Clinical methodology

Our approach combines advanced Large Language Models (LLMs) with established clinical frameworks to provide safe, consistent, and explainable triage support.

Hybrid Intelligence Model

MindBridge does not rely solely on generative AI for decision-making. We use a hybrid architecture:

1. Generative Layer

Uses LLMs to understand natural language, build rapport, and extract clinical entities (symptoms, duration, severity) from patient conversations.

2. Rules-Based Guardrails

Deterministic clinical rules map extracted entities to risk bands. This ensures that critical keywords (e.g., "suicide", "overdose") trigger immediate escalation regardless of the LLM's output.

Risk Stratification Framework

Risk BandScoreDefinitionTypical Action
Low0 - 25No immediate risk identified. General distress or mild symptoms.Routine appointment booking. Self-help resources.
Moderate26 - 50Significant distress, functional impairment, but no immediate safety risk.Priority appointment based on clinic availability.
High51 - 75Severe symptoms, potential safety concerns, or rapid deterioration.Urgent review per clinical protocols. Crisis plan activation.
Crisis76 - 100Imminent risk of harm to self or others.Immediate emergency services (000/911). Direct clinician alert.

Key Risk Factors

Our system continuously monitors the conversation for specific risk markers, including but not limited to:

  • Intensity of negative emotion (hopelessness, despair)
  • Presence of a specific plan or intent for self-harm
  • Access to means (medication, weapons)
  • History of past attempts
  • Lack of social support or protective factors
  • Substance use or intoxication

Known Limitations

  • Text-Only Analysis: The system cannot currently analyze tone of voice, body language, or silence, which are critical clinical cues.
  • Cultural Nuance: While we train on diverse datasets, idioms and cultural expressions of distress may occasionally be misinterpreted.
  • Complex Comorbidities: Highly complex cases with multiple overlapping physical and mental health conditions may require human review for accurate triage.