Category Archives: Technology

Can we measure muscle wasting in the ICU using Ultrasound?

Zudin Puthucheary (London, UK)

US can track changes in muscle mass of critically ill pts

Rectus femoris cross-sectional area (RF CSA) validated against Fibre CSA (gold standard – muscle biopsy of vastus lateralis) and Ratio of Protein:DNA in myocyte (biochemical gold standard in muscle mass measurement; UNAFFECTED by hydration status)

RF CSA and Fibre CSA both underestimate muscle wasting compared to myocyte Protein:DNA

–>> muscle wasting visualised on Ultrasound is likely to also be an underestimate


US sensitively discerns muscle loss between pts with different severity of illness – significantly greater in pts with higher number of failed organs


US + Bx of muscle, repeated 10d later –> smaller, brighter RF = cellular infiltrate + myonecrosis (present in up to 40% of critically ill pts)  — US is a non-invasive tool to detect changes in Muscle QUALITY and has been used in the paediatric population for over a decade

Limitations with using US to measure Muscle MASS – How to assess?


Changes in RF CSA are associated with changes in muscle strength (MRC score)

This change is not seen in MLT i.e. MLT is not associated with muscle function in critical illness

MLT also does not correlate with muscle mass (measured on CT)

–> MLT is not a good indicator of muscle mass in critically ill population

Lack of standardisation in examination technique and reporting are greatest barriers to external validity

–> reliability, reproducibility, accuracy should not be assumed

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

Data Science Masterclass

The field of machine learning, science that studies the design of algorithms that can learn, is advancing rapidly and is becoming widespread in critical care medicine given the large amounts of data collected routinely in the intensive care units. Typical tasks are concept learning, function learning or “predictive modeling”, clustering and finding predictive patterns. These tasks are learned through available data that are observed through experiences or instructions.

The goal of this Data Science masterclass is to teach doctors and other health care professionals basic concepts and skills and give tools for working more effectively with data. Moreover, in the literature there is an increasing number of papers describing AI/machine learning algorithms and prediction models so clinicians and other healthcare provides must know the key concepts of Data Science to correctly interpret results.

The Data Science masterclass was a very interactive and practical course were participants have the possibility to discover insights about large, rich and complex data sets, to find new ways to answer clinical questions using large datasets of electronic health records, to cooperate with specialists of different fields and to learn more about the potential of medical data, machine learning and predictive modelling that could provide new insights and improve patient care.

To start familiarizing with Clinical Data Science for Critical Care you need

  1. a laptop
  2. to install R and R studio
  3. to have or to sign up for a Google docs account (optional)
  4. to download and install a spreadsheet software

Moreover, you need to have an understanding of how files and folders (directories) are named on your computer because unlike your usual habit of pointing and clicking to open something you don’t have a graphical user interface (GUI) and you will need to start writing instructions/scripts in the R terminal.

What is R?

R is a free cross-platform (UNIX platforms, Windows and MacOS) software environment for statistical computing and graphics well suited to data analysis. R is not graphical (GUI) instead is based on scripts and the learning curve might be steeper than with other software. Working with scripts forces you to have deeper understanding of what you are doing.

Why R?

3 good reasons:

  1. You can do anything in R
  2. Science should be reproducible
  3. You have a vast support network

People think R is hard because it’s not a graphical user interface (GUI) and you have to describe what tasks you want the computer to complete in text, using the R language.

Data pipeline

Building data pipelines is a core component of data science. Data pipeline is a set of actions that extract data (or directly analytics and visualisation) from various sources to produce an output (tables, plots, manuscripts, presentations) thanks to a R script. 

After obtaining data from electronic health records databases, web servers, logs, online open-source repositories you have your data in a spreadsheet, you write instructions/scripts using the R language and you obtain an output: a table, plot or entire manuscript. You can change your data, or add new data, and run the script another time and instantly you regenerate the output.

Data preparation

Data preparation is the combination of data cleaning and data modelling. To be able to describe, plot, and test data must be tidy following the rule that “Each column is a variable. Each row is an observation.”. Data preparation includes variable re-naming, extract numbers and strings, parsing dates, columns to rows, missing and duplicate values.

Types of data: Not all data is equal, aim for consistency in every column, never try to record more than 1 type in a column: integers, decimals, strings, datetime, booleans, factors, try to think like a computer.

Data visualisation

Complex ideas must be communicated with clarity, precision and efficiency with storytelling, decluttering, avoid misleading and pie chart horror, scaling up and rational use of colours.

