35% and rising

As part of our project review, September 30th, Rob Donald presented a report on the results of his research with the Bayesian Neural Network.

For us mere mortals the concept really isn’t easy to understand, but Rob found an excellent way to illustrate how the BANN works. It went something like this:

Take a full set of measures for an individual patient and present them to 100 clinicians. How many of the 100 predict an adverse hypotension event will occur in the next hour?

That’s what the BANN is doing – being 100 clinicians with different opinions.

In Rob’s research the clinicians are already getting it right 35% of the time, but this is only the beginning.

When he’s added the functionality for the prediction engine these clinicians will be getting it right much more often.

Clinician led research

After our project review, Ian Piper and I had an interesting discussion which we’d like to get everybody’s views on.

It seemed to us a lot of innovation in health care starts with the wrong people, and ends in the wrong place! Manufacturers of drugs and devices decide which innovations are required, and have to engage the clinical community in the benefits post clinical trials.

How much more appropriate would it be for clinicians to decide the innovations required, and associated benefits, and work with suppliers on the research?

Surely the innovation process would deliver more, with less cost and effort, and achieve more for all stakeholders, if clinicians could drive the process.

Research snapshot – the Bayesian Neural Network

The research is now well into the second phase – analysis of the historic data set.

In the first phase the statisticians and clinicians analyzed the historic data set looking for measures which could be used as standard indicators of adverse events.  This was a critical step in the project.  Only with common indicators across all cases, and agreement with the clinicians as to the reliability of those indicators in medical terms could the team proceed the the current phase.

The original project plan anticipated phase one would identify linear combinations of measures.  These combinations would guide the second phase analysis as to the time windows, prior to the event, appropriate for detailed analysis by the BANN.

Unfortunately things turned out to be much more complex.  The onset of the adverse event wasn’t indicated  in the linear combinations of any of the measures.  The idea of establishing a fixed time window for further analysis, shortening the duration of the second phase turned out to be false hope.

Now the second phase needs to compare relationships between up to eight variables through the smallest possible time windows prior to the event, and trend each of those relationships through one hour or more.  And do this across 200+ cases.

This work is now underway.  In order to get the computers processing these very complex queries in a matter of hours rather than weeks our computer scientists needed to add their skills to those of the clinicians and mathematicians.

Whilst this particular research is focused on finding indicators of upcoming hypotension, the clinicians are expecting a number of additional results which will help guide the care of TBI patients.

Clinical Interfaces Review

On July 15th Debrah Scobie demonstrated the initial clinical interfaces for users to review and suggest enhancements.

In the near future we’ll be including in this post a video of the demonstration