LUNA is an AI-powered assistant designed to help mental health clinicians rapidly synthesize insights from unstructured and structured patient data, such as therapy session transcripts, clinical notes, and standardized assessments (e.g., GAD-7, PHQ-9). It uses a RAG (Retrieval-Augmented Generation) model to retrieve data, classify it, and respond to targeted clinician prompts with visualized, explainable insights.

The Challenge

Clinicians often face:

• Limited time to review large volumes of session content

• Difficulty identifying patterns buried in qualitative data

• Gaps in documentation that slow down decision-making

LUNA solves this by surfacing what’s known, what’s missing, and what matters—automatically.

How LUNA Works

1. Data Input. Clinicians upload patient artifacts:

• Session transcripts (TXT, JSON)

• Standard assessments (e.g., phq9.json, gad7.json)

• Notes (SOAP format, free text)

2. AI Scanning and Category Scoring. Upon upload, LUNA uses NLP and embeddings to evaluate content across 5 domains:

• Emotional Functioning

• Cognitive Profile

• Behavioral Patterns

• Interpersonal Functioning

• Somatic Health

It calculates an “understanding score” (0–100%) for each based on keyword detection, emotional tone, document richness, and source diversity.

Sample Output:

{

"patient_id": "patient-0456",

"categories": {

"emotional": { "score": 85, "last_updated": "2025-05-12" },

"behavioral": { "score": 42 },

"cognitive": { "score": 72 },

"interpersonal": { "score": 60 },

"somatic": { "score": 0 }

}

3. Dynamic Prompt Display: Each category shows one of three states:

• No data → Show guidance prompts:

• “What questions should I ask to detect somatic stress?”

• Partial data → Show exploration prompts:

• “What avoidance behaviors have emerged, and what might be missing?”

• Sufficient data → Show analysis prompts:

• “Summarize emotional tone shifts across the last 5 sessions.”

Example Prompt + Insight Flow

Clinician Prompt: “Summarize how anxiety and low mood have evolved across sessions.”

LUNA Response:

• Summary: “Anxiety has decreased slightly over the past 3 sessions (GAD-7: 16 → 12). Mood has improved with less frequent self-critical language.”

• Visual: Line chart of GAD-7/PHQ-9 trend

• Verbatim: “I slept through the night for the first time in weeks.” – Session 4

• Reasoning: “This suggests that behavioral activation and grounding exercises are having an impact.”

• Source Reference: Session 4 transcript, timestamp 00:17:42

• Next Prompts:

• “What events correlate with emotional improvement?”

• “Are there emotional states the patient still avoids discussing?”

• “Compare this profile to similar GAD+MDD cases.”

Frontend & Interaction Design

• Smart prompt panel with auto-filtered categories

• Insight cards with expandable visual and source links

• Follow-up prompts appear contextually at the bottom of each insight

• Real-time refresh when new data is uploaded

Actionable AI Recommendations & Task Feedback Loop

LUNA AI doesn’t just generate insights—it also recommends actions that may require clinician follow-up. To support accountability and closed-loop decision-making, LUNA integrates a recommendation tracking system that links actions, reminders, and AI prompts.

How It Works

1. AI Suggests a Follow-Up Action. When generating a recommendation, the AI can label it as actionable, meaning it expects the clinician to do something (e.g., assign a journaling exercise, run a GAD-7 reassessment, discuss boundaries).

Example Insight Response JSON:

{

"insight_id": "insight-3024",

"category": "emotional",

"summary": "Patient shows persistent low mood with emerging emotional flattening.",

"recommendation": {

"type": "action_required",

"description": "Introduce emotion labeling exercise and track engagement over next 2 sessions.",

"action_id": "action-7291",

"due_by": "2025-05-19",

"status": "pending"

},

"follow_up_prompts": [

"Has the patient responded to the emotion labeling task?",

"Is there a shift in tone or range since the intervention?",

"Compare emotional expressiveness before and after task assignment."

]

}

2. Action Passes to Backend. When an action is recommended:

The AI submits the action_id, description, and due_by to the backend task tracker.

Status is logged as pending and linked to the patient profile.

3. Notification System. Frontend monitors task status and:

Displays reminders (e.g., “Follow-up task: Check emotional response to journaling – due today”)
Can optionally alert the user via in-app badge or notification drawer

4. Completion Options. LUNA can mark a task complete in two ways:

Manual Trigger: Clinician prompts the AI with a follow-up prompt (e.g., “Has the patient completed the labeling task?”). LUNA matches patterns or updates the status via inference.

Automatic Trigger: The system detects that relevant data (e.g., new transcript + evidence of engagement) has been added. LUNA parses the new input and updates the status if criteria are met.

Backend task update:

{

"action_id": "action-7291",

"status": "completed",

"completed_on": "2025-05-18T16:45:00Z",

"verified_by": "transcript:session-6:00:12:34"

}

Benefits

Supports actionable recommendations without losing context
Promotes follow-through and makes clinical AI assistive, not passive
Encourages reflection and re-prompting based on real patient progress

Impact

• Reduces time to insight for therapists by 60–80%

• Promotes consistency in assessment and tracking

• Enhances safety and traceability with source-linked AI responses

• Allows new clinicians to onboard quickly to complex cases

LUNA Clinical Insight Engine