Monitoring the Injured Brain

23 -25 June in Dublin, Ireland.

The Sixth International Conference on Condition Monitoring and Machinery Failure Prevention
Technologies

Iain Chambers presented the Avert-IT project and it’s work in the area of monitoring injured brains.

As a result of the presentation Iain has been asked to publish a paper detailing the research and the possibilities of extending the BANN technology to the condition monitoring sector.

Philosophical Transactions of The Royal Society

June 1st 2009

The Royal Society today published a paper presented by our colleagues as follows:

Federating distributed clinical data for the prediction of adverse hypotensive events

Abstract

The ability to predict adverse hypotensive events, where a patient’s arterial blood pressure drops to abnormally low (and dangerous) levels, would be of major benefit to the fields of primary and secondary health care, and especially to the traumatic brain injury domain. A wealth of data exist in health care systems providing information on the major health indicators of patients in hospitals (blood pressure, temperature, heart rate, etc.). It is believed that if enough of these data could be drawn together and analysed in a systematic way, then a system could be built that will trigger an alarm predicting the onset of a hypotensive event over a useful time scale, e.g. half an hour in advance. In such circumstances, avoidance measures can be taken to prevent such events arising. This is the basis for the Avert-IT project (http://www.avert-it.org), a collaborative EU-funded project involving the construction of a hypotension alarm system exploiting Bayesian neural networks using techniques of data federation to bring together the relevant information for study and system development.

The full text and of names of contributors can be accessed Here

Another Question Answered

All the clinicians involved in our project are specialists in neurosurgery. Hypotensive events are a key issue for them, because of the damage caused to the brain.

But a major question for us has been will our research result in a technology which could apply in general intensive care. Specialist neurosurgery units account for only a small proportion of the total ICU beds. We’ll find it much easier to justify a commercial operation if all ICU beds are potential installations.

Great News – Physionet based at MIT and funded by the National Institutes of Health’s NIBIB and NIGMS, recently announced its “challenge” for 2009 – Predicting Acute Hypotensive Episodes.

As a basis for the challenge Physionet provides a data set collected during intensive care of cardiology patients. The challenge itself isn’t appropriate for our research – we’re targeting a different definition of episode -but the instructions for the challenge provides an interesting perspective. Here’s and extract.

“Among the most critical events that occur in intensive care units (ICUs), acute hypotensive episodes require effective, prompt intervention. Left untreated, such episodes may result in irreversible organ damage and death. Timely and appropriate interventions can reduce these risks. Determining what intervention is appropriate in any given case depends on diagnosing the cause of the episode, which might be sepsis, myocardial infarction, cardiac arrhythmia, pulmonary embolism, hemorrhage, dehydration, anaphylaxis, effects of medication, or any of a wide variety of other causes of hypovolemia, insufficient cardiac output, or vasodilatory shock. Often the best choice may be a suboptimal but relatively safe intervention, simply to buy enough time to select a more effective treatment without exposing the patient to the additional risks of delaying treatment.

Of the 2320 patients whose monitored waveforms and accompanying clinical data were included in the MIMIC II Database as of December 2008, arterial blood pressure was recorded in 1237 (53%); among these 1237 patients, 511 (41%) experienced recorded episodes of acute hypotension (as defined below) during their ICU stays. The mortality rate for these 511 patients is more than twice that of the MIMIC II population as a whole. To the extent that one might forecast acute hypotensive episodes in the ICU, there is a possibility of improving care and survival of patients at risk of these events. “

Just What Is Prediction

When it comes to forecasting the future we can get into hot water in a hurry.  After all no two economists can agree on anything, I’m told.  The collapse in financial markets recently seems to bear that out!

Forecasting seems to be an especially risky activity when the results challenge accepted thinking.  Highly qualified scientists who learn one way of doing things aren’t easily persuaded somebody has a better way of working things out.

When a new approach comes along it automatically gets associated with reading tea leaves or astrology.

We’re experiencing a little bit of this at the moment, not at all because our audience is Luddite, precisely the opposite. It’s extremely well qualified in its field.  What we’re doing verges on “black magic” mostly because we can’t easily explain what’s going on.

At a recent meeting we got into a discussion about whether our prediction system is “early detection” or “forecasting”.

The “early detection” guys were clinicians – people who know enough about physiology to be able to know if one thing is happening it will continue to deteriorate and finish up as something else.  This knowledge is based on experience, of course.  The result is a prediction based on what has happened before.

The “forecasting” guys were engineers.  These are people who “model” various parameters and use math to calculate the probability of outcomes.  The result is a level of confidence regarding a prediction of what should happen, based on known parameters.

Both are valid approaches.  Unfortunately they both suggest our approach doesn’t meet the standards demanded by science.  Basically we can’t explain the inputs to the calculations or how the outputs are derived.

Actually we aren’t much different to the strange old lady reading the tea leaves

Luckily the other group we’re working with – Neural Network experts – have come across the problem before.  They give us confidence we’re on the right track.

Bayesian Neural Networks add value precisely because they get into stuff mere humans can’t cope with.  They process literally millions of “what if, then” questions and provide us with answers meeting our criteria.

They recognize patterns that traditional science can’t see.

We’re now in a place where, as humans, we can’t explain exactly what the BANN has found.  Our only quality check is how accurate the pattern recognition proved to be.

We do know what the software is doing – we built it after all.  But digging into the calculations and results of the millions of calculations would take more years than any of us have left.

Our BANN is questioning the physiology parameters down to a level the human brain can’t cope with.  It’s finding stuff that one day others will understand. But in the meantime we can only decide whether to believe it, or not, based on the accuracy of it’s predictions.

We can’t explain to the clinicians and engineers the association between physiology parameters.

We can only tell them when we’re seeing predictions which turn out to be right more times than they’re wrong.

Those of us who have been around computers for long have always known there would be a time when the machine could take over.

It seems Avert-IT is finding that point.