---
id: wiki-2026-0508-asd-intervention
title: ASD Intervention (AI-Assisted)
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [자폐 스펙트럼, ASD, autism, neurodiversity, social robot, AAC, emotion recognition, social skills training]
duplicate_of: none
source_trust_level: B
confidence_score: 0.83
verification_status: conceptual
tags: [accessibility, asd, autism, neurodiversity, ai-for-good, social-robot, aac, emotion-recognition, ethics]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: Python / Swift / TypeScript
  framework: Vision API / Speech / VR
---

# ASD Intervention (AI-Assisted)

## 📌 한 줄 통찰
> **"매 social barrier 의 digital companion"**. 매 ASD 의 communication / emotion 의 difficulty 의 AI 의 supplement. 매 NESCA / VR / robot / AAC. 매 supplement only — 매 human therapist 의 substitute X. 매 neurodiversity-affirming 이 새 paradigm.

## 📖 핵심

### 매 ASD 정의
- 매 DSM-5: 매 social communication + 매 restricted/repetitive behavior.
- 매 spectrum: 매 mild ↔ 매 severe.
- 매 1 in 36 (CDC 2023 US).
- 매 male:female 약 4:1 (under-diagnose 의 female).

### 매 핵심 challenge
1. **Communication**: 매 verbal / non-verbal 의 difficulty.
2. **Social cognition**: 매 ToM (theory of mind), 매 emotion read.
3. **Sensory**: 매 over/under-sensitivity.
4. **Routine**: 매 change 의 distress.
5. **Executive function**: 매 planning / flexibility.

### 매 evidence-based intervention
- **ABA** (Applied Behavior Analysis): 매 controversial.
- **DIR/Floortime**: 매 child-led play.
- **PECS** (Picture Exchange): 매 visual.
- **Speech / OT**: 매 standard.
- **Social skills group**.

→ 매 controversy: 매 ABA 의 normalization 의 critique (neurodiversity movement).

### 매 AI 의 응용

#### Emotion recognition (computer vision)
- 매 webcam / smart glass.
- 매 facial expression → text / audio cue.
- 매 Brain Power, 매 Empowered Brain.

#### Social skill training (VR)
- 매 safe rehearsal environment.
- 매 job interview / classroom / store.
- 매 Floreo, 매 BrainPOP (research-stage).

#### Robot companion
- **NAO**, **Kaspar**: 매 humanoid 의 인내 의.
- **Milo**, **Moxie**: 매 child-targeted.
- 매 emotion 의 consistent + 매 patient.

#### AAC (Augmentative & Alternative Communication)
- **Proloquo2Go**: 매 symbol-based.
- **TouchChat**: 매 communication board.
- 매 LLM 의 personalization.

#### Sensory regulation
- **Stimming-aware UI**: 매 minimize visual / audio overload.
- **Customizable**: 매 brightness / volume.
- **Predictability**: 매 visual schedule.

#### Behavioral analytics
- **Observe behavior pattern**.
- **Trigger detection** (anticipate meltdown).
- **Outcome tracking**.

### 매 ethical concern
1. **Substitute risk**: 매 human therapist 의 replace 의 X.
2. **Privacy**: 매 child data 의 sensitive.
3. **Bias**: 매 white male sample 의 train.
4. **Neurodiversity**: 매 cure framing 의 critique.
5. **Surveillance**: 매 always-on monitoring.
6. **Consent**: 매 child 의 capacity.
7. **Autonomy**: 매 user-driven > forced compliance.

### 매 Neurodiversity affirming
- 매 ASD = 매 difference, 매 disorder X (some view).
- 매 strength: 매 pattern, 매 detail, 매 honesty.
- 매 AI design: 매 accommodate, 매 normalize 의 X.
- 매 community input (autistic people 의 lead).

→ "Nothing about us without us."

