Personalising Care: Machine learning from pressure waves (ICP)

Personalising Care: Machine learning from pressure waves (ICP)

Soojin Park, Associate Prof. Neurology

Division of Neurocritical Care, Columbia University, NYC, USA

Watch on demand:


  • 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..)


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..)


  • 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


  • 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

Blog by Chris Tomlinson:

Anaesthetist & Critical Care Registrar

‍ PhD Candidate at UCL UKRI Centre for AI-enabled Healthcare | LinkedIn | @tomlincr