MINORIA RESEARCH

MINORIA’s approach balances technical efficiency with clinical rigor, enabling SLMs to augment—not replace—mental health professionals while addressing resource disparities in PTSD care.

Data Curation

Diverse Trauma Narratives

Create synthetic dialogues covering PTSD symptoms (e.g., flashbacks, avoidance) and cultural contexts using LLM-generated personas 2.
Example: TIDE dataset’s 500 personas with varied trauma histories (combat, assault, disasters) and demographics 2.

Clinical Validation

Psychologists review scenarios for emotional plausibility and trauma sensitivity (e.g., avoiding retraumatizing language) 3.
Annotate linguistic markers of PTSD (disfluencies, avoidance semantics) from clinical interview transcripts 3.

Model Architecture Selection

Teacher-Student Framework

Use large models (e.g., Claude Sonnet 3.5) to generate empathetic responses via Diverse Response Inpainting, creating training data for SLMs 1.
Optimize for compact architectures (0.5B–5B parameters) to enable on-device processing and reduce latency 2.

Domain-Specific Embeddings

Fine-tune embeddings on trauma narratives (e.g., LLaMA embeddings + neural networks achieve 0.700 F1 in PTSD detection) 3.

Training & Fine-Tuning

Empathy-Centric Objectives

Train models to excel in three empathy factors:

Emotion recognition: Identify distress cues (e.g., “I can’t sleep because of the nightmares”) 2.
Distress normalization: Validate experiences without pathologizing (e.g., “Your reaction makes sense given what happened”) 2.
Supportive reflection: Suggest clinically aligned strategies (e.g., grounding techniques) 1.
Bias Mitigation : Implement debiasing protocols using adversarial training to reduce misdiagnosis risks in minority populations 3.

Validation & Evaluation

Clinical Metrics

Assess against DSM-5 criteria using tools like the DAIC-WOZ dataset (domain-specific models achieve 0.754 AUC) 3.
Human evaluations rate responses on empathy, safety, and cultural relevance (e.g., 29% referral increase in minority trials) 2.

Demographic Sensitivity Testing

Test performance across age, education, and trauma type subgroups (e.g., older adults prioritize distress validation) 2.

Deployment & Monitoring

Ethical Guardrails

Implement transparent decision logs for clinician review and verify treatment recommendations 2.
Use federated learning to update models without compromising patient privacy 2.

Continuous Learning

Incorporate clinician feedback loops to refine responses (e.g., adaptive learning from therapist corrections) 3.