Imbiriba — JAMA Network Open 2023 (AIC Extension)
Citation: Imbiriba, T., Demirkaya, A., Singh, A., et al. (2023). JAMA Network Open, 6(12), e2348898.
DOI: 10.1001/jamanetworkopen.2023.48898
Access: AIC study — contact authors (Imbiriba et al.), check SFARI Base
Tier: 1
Overview
Largest ASD aggression prediction study to date. Replicated and extended Goodwin 2019 across 4 inpatient sites. 70 participants (ages 5-19, 88.6% male), 429 sessions, 497 hours, 6665 aggressive behaviors.
Signals
Device: Empatica E4 wristband (non-dominant wrist) - EDA (electrodermal activity) - BVP (blood volume pulse → HR/HRV) - ACC (accelerometry) - TEMP (skin temperature)
Labels
- Self-injurious behavior (SIB) — 3983 events (59.8%)
- Emotion dysregulation (ED, including tantrums and meltdowns) — 2063 events (31.0%)
- Aggression toward others (ATO) — 619 events (9.3%)
Performance
- Best classifier: logistic regression
- AUROC: 0.80 predicting 3 minutes before onset (95% CI, 0.79-0.81)
- Compared: SVM, neural networks, domain adaptation
Key Findings
- Replicated Goodwin findings at larger scale
- First study to separately predict SIB, ED, and ATO from wearable data
- Logistic regression consistently best across all experiments
- Domain adaptation improved cross-subject generalization
Linked To
- Paper: imbiriba-2023
- Dataset: goodwin-aic (predecessor study)