1. Predicting Vasospasm after Subarachnoid Hemorrhage Using High-Frequency Physiological Data Soojin Park MD
Murad Megjhani PhD
Hans-Peter Frey PhD
Edouard Grave PhD
Chris Wiggins PhD
Noemie Elhadad PhD
K01 ES026833

2. Conflict of Interest Disclosure & Acknowledgement
The co-authors have nothing to disclose with regard to commercial interests.
I do not plan on discussing unlabeled/investigational uses of a commercial product.

3. Multiple monitoring modalities of brain function and physiology
ICP
CPP
TCD
PbtO2
ETCO2
Microdialysis
Thermal Diffusion
CBF
NIRS
Laser Doppler Flowmetry
SJVO2
Cortical EEG
Depth Electrodes
AMAZING OPPORTUNITY as we learn more about the BRAIN and its COMPLEXITY
We have more devices and measurements available to RESEARCHERS and CLINICIANS.
There is also an amazing opportunity for DATA SCIENCE in how to make sense of the Heterogeneous data, with different sampling rates and missing data issues.
Heterogeneous streams of data
Different sampling rates
Missing data issues

4. Additional complexity of ICU environment
As a stroke and critical care neurologist focused on the complexity of the brain.
I face the daily challenge of additional mundane complexity of the ICU environment with all its competing demands on attention.
In the time-pressured environment of critical care, providers require a great
deal of cognitive fortitude to overcome the vital threshold of gaining even a holistic
view of the patient

5. Looking for Actionable Knowledge
And so what I and we all as a community are looking for is how to translate BRAIN and PHYSIOLOGY monitoring data into ACTIONABLE KNOWLEDGE.

6. Purpose of Neuromonitoring = NCC
PREVENT  IDENTIFY  TREAT
Assess extent of primary brain injury
Detect secondary brain injury – early enough, START RX
Measure effect of interventions – STOP RX
Prognosticate recovery/outcome
Purpose of NCC is detecting secondary brain injury

7. Purpose of Neuromonitoring = NCC
PREVENT  IDENTIFY  TREAT
Assess extent of primary brain injury
Detect secondary brain injury – early enough, START RX
Measure effect of interventions – STOP RX
Prognosticate recovery/outcome

8. Vasospasm after Aneurysm Rupture
14.5 per 100,000 in US
Compared to other strokes, younger patients
Substantial burden on health care resources
Long term functional and cognitive disability
Common disease entity (25% patient-days)
Timely interventions for VSP to prevent stroke:
Prediction scales: resource utilization (intensity of monitoring)
First 14 days: Detect preclinical or early VSP with TCD
One type of secondary brain injury that is particularly significant to study is VSP after aneurysm rupture.
This disease affects younger patients than other types of stroke so leads to a substantial burden on health care resources.
Poor outcome is caused by VSP which leads to stroke if not detected early enough.

9. Vasospasm Prediction based on admission CT
Baseline risk score: Fisher, Modified Fisher Scale Used in clinical care Advantageous for simplicity
Symptom
Infarct
Vessel Narrowing
Reported in Liter.
Scale
Within study
Angiographic VSP X
50-70%
Fisher
27/41 (66%)
Delayed Cerebral Ischemia
19-54%
Claassen
54/276 (20%)
Symptomatic VSP
20-40%
Modified Fisher
451/1355 (33%)
There are established and validated baseline risk scores for VSP
That are used daily in clinical care
The data they are using are admission CT
And they predict for VSP as defined by symptoms or stroke or vessel imaging.

10. ‘It’s tough to make predictions, especially about the future’ (Danish, unknown)
MFS 1 (low grade)  24% Sx VSP
MFS 4 (high grade)  40% Sx VSP (OR 2.20)
Comparing two patients:
Patient with MFS times more likely to develop VSP than patient with MFS 1
Individual patient:
24% of patients with MFS 1 developed symptomatic VSP
Even when someone is in the highest risk category for VSP
There is still a lot of uncertainty as to whether that individual will develop VSP
Again, this lack of ACTIONABLE KNOWLEDGE is crippling us.
Figures from Frontera Neurosurgery 2006

11. Novel sources of data for prediction
Electronic Health Record (EHR)
Continuous Physiologic and Brain Monitors
Are there patterns in high frequency time series data that are informative for VSP classification?
No hypothesis of what, if any, pattern exists
Orders
Procedures
Phenotype
Assessments
Laboratory
Radiology
Physiology
Brain monitors
We have more data than we are using in these baseline scores.
And we have better methods that we can leverage to make sense of this data.

12. VSP prediction Challenges for VSP prediction
Infinite Dimensional Space
Project it to an infinite dimension and separate it with a hyperplane
Linearly Separable
Data points that are not
Linearly Separable (complex
non linear boundary)
Dealing with
non-linearity
This is a picture of classic types of interactions among features
In our case we suspect we have non-linear interactions
But our biggest question is one of feature engineering.
Of all these novel data with all of their complexity, what exact features do we want to incorporate in our classifier.
The added challenge is that we have continuous monitoring that is temporal in nature.
Challenges for VSP prediction
Feature engineering (which variables are discriminative)
Temporal prediction (how to translate continuous stream into actionable feature)

13. Random Feature Idea Random Kitchen Sinks (RKS)
Classifier
Project to infinite dimensional space
Create a gram matrix of size nxn
As n increases, becomes
computationally challenging
Big Data
Proposed
That question of feature engineering is a classical problem in machine learning
If we want to stay agnostic in our hypothesis and want to discover features that are discriminative in our classification, then this idea of generating lots of random features and seeing what sticks is one that has shown to be extremely successful.
Randomly Featurize
Classifier
Project to finite
low dimensional space
Rahimi. Weighted Sums of Random Kitchen Sinks:
Replacing minimization with randomization in learning.
In: NIPS vol Citeseer (2008)

14. Making RKS temporal Random featurization etc. Convolution
Different downsampling rates
Varying temporal lengths for random kernels RR
The novelty is that we applied this random featurization to temporal data.
etc.
Multiply at each time point  take the integral of the multiplications
The higher your convolution, the better the pattern fits your data
Random featurization
extracting features at different scales
to capture time varying characteristics
for different variables

15. These are the results on our dataset, one of the largest in the nation of high frequency SAH data.
We are able to show that random featurization of temporal physiologic data significantly improves prediction above established clinical risk scores.
mRMR
Under review

16. What if our prediction model is a detection model?
A problem with building prediction models using temporal data, especially when the syndrome has a subclinical onset (such as vasospasm or sepsis), is that the clinical expert detection corrupts the data.
Detecting clinical intervention is not useful

17. Conclusions
Higher frequency time series data are superior to lower frequency data for discriminatory feature generation in VSP classification. “Automateable” monitor data prior to peak VSP period shows promise to provide individualized prediction of VSP. This might be an informative feature extraction approach with universal subset of physiologic parameters (HR, RR, SBP, DBP, O2 sat). Syndromes with insidious onset are a peculiar case for data mining experiments.

18. Acknowledgements Murad Megjhani PhD - Neurology
Hans-Peter Frey PhD - Neurology
Edouard Grave PhD – Biomedical Informatics
Chris Wiggins PhD – Applied Physics and Applied Mathematics
Noemie Elhadad PhD – Biomedical Informatics
BD2K K01 ES026833