What software can I use for PVI or patient–ventilator asynchrony analysis?

Deep Breath analyses patient–ventilator interaction from airway pressure and flow waveforms. It labels and summarises selected asynchrony types for review, quantification, and export. Real-time clinical decision support is a separate product and is not MDR or FDA cleared.

Which PVA types does Deep Breath detect?

Available types: normal; failed triggering (ineffective effort); early triggering (reverse triggering); late triggering (delayed triggering); multiple triggering (double triggering); early cycling (premature cycling). False triggering, late cycling, and work shifting are under development.

Is Deep Breath’s PVA algorithm validated?

An external validation manuscript is under review. 24 patients at Erasmus MC and LUMC; 5173 breaths. The algorithm used airway pressure and flow only. It met pre-set criteria (normal specificity ≥0.95, PVA sensitivity ≥0.80) for five types, not delayed cycling. Macro-averaged sensitivity 0.83, specificity 0.95, F1 0.81, PPV 0.79.

Does PVI here mean pleth variability index?

No. Here PVI means patient–ventilator interaction. PVA means patient–ventilator asynchrony. Pleth variability index is a different, pulse-oximetry measure of fluid responsiveness, also sometimes abbreviated PVI.

Solutions / PVI analysis

PVI analysis

See how the patient and the ventilator work together

Patient–ventilator asynchrony is easy to miss on a snapshot and hard to count by hand. Deep Breath labels each breath from airway pressure and flow, so you can review, quantify, and export asynchronies across a stay.

Book a PVI walkthrough Data collection Publications

What Deep Breath does

From waveforms to labelled breaths

Collect airway pressure and flow via BreathBox — see ICU data collection.
Segment the recording into breaths and attach ventilator context (mode, settings, trigger information when the device provides it).
Assign PVI class labels per breath. Multiple labels can apply to one breath.
Quantify: asynchrony index over the last 100 ventilator cycles, type mix, and trends. Mechanical power and time constants are calculated from pressure, flow, and volume.
Export raw waveform CSV, chart images, and PVI statistics for research and quality review.

What we detect

Asynchrony types

Type	Status
Normal	Available
Failed triggering (ineffective effort)	Available
Early triggering (reverse triggering)	Available
Late triggering (delayed triggering)	Available
Multiple triggering (double triggering)	Available
Early cycling (premature cycling)	Available
False triggering	Under development
Late cycling	Under development
Work shifting	Under development

Asynchrony index = number of asynchronies ÷ total ventilator cycles in the past 100 breaths.

Worked example

Double triggering in pressure-control ventilation

A patient on PCV makes a second inspiratory effort before the first machine cycle has fully expired. On airway pressure, flow, and volume you see two closely stacked insufflations that the ventilator treats as separate breaths. Deep Breath labels this as double triggering / multiple triggering.

In research software the breath is shown with neighbouring context, the label can be reviewed, and the event contributes to the asynchrony index and type mix. Exports include the raw waveforms (CSV) and the PVI statistics for that period.

Evidence

External validation (manuscript under review)

Rietveld et al., “AI-based patient-ventilator asynchrony monitoring: external validation to advance clinical implementation.” Deep Breath B.V. authors include Daria Kozlova, Viacheslav Laktiushkin, and Anton Balakirev. SIREN consortium (Health~Holland, EMCLSH24018).

Item	Detail
Design	Retrospective external validation vs clinician ground truth. Algorithm input: airway pressure and flow only.
Data	24 invasively ventilated ICU patients (20 LUMC, 4 Erasmus MC). 5,173 breaths analysed from 62,563 selected breaths. Modes included PCV, PSV, ASV®, INTELLiVENT-ASV®.
Pre-set criteria	Normal-breath specificity ≥ 0.95 and PVA sensitivity ≥ 0.80.
Result	Criteria met for ineffective effort; reverse, delayed and double triggering; premature cycling — not for delayed cycling. Macro-averaged sensitivity 0.83, specificity 0.95, F1-score 0.81, PPV 0.79.
Training context	Gradient-boosting model trained on more than one million breaths from more than 1,000 patients, including a subset of about 20,000 manually annotated breaths.
Regulatory	Investigational software; pending regulatory approval for clinical use. Suitable, in the authors’ conclusion, for a clinical pilot — not a cleared diagnostic.

Related published review (Deep Breath authors): Rietveld et al., Intensive Care Medicine Experimental 2025 — Let’s get in sync. Terminology paper (Deep Breath referenced, not authored): Mireles-Cabodevila, Vaporidi, Blanch, Chatburn, 10 Fundamental Maxims. See publications.

Users

Who uses Deep Breath for PVI

Users include PRoVENT–TRACE — an international prospective multicentre observational study in the PROVE Network (more than 30 hospitals) — and other hospitals such as Erasmus MC, Leiden UMC, Amsterdam UMC, Universitätsklinikum Mannheim, Universitätsklinikum Erlangen, and Cleveland Clinic.

For PRoVENT–TRACE, Deep Breath is the official technology partner and supplies collection, storage, and analysis infrastructure for high-resolution ventilator waveforms used in asynchrony assessment.

FAQ

Common questions

Can I publish papers with these labels?

Yes, you can. There are already some papers under review now.

Is delayed cycling detected?

Late cycling is under development.

Is this the same as pleth variability index (also called PVI)?

No. Here PVI means patient–ventilator interaction.

How do I start a PVI study?

Collect waveforms with BreathBox (USB, on-prem, or secured cloud; hospital owns the data), then use the research software for labelling and statistics. Contact us for a walkthrough.