Enhancing mechanical ventilation management with AI: Computer vision for automated detection of ventilatory modes,...

To evaluate the performance of an artificial intelligence (AI)-based decision support platform called NexoVent, which uses computer vision to automatically detect ventilator modes, parameters, and patient-ventilator asynchrony (PVA) from ventilator screen images in real time.

This observational study was conducted in the ICU of a tertiary care hospital. Images from Servo-i and Servo-s ventilators in PCV mode were acquired using standard mobile devices under various clinical conditions. The NexoVent platform used pre-processing filters, optical character recognition (OCR), and waveform analysis to extract alphanumeric and waveform data. Six types of PVA were evaluated: premature cycling, delayed cycling, ineffective effort, double triggering, flow starvation, and excessive flow. Performance was compared to expert consensus, which served as the reference standard.

A total of 621 respiratory cycles were analyzed to evaluate the accuracy of NexoVent in detecting ventilator mode and alphanumeric ventilator parameters. NexoVent identified ventilator parameters with an overall accuracy of 95.4 % and detected ventilator modes with an accuracy of 94.0 %. The system accurately detected asynchronies with performance ranging from 81.6 % (delayed cycle) to 97.8 % (ineffective effort). All analyses were performed using images only, without any direct interface to the ventilator hardware or software.

NexoVent accurately detects ventilatory data and multiple forms of PVA using non-invasive, image-based computer vision. These findings support the platform's potential to improve mechanical ventilation management and provide real-time clinical decision support in various ICU settings, especially where expertise or device interoperability is limited.