Visualisation is a fundamentally human activity. A good visualisation will show you things that you did not expect, or raise new questions about the data. A good visualisation might also indicate you that you’re asking the wrong question or you need to collect different data. 

Statistical modeling

Models are complementary tools to visualisation. Once you have made your questions sufficiently precise, you can use a model to answer them. Machine learning algorithms are divided in three categories:

  1. supervised: model training, focused on predictive tasks (e.g. risk of death, readmission, length of stay, early deterioration, …);
  2. unsupervised: discovery of latent structure/subclasses in a dataset, useful to define subgroups and phenotypes;
  3. reinforcement learning:virtual agents ought to take actions in an environment so as to maximize some notion of cumulative reward. This is the most immature branch of machine learning.


The last step of data science is communication, a critical part because It doesn’t matter how well your models and visualisation have led you to understand the data unless you can also communicate your results to others.

Tips in case of error messages

If you encounter any error messages during your Data Science practice just try copy and past your error message into and in most of the times you’ll find an answer.


Most of the material and sample code used in this Data Science masterclass is available online here

The suggested book for starting learning R for Data Science is “R for Data Science” and is available online here Moreover, remember that Google is your friend.


To conclude, my infographic from masterclass in Data Science at summarising the key concepts. Follow me on Twitter: Scquizzato Tommaso @tscquizzato.

ESICM Datathon: Day 3

Hemodynamics in septic shock Speaker: G. Baselli
Amazing how a ‘routine’ haemodynamic system can be viewed so differently from the point of view of an engineer/data scientist/medic

In conclusion

  1. Autonomic nervous system function and CV regulation is dynamic and hence the data needs to be dynamic and not a constant
  2. Need to pick out the meaningful physiological parameters to feed into the machine learning algorithm
  3. Important to have large open-access databases
  4. These databases need to integrate multi-scale information both in dimension and in time

Neurointensive care Speaker: A. Ercole

The concept of cerebral perfusion pressure = MAP – ICP is an example of a simple mathematical concept

We measure what we can, NOT what we should

Perhaps the autoregulation status of the TBI pt is more important – cerebrovascular pressure reactivity (PRx) Link

Data studies needs the same robustness as any other drug studies

Data Access quality and Curation for Observational Research Designs

Reinforcement learning in sepsis Speaker: M. Komorowski

The Artificial Intelligence Clinicians learns optimal treatment strategies for sepsis in intensive care

Can a computer help the clinician do the right thing?

Looked at 2 groups of treatment in septic patients – vasopressor use and fluid administration.

In general pts received more IV fluids and less vasopressors than recommended by the AI policy.

The machine considered much more variables compared to the human clinician


  • Reinforcement learning could provide clinically interpretable treatment suggestions
  • The models could improve outcomes in sepsis
  • Flexible framework transferable to other clinical questions

Gradient-boosted decision trees Speaker: C. Cosgriff

“Its not about doctors versus computers, its doctors with computers versus doctors without computers – @cosgriffc”

If you don’t know what gradient boost is (I don’t) but would like to find out more OR want to learn some basic coding in R/Python/SQL, have a look at

The power of XGBoost for the ICU is to shift towards ‘human intelligence’, supporting clinicians and intensivists (NOT REPLACING THEM)


  • Introduction to Statistical Learning
  • Elements of statistical learning
  • Deep Learning with Python
  • xgBoost: A scalable tree boosting system

Clinical research and AI Speaker: R. Furlan

This is highlighted by a local project looking at the ability to diagnose syncope using natural language algorithms of EHR compared to human clinicians

Moving models to the bedside Speaker: P. Thoral

Insight into the challenges on bringing machine learning and artificial intelligence models to the bed side.

Highlights the various legislative requirement with regards to introducing system (medical software is still considered a medical device)

Need to engage the stakeholders

The future: physicians, engineers, machines Speaker: R. Barbieri

Various scenarios of the future predicted by Prof Barbieri


  • Technology used in warfare
  • Technology cannot overcome environmental disasters

Best Scenario

  • Technology cures, predicts and prevents diseases
  • No need for ICUs


  • Technology blunts human knowledge
  • Humans lose ability to think
  • Technology takes over but are unable to make critical decisions

Oligarchy (aka BladeRunner)

  • Technology and knowledge controlled by a few
  • Progress without common wealth
  • Humans and machines start merging


  • Machines takes over
  • Humans and human nature are irrelevant