Validation of an automated PVA detection algorithm

What this abstract is

A validation of an AI-based algorithm, named Deep Breath in the abstract, that detects multiple patient–ventilator asynchrony (PVA) types in real time: ineffective effort, reverse, double, and delayed triggering, and premature and delayed cycling.

Ventilator and esophageal-pressure (Pes) waveforms were collected from 24 ventilated patients. 5,169 breaths were labelled for PVA by three clinicians from a pool of twelve. Ground truth was majority vote or consensus. The algorithm classified the same breaths using only ventilator waveforms (no Pes).

Validation criteria were at least 0.95 specificity on normal breaths and at least 0.8 recall per PVA type. The authors reported macro recall 0.83, specificity 0.96, and F1-score 0.81. Minimum PVA recall was 0.8 (delayed triggering); normal-breath specificity was 0.97. Both met the pre-set criteria.

This is a conference abstract in an ERJ Open Research supplement, not a full original article. It is research evidence for PVA classification. It is not MDR or FDA clearance of a bedside decision-support product.

Deep Breath note

The abstract names Deep Breath as the algorithm validated for real-time detection of multiple PVA types on ventilator waveforms only (24 patients) against expert labels with esophageal pressure. Sensitivity / specificity / F1 versus those labels:

Class	Sensitivity	Specificity	F1
Normal	0.84	0.97	0.86
Ineffective effort	0.89	0.99	0.84
Reverse trigger	0.85	0.99	0.84
Delayed trigger	0.80	0.97	0.75
Double trigger	0.89	1.00	0.80
Premature cycling	0.84	0.98	0.79