The summary of Dr. Manasi Nandi's presentation from the Jun 11-12th 2019 event Data-driven systems medicine at Cardiff University Brain Research Imaging Centre.
Dr. Manasi Nandi (King's College London) - Data-driven systems medicine
1. Extracting new information from old signals
Manasi Nandi PhD, FBPhS, FHEA
Senior Lecturer Integrative Pharmacology
King’s College London, U.K.
manasi.nandi@kcl.ac.uk
Google images used throughout
2. Mahomed; Med Chir Trans. 1874;57:197-228. The Etiology of Bright's Disease and the Prealbuminuric Stage
Images courtesy of King’s College London; Foyle Special Collection library.
Since the mid 19th century we have known there
is extra information in waveform shapes
4. Source: Chung, M.K., and Rich, M.W. Introduction to the cardiovascular system. Alcohol Health and Research World
14(4):269–276, 1990.
5. • As individual control systems in a plane
start to fail, so the plane shakes, sways,
wobbles, loses altitude and finally
crashes….
• Similarly, in the human body, early failure
of control systems may not be detectable,
such that when we ‘diagnose‘ a patient –
they have already ‘crashed’
Early detection in ‘stable’ phase is
better than late detection in ‘crash’
7. Summary so far
Can we make better use of the
high fidelity - ‘big data’ that we
routinely collect?
Would this more sensitively tell
us about CV changes in
health and disease?
11. Can we plot blood pressure data in 3D?
We only have 1 data stream..
but this can be split by
using time delays
Floris Takens –
Mathematician
1981
Philip Aston
Mathematician
University of Surrey
2013
Can we plot blood pressure data
in 3D?
13. Use 𝜏 to get (𝑥, 𝑦, 𝑧) coordinates
(101.309, 117.469, 125.191)
(102.622, 118.354, 125.175)
(103.827, 119.239, 124.977)
(104.956, 120.063, 124.489)
…. etc
Step 1: Replot data into 3D
14. Step 2: Project to 2D
Change the coordinates to get a 2D attractor
3D attractor 2D attractor
Step 2: Project and view in 2D
16. If:
X = 108 mmHg
y = 120 mmHg
Z = 138 mmHg
100110120130140
140
120
100
100
110
120
130
140
X
y
z
x
140 130 120 110 100y
100
110
120
130
140
100
120
z
100
120
140
100
120
140
140
120
100
140
120
100
BP(mmHg)
Time (seconds)
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
z
y
x
1
2
3
4 5 6
Attractor reconstruction
Projection
Symmetric
17. Extracting measures from the attractor…..
How big is it?
How dense are
the hot spots?
What is the
angle of
rotation?
How wide are
the sides?
and relating these to physiological change…..
but applicable to deep learning image analysis to quantify attractor
…all features are quantifiable
18. BP ECG
Pulse Oximetry Intra Cranial Pressure
Central Venous Pressure Respiratory
Courtesy of Gary Chaffey – mathematicians, U of Surrey ; Physionet and other open access online sources
Agnostic to waveform..
19. Telemetry continuous
waveform data 1000Hz
BP software
Extraction of SBP, DBP,
HR, HRV and PP
every half hour
AR coding
Extraction of AR
measures, size, form,
density etc.
every half hour
10am-4pm naïve
10am-4pm sepsis
Preclinical sepsis modelling….
25. Summary
• Periodic physiological data (BP, ECG) is routinely sampled at high fidelity
• Attractor reconstruction extracts new information from these signals
providing quantitative measures of waveform shape and variability
• These can enhance the sensitivity of detecting cardiovascular change
• Numerous potential applications in healthcare and biomedical research
setting
• Appropriate for machine learning and A.I approaches from large datasets
• Two communication between disciplines is extremely important –
mathematicians need to perform physiologically meaningful analysis
• Attractor ‘signatures’ must be translated to physiological meaning – not
just a ‘black box’ of new numbers…
Summary
26. Disclosure: Nandi co-inventor: Delay Coordinate Analysis of Periodic Data WO/2015/121679.
http://ehealth.kcl.ac.uk/cardiomorph/
Aston et al., Physiol Meas. 2018 Mar 1;39(2)
Nandi et al., Physiol Meas. 2018 Oct 30;39(10)
