate real-world redundancy and complementarity in symptom reporting.
2. Retrieval-Augmented Generation (RAG):
   • The system retrieves relevant sections from the DSM-V (Diagnostic and Statistical Manual of Mental Disorders) based on the semantic similarity between the user's dialogue and diagnostic criteria.
   • This mitigates hallucinations by grounding the LLM's output in authoritative medical standards.
3. Chain-of-Thought (CoT) Reasoning:
   • The LLM is guided through a four-step reasoning process: (1) Analyze emerging symptoms, (2) Match symptoms to retrieved DSM-V criteria, (3) Estimate severity, and (4) Predict the final diagnosis.
   • This ensures the output is a structured, logical argument rather than a raw classification.

C. System Output

The AI Psychiatrist generates a fixed-format JSON report containing:

• Binary classification (Depressed/Not Depressed).
• Severity degree estimation (Normal, Mild, Moderate, Severe).
• Symptom-to-standard alignment (mapping specific symptoms to DSM-V criteria).
• Subtype categorization.
• Personalized treatment strategies and care advice for both patients and families.

3. Key Contributions

1. Doctor-Patient-Family Paradigm: Introduced a multi-role interaction framework that leverages family observations to complement patient self-reports, reducing the risk of inaccurate self-reporting.
2. Enhanced LLM Capabilities:
   • Developed prompt engineering to synthesize counseling data from social media.
   • Integrated RAG and CoT to align LLM outputs with professional diagnostic standards (DSM-V), significantly reducing hallucinations and improving interpretability.
3. Comprehensive System Validation: Validated the system through both quantitative metrics (classification accuracy) and qualitative assessments by clinical psychiatrists.

4. Experimental Results

The system was evaluated on three datasets: MMDA (Chinese clinical interviews), DR (Reddit posts), and DepSeverity (Reddit posts with severity labels).

• Binary Classification:
  • The "Both" (Patient + Family) paradigm achieved the highest F1 scores across all datasets (e.g., 0.862 on MMDA, 0.890 on DR), outperforming patient-only, family-only, and baseline models (BiLSTM, RoBERTa, GPT-4).
  • The multi-role approach proved superior in capturing comprehensive information.
• Severity Estimation:
  • InterMind demonstrated superior performance in identifying Moderate and Severe depression compared to baselines, which often failed to detect these categories.
  • On the MMDA dataset, the system achieved a weighted F1 of 0.619, outperforming GPT-4 (0.593).
• Report Quality (Expert Evaluation):
  • Five clinical psychiatrists evaluated the reports. InterMind outperformed GPT-4 in Symptom-Standard agreement (0.966 vs. 0.954 on MMDA) and Severity estimation.
  • While GPT-4 was more conservative (high precision, low recall), InterMind aligned better with ground truth labels, aiding early detection.
• Ablation Study:
  • Adding RAG improved recall by grounding the model in DSM-V.
  • Adding CoT significantly boosted recall and F1 scores by enforcing logical reasoning steps.
  • The combination of RAG + CoT yielded the best results, confirming their necessity for professional report generation.
• Counseling Competency:
  • In subjective evaluations by 18 psychology-trained volunteers, InterMind outperformed GPT-4 in consistency, engagement, and overall performance, providing more natural and supportive interactions.

5. Significance and Limitations

Significance:

• Clinical Utility: InterMind moves beyond simple classification to provide interpretable, actionable diagnostic reports that assist doctors rather than replace them.
• Holistic Care: By integrating family perspectives, the system addresses the "blind spots" of patient self-reporting, leading to more objective assessments.
• Scalability: The use of RAG and CoT offers a blueprint for adapting LLMs to other specialized medical domains where hallucination control and adherence to guidelines are critical.

Limitations:

• Data Simulation: The family-perspective data was simulated by LLMs based on existing sources rather than collected from real-world family interactions, which may not fully capture real-world dynamics.
• Clinical Validation: While validated by psychiatrists in a study setting, the system requires further testing in live clinical environments with real doctor-patient-family triads to confirm long-term efficacy.