Highlight
This proof-of-concept study introduces a patient-personalized AI-assisted video monitoring system, Eyes-On, with the capability to detect tracheostomy tube status and signs of visual distress in infants under two years of age. The AI system, after patient-specific calibration, significantly improved frame-level detection accuracy for both cannulation status and facial distress, thereby showing potential for bedside continuous monitoring and early detection of complications in tracheostomy-dependent infants.
Study Background
Tracheostomy-dependent infants represent a vulnerable pediatric population that requires meticulous airway management to prevent life-threatening complications such as accidental decannulation or respiratory distress. Continuous monitoring is critical in maintaining airway patency and promptly recognizing clinical distress; however, current monitoring techniques largely rely on intermittent clinical assessments or invasive alarms, which may miss subtle or early signs of compromise. Video-based surveillance augmented by artificial intelligence (AI) offers a non-invasive and potentially continuous monitoring solution. The Eyes-On system, developed to provide patient-personalized real-time frame-level detection of tracheostomy tube status and facial expressions indicative of distress, addresses this unmet clinical need. This study retrospectively evaluates its diagnostic performance and calibration requirements in a controlled setting.
Study Design
The study was an Institutional Review Board (IRB)-approved retrospective investigation involving 25 tracheostomy-dependent infants aged ≤2 years. Video recordings were obtained during routine tracheostomy care and tube changes—a critical clinical scenario for distress and tube displacement risk. The dataset was split at the patient level into 22 infants for model development and 3 infants withheld for a staged validation approach. From the edited video clips, 10,000 frames were extracted and annotated independently by two blinded pediatric otolaryngologists to generate ground-truth labels for tracheostomy tube cannulation status and facial distress indicators.
The AI pipeline incorporated two components: (1) a YOLOv11-based deep learning detector that identified cannulation status from video frames, and (2) a facial distress classifier harnessing facial landmark features and action unit signals analyzed through computer vision methods. Initially, generalized pretrained models trained on the development cohort were evaluated on held-out infants to simulate unseen patient scenarios. Subsequently, patient-specific calibration was implemented by adjusting decision thresholds using decision curve analysis to optimize sensitivity and specificity for each patient.
Key Findings
The pretrained cannulation detection model achieved moderate performance with an overall accuracy of 73.6%, sensitivity of 80.6%, specificity of 66.7%, mean average precision at 50% IoU (mAP@50) of 64.5%, and a 23.62% rate of frames where the tube was not detected. Notably, after patient-specific calibration, the evaluable-frame performance improved markedly to an accuracy of 94.0%, sensitivity of 99.7%, specificity of 87.4%, with an area under the receiver operating characteristic curve (AUROC) of 0.962 and area under the precision-recall curve (AUPRC) of 0.918, indicating excellent discriminative ability.
Similarly, the pretrained facial distress classifier exhibited strong baseline performance with accuracy of 82.5%, sensitivity of 87.7%, specificity of 77.0%, and a 21.8% failure rate in face extraction from video frames. Calibration further enhanced the evaluable-frame distress classification results to an accuracy of 96.0%, sensitivity of 97.4%, and specificity of 94.6%, with outstanding AUROC (0.993) and AUPRC (0.993), reflecting near-perfect performance.
This evidence supports that patient-personalized threshold tuning substantially elevates AI system reliability and robustness, likely by accounting for individual anatomical variability, positioning, and camera angles common in this pediatric cohort.
Expert Commentary
These findings compellingly demonstrate the feasibility of AI-assisted video monitoring for tracheostomy-dependent infants, a population for whom clinical surveillance is both challenging and critical. The use of a deep learning framework combined with patient-specific calibration represents a novel and pragmatic approach to enhance detection accuracy beyond generic models. By integrating facial action unit analysis, the system also captures nonverbal distress cues that might otherwise be missed during routine assessments.
However, it is important to note that the study was retrospective and focused on frame-level rather than event-level analysis. Clinical implementation will require prospective validation in real-world settings to establish effectiveness in recognizing actionable events, triggering alarms, and integrating with standard monitoring modalities. Furthermore, the system’s approximately 20-24% frame exclusion rates due to non-detection of tubes or face extraction failures highlight ongoing challenges with video quality, occlusion, or patient movement that must be addressed.
Nevertheless, this work aligns with the growing literature supporting AI applications in pediatric airway management and critical care, as recent studies have reported on AI tools for automated distress detection and device monitoring, suggesting broader potential for improving patient safety and clinical outcomes through continuous AI-driven surveillance.
Conclusion
The Eyes-On AI-assisted video monitoring system demonstrates promising accuracy for personalized frame-level classification of tracheostomy tube status and infant distress in a proof-of-concept retrospective study. Patient-specific calibration emerged as a key factor to optimize model performance, underscoring the importance of individualized approaches in pediatric AI applications. While these results provide a compelling foundation, prospective clinical trials are imperative to validate event-level detection capabilities, alarm thresholds, and integration with existing clinical workflows before routine use can be recommended. If successfully implemented, such AI technology could revolutionize continuous respiratory monitoring in tracheostomy-dependent infants, potentially reducing morbidity and improving timely intervention.
Funding and Clinical Trials
The study was conducted with institutional support and approved by the appropriate IRB. No external funding or clinical trial registration information was reported.
References
1. Cecola CF, Settoon C, Buck LS, Evans AK, Dunham ME. AI-Assisted Video Monitoring for Tracheostomy-Dependent Infants: A Proof-Of-Concept Study. Laryngoscope. 2026 Jun 5;136(10):4198-4208. PMID: 42248801.
2. Bozkurt S, et al. AI in Pediatric Airway Management: Current Status and Future Perspectives. Pediatr Pulmonol. 2024;59(3):605-619.
3. Johnson AM, et al. Deep learning for non-verbal distress recognition in infants and children. J Clin Monit Comput. 2023;37(2):341-350.
