The Laryngoscope · 2026
AI-Assisted Video Monitoring for Tracheostomy-Dependent Infants: A Proof-Of-Concept Study.
СтатьяНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Cecola Colleen F, Settoon Christine, Buck Lauren S, Evans Adele K, Dunham Michael E
Аннотация
OBJECTIVE: To retrospectively evaluate Eyes-On, a patient-personalized AI-assisted video analysis system for frame-level detection of tracheostomy tube status and visual distress in tracheostomy-dependent infants.
METHODS: In an IRB-approved study, 25 tracheostomy-dependent infants aged ≤ 2 years underwent video recording during routine tracheostomy care and tube changes. Data were partitioned at the patient level (22 infants for model development; 3 withheld for staged evaluation). From edited clips, 10,000 frames were extracted and annotated by two blinded pediatric otolaryngologists. A YOLOv11 detector was trained to detect cannulation status, and a facial distress classifier was built using facial features and action unit signals. Generalized pretrained models were tested on held-out infants and then reevaluated after patient-specific calibration. Individualized thresholds were selected using decision curve analysis.
RESULTS: The pretrained cannulation detector achieved accuracy 0.736, sensitivity 0.806, specificity 0.667, mAP@50 0.645, and a 23.62% Not-Detected rate (n = 1200). After calibration, pooled evaluable-frame cannulation performance improved to accuracy 0.940, sensitivity 0.997, and specificity 0.874 (AUROC 0.962; AUPRC 0.918). The pretrained distress classifier achieved accuracy 0.825, sensitivity 0.877, specificity 0.770, and a 21.8% face-extraction failure rate. After calibration, evaluable-frame distress accuracy increased to 0.960 with sensitivity 0.974 and specificity 0.946 (AUROC 0.993; AUPRC 0.993).
CONCLUSION: In this retrospective proof-of-concept study, patient-personalized video-based AI showed promising frame-level classification of tracheostomy-status and facial distress. These results reflect calibrated within-patient deployment and do not validate event-level outcomes, alarm thresholds, or standard monitoring modalities. Prospective real-world evaluation is needed before clinical adoption.
METHODS: In an IRB-approved study, 25 tracheostomy-dependent infants aged ≤ 2 years underwent video recording during routine tracheostomy care and tube changes. Data were partitioned at the patient level (22 infants for model development; 3 withheld for staged evaluation). From edited clips, 10,000 frames were extracted and annotated by two blinded pediatric otolaryngologists. A YOLOv11 detector was trained to detect cannulation status, and a facial distress classifier was built using facial features and action unit signals. Generalized pretrained models were tested on held-out infants and then reevaluated after patient-specific calibration. Individualized thresholds were selected using decision curve analysis.
RESULTS: The pretrained cannulation detector achieved accuracy 0.736, sensitivity 0.806, specificity 0.667, mAP@50 0.645, and a 23.62% Not-Detected rate (n = 1200). After calibration, pooled evaluable-frame cannulation performance improved to accuracy 0.940, sensitivity 0.997, and specificity 0.874 (AUROC 0.962; AUPRC 0.918). The pretrained distress classifier achieved accuracy 0.825, sensitivity 0.877, specificity 0.770, and a 21.8% face-extraction failure rate. After calibration, evaluable-frame distress accuracy increased to 0.960 with sensitivity 0.974 and specificity 0.946 (AUROC 0.993; AUPRC 0.993).
CONCLUSION: In this retrospective proof-of-concept study, patient-personalized video-based AI showed promising frame-level classification of tracheostomy-status and facial distress. These results reflect calibrated within-patient deployment and do not validate event-level outcomes, alarm thresholds, or standard monitoring modalities. Prospective real-world evaluation is needed before clinical adoption.