Enhanced Detection of Patient-Ventilator Asynchrony Using Multi-Input Convolutional Neural Networks

Patient-ventilator asynchrony (PVA) can be common in mechanical ventilation (MV), but it is underdiagnosed. Automated PVA detection methods are limited in applicability and interpretability of results. This study deploys a tri-input convolutional neural network (TCNN) with class activation mapping to provide better detection and interpretability in PVA classification. Its performance is compared to that of a commonly used Dual-input CNN model. Each CNN model was trained with a total of 56,000 augmented breaths comprising of 7 clinically recognised PVA classes, including: 1) flow asynchrony; 2) reverse triggering; 3) premature cycling; 4) double triggering; 5) delayed cycling; 6) ineffective efforts; and 7) auto triggering. These augmented breaths were derived from 11 patients, each contributing 500 clinician-labelled breaths. Gradient-weighted class activation mapping (CAM) was embedded in the model to highlight key regions of the respiratory waveforms influencing model decision-making, improving result interpretability. Analysis of the patient cohort shows the TCNN model achieved marginally higher accuracy and specificity, while the DCNN model achieved slightly greater sensitivity [Sen, Spe, Acc: TCNN - 98.3%, 83.9%, 89.5% vs DCNN - 98.5%, 82.8%, 88.9%]. Incorporating a volume feature learning stream in the TCNN enhances result interpretability, where the CAM heatmaps are aligned with underlying physiological processes, particularly for auto-triggered breaths caused by air leaks from the ventilation circuit. Overall, both CNN models demonstrate good PVA detection performance, with marginal improvements from the TCNN model. Both models demonstrate the potential for integration into standard care practices, where continuous monitoring of PVA could offer clinicians enhanced insights into patient-ventilator interactions that are not currently available today. The overarching goal is to use modern tools to improve MV treatment and facilitate more informed decision-making.