ESICM LIVES2020: Personalising Care: ML from ICP waves

Below you can find my notes from this talk at the ESICM LIVES 2020.

Thanks to Adrian Wong (@avkwong) for inviting me to blog on the ESICM website where you can find the full version of the below, including some slides.


💬 Personalising Care: Machine learning from pressure waves (ICP)
🩺 Soojin Park, Associate Prof. Neurology
🏥 Division of Neurocritical Care, Columbia University, NYC, USA
📺 Watch on demand: https://lives2020.e-lives.org/media/machine-learning-pressure-waves


❗ Motivation
· Acute hydrocephalus affects ~37k pts/yr in USA
· Rx = EVD, but 1/5th develop infection ventriculitis
· Risk of ventriculitis ↑ with duration and frequency of CSF sampling (by which diagnosis made..)

❓ Question
Can we find a way of using physiological information contained in ICP waveform to develop a method for detecting ventriculitis, without having to sample CSF?

Park reminds us of the normal ICP waveform (exam revision déjà vu..)

And how it’s morphology changes with ↑ICP

This alteration in waveform morphology with ↑ICP has a biologically plausible mechanism in ventriculitis.

⭐ Goal 1
Examine changes in ICP waveform morphologies prior to ventriculitis

· Dataset = only patients WITH ventriculitis
· Collaboration with group experienced in ICP waveform big data, however their pre-processing identified abnormal waveforms as artefactual!
· ⚠Problem = vague definition of ventriculitis
· Used ‘gold standard’ of limiting it to those with culture-positive CSF
· n = 19 pts
· ❗ Park mentions that CSF is cultured 3/w at this institution, perhaps not usual practice – CT: worth considering this in the context of their motivation

· ⚠ EVDs left open to drainage most of the time, typical practice across other institutions, thus waveform only intermittently present when EVD clamped by nurse
· ❓ Challenge = automating identification of waveforms (CT: I note solution was not to get desperate medical student to manually sift data in exchange for ‘research experience on their CV..)

🔧 Methods
· Dominant pulses extracted using Morphological Clustering Analysis of ICP Pulse
· Before / During / After ventriculitis (i.e culture-positive CSF)
· Morphologically similar groups obtained by hierarchical k-means clustering
· Dynamic Time Warping used as a ‘distance’ metric to correct for speed (HR), see below
· Meta-clusters determined by clinicians, see figure B below.
· Bi/triphasic (green)
· Monophasic/tombstone (yellow)
· Artefactual (red)
· = supervised learning

📜 Results
· Prior to ventriculitis majority of pulses had physiological tri/biphasic appearance
· During ventriculitis this dropped from 61.8 > 22.6%, a statistically significant change, which persisted
· ✨ Most importantly this change occurred a full day before the ventriculitis was clinically detectable

⭐ Goal 2
Leverage time-varying dominant pulses of ICP from hourly EVD clamping data into a detection model of ventriculitis
· Collaboration:
· Columbia Vangelos College of Physicians & Surgeons
· R Adams Cowley Shock Trauma Center, University of Maryland
· Aims:
· Improve performance and generalisability of model to other institutions data
· Work in submission therefore not shown
· Collaborators sought, see email below:

🚩 Concluding Remarks
· Example presented for ICP but process generalisable to other waveforms, of which there are many in ICU!


🙋‍♂️ My thoughts:
· I’ve often been disappointed at how little waveform data is actually stored from ICU monitors
· Perhaps I shouldn’t be given the general lack of high-quality ICU data (see data sharing session) and huge storage requirements
· Most of the ‘high resolution/granular/insert other buzzword here’ EHRs I’ve come across sample at a frequency ~ 1 hz (c.f. 125-250 hZ in this study)
· Starting point for those interested in waveform data in ICU = MIMIC-III Waveform Database
· Be warned this is truly big data!

Dr Chris Tomlinson
Dr Chris Tomlinson
👨‍⚕️ Critical Care Doctor | 🎓 PhD @ UCL CDT in AI-enabled Healthcare

Related