Editing Ai Mental Health Dx (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''Depression screening NLP from clinical notes:'''
<syntaxhighlight lang="python">
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F
from openai import OpenAI

# MentalBERT — BERT pre-trained on mental health text (Reddit MH forums + clinical notes)
tokenizer = AutoTokenizer.from_pretrained("mental/mental-bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained(
    "mental/mental-bert-base-uncased"
)

def screen_for_depression(text: str) -> dict:
    """Screen text for depression severity using MentalBERT."""
    inputs = tokenizer(text, return_tensors="pt", max_length=512,
                       truncation=True, padding=True)
    with torch.no_grad():
        logits = model(**inputs).logits
    probs = F.softmax(logits, dim=1)[0]
    labels = model.config.id2label
    return {labels[i]: float(p) for i, p in enumerate(probs)}

# LLM-based PHQ-9 extraction from clinical notes
client = OpenAI()
def extract_phq9_from_note(clinical_note: str) -> dict:
    """Extract PHQ-9 items from unstructured clinical note."""
    prompt = f"""Extract PHQ-9 depression screening scores from this clinical note.
For each PHQ-9 item (1-9), score 0-3 if mentioned, or 'not mentioned'.
Return JSON with: {{phq9_items: {{...}}, total_if_complete: int|null, flags: [...]}}

Clinical note:
{clinical_note}"""
    response = client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[{"role":"user","content":prompt}],
        response_format={"type":"json_object"}, temperature=0
    )
    return response.choices[0].message.content

# Suicidality risk keyword escalation (rule + ML hybrid — safety-critical)
HIGH_RISK_PHRASES = [
    "suicidal ideation", "wants to die", "plan to kill", "suicide attempt",
    "self-harm", "cutting", "overdose intentional", "no reason to live"
]
def detect_crisis_risk(text: str) -> dict:
    """Rule-based + ML hybrid crisis detection — conservative (high sensitivity)."""
    text_lower = text.lower()
    rule_triggered = any(phrase in text_lower for phrase in HIGH_RISK_PHRASES)
    ml_result = screen_for_depression(text)
    return {
        'rule_triggered': rule_triggered,
        'ml_risk_score': ml_result.get('high_risk', 0),
        'escalate': rule_triggered,  # Rules always take precedence for safety
        'action': 'immediate_clinical_review' if rule_triggered else 'standard_care'
    }
</syntaxhighlight>

; Mental health AI tools
: '''Chatbots / DTx''' → Woebot, Wysa, Mindstrong (passive sensing), Headspace for Work
: '''Clinical NLP''' → MentalBERT, BERT-MH; Columbia CSSRS NLP (academic)
: '''Speech analysis''' → Sonde Health (depression from voice), Ellipsis Health, Canary Speech
: '''Population screening''' → Crisis Text Line ML (text message triage), Safe Messaging AI
: '''Passive sensing research''' → StudentLife dataset (Dartmouth), TILES dataset
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">