The following content has been provided by the University of Erlangen-Nürnberg.

So today I'm going to talk about the design of an application-specific processor where

we modified the instruction set in order to support very advanced algorithms for real-time

medical monitoring applications.

So the motivation is a condition called pericardial tamponade in which the sac around the heart

starts to fill up with fluid and this impedes the ability of the heart to properly plump

blood.

This is a very good motivating example for real-time medical monitoring because this

condition is 100% fatal if left untreated, yet if you can detect it, it is very, very

easy for a doctor to treat. You simply puncture the outer membrane of the heart with a needle

and you drain the fluid out and everything is fine after that.

Now the problem is that this can be detected and treated easily for a patient under doctor's

care. So for example, this might happen due to a small mistake made during surgery, the

doctor accidentally cuts something, the fluid starts to flow. So in that context it's very,

very easy to quickly detect and quickly fix.

But this is also something that can just happen to a person if they're walking around. And

typically if this starts happening and a doctor doesn't know and you fall down and an ambulance

is called and you get taken to the hospital and they start diagnosing you, you have a

much higher likelihood of dying because this is a real-time problem. You have to detect

that it occurs and get a needle in there fast enough. And if people don't know how to look

for this, it can be very hard.

So what we want to do is to design a wearable computing device that can analyze signals

from the human heart and the human respiration system that can detect very quickly this and

many other conditions that could be potentially harmful. So it turns out that under normal

operating conditions, your heart and your respiratory system operate completely independently.

They do not affect one another. But when tamponade tech starts to occur, the pulse and the respiration

interferes with the pulse. Now this is fairly easy to detect if you understand what you're

looking for in terms of signal processing and you can easily design a computing system