## 💻 패턴

### Emotion recognition (CV API)
```python
from azure.cognitiveservices.vision.face import FaceClient

face_client = FaceClient(endpoint, credentials)

def detect_emotion(image):
    faces = face_client.face.detect_with_stream(
        image, return_face_attributes=['emotion'],
    )
    if not faces: return None
    
    emotions = faces[0].face_attributes.emotion
    top = max(emotions.__dict__.items(), key=lambda x: x[1])
    return top[0]  # 매 'happiness', 'sadness', 'anger', ...

# 매 caption 의 supportive (not invasive)
emotion = detect_emotion(camera_frame)
if emotion:
    show_subtle_caption(f'They might be feeling: {emotion}')
```

### AAC builder (LLM-augmented)
```python
def suggest_phrase(intent, context, recent_words=[]):
    prompt = f"""User wants to express: {intent}
Context: {context}
Recent words: {recent_words}

Suggest 4 short phrases (≤6 words each) the user could send.
Match their typical voice based on recent words."""
    
    return llm.generate(prompt).split('\n')[:4]

# 매 user 의 click 의 word → 매 prediction.
```

### Sensory-friendly UI
```tsx
// 매 settings 의 user-controllable
<Settings>
  <Toggle label="Reduce motion" value={reduceMotion} />
  <Toggle label="High contrast" value={highContrast} />
  <Slider label="Volume cap" min={0} max={100} value={volumeCap} />
  <Toggle label="Predictable schedule" value={predictableSchedule} />
  <Toggle label="Less notifications" value={lessNotif} />
</Settings>

// 매 ApplyAccessibility 의 propagate.
```

### Visual schedule (predictability)
```ts
type ScheduleItem = {
  time: string;
  activity: string;
  icon: string;
  duration_min: number;
};

function renderSchedule(items: ScheduleItem[]) {
  return (
    <div role="list">
      {items.map((item, i) => (
        <Card key={i}>
          <img src={item.icon} alt={item.activity} />
          <h3>{item.activity}</h3>
          <p>{item.time} ({item.duration_min} min)</p>
          {i === currentIndex && <Highlight>NOW</Highlight>}
        </Card>
      ))}
    </div>
  );
}
```

### Trigger detection (behavioral pattern)
```python
def detect_overload_risk(sensor_data, window=30):
    """매 heart rate + skin conductance + recent stim count → meltdown risk."""
    hr = sensor_data['heart_rate'][-window:]
    eda = sensor_data['eda'][-window:]
    stim_count = count_stims(sensor_data['accelerometer'][-window:])
    
    risk = (
        np.mean(hr) > BASELINE_HR + 20 and
        np.mean(eda) > BASELINE_EDA + 0.5 and
        stim_count > 5
    )
    
    if risk:
        suggest_break()
        notify_caregiver(consent_required=True)
    return risk
```

→ 매 child consent + caregiver consent + 매 invasive 의 X.

### Privacy-preserving local processing
```python
# 매 cloud upload X — 매 edge inference
import tflite_runtime.interpreter as tflite

interpreter = tflite.Interpreter(model_path='emotion_model.tflite')
# 매 raw frame 의 leave 의 X. 매 label 만 의 leave (with consent).
```

## 🤔 결정 기준
| 응용 | Approach |
|---|---|
| Emotion | CV + supportive caption |
| Social practice | VR safe environment |
| Companion | Robot (NAO, Moxie) — 보완 |
| Communication | AAC + LLM suggest |
| Sensory | Customizable + local |
| Behavioral | Edge ML + consent |
| Therapy | 매 therapist + 매 AI tool 의 supplement |

**기본값**: 매 user-driven + 매 consent + 매 local processing + 매 neurodiversity affirming.

## 🔗 Graph
- 부모: [[Accessibility (A11y)|Accessibility]] · [[AI-for-Good]]
- 변형: [[Emotion-Recognition]] · [[Social-Skills-Training]] · [[AAC]] · [[Social-Robot]]
- Adjacent: [[Anthropomorphism]]

## 🤖 LLM 활용
**언제**: 매 AAC supplement. 매 social practice prompt. 매 visual schedule generation. 매 sensory-friendly content.
**언제 X**: 매 diagnosis (의사). 매 therapy 의 substitute. 매 child 의 consent X 의 deployment.

## ❌ 안티패턴
- **Cure framing**: 매 normalization 의 push.
- **Substitute therapist**: 매 over-reliance on AI.
- **Invasive monitoring**: 매 always-on without consent.
- **Cloud-only**: 매 child data 의 leak.
- **Generic UI**: 매 sensory difference 의 ignore.
- **Forced compliance**: 매 ABA-style 의 control.
- **No autistic input**: 매 community 의 ignore.

## 🧪 검증 / 중복
- Verified (peer-reviewed ASD research, neurodiversity literature).
- 신뢰도 B.
- Related: [[Accessibility (A11y)|Accessibility]] · [[AI-for-Good]] · [[Humane-Tech]] · [[Anthropomorphism]].

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — intervention type + ethics + neurodiversity + 매 emotion recognition / AAC / sensory UI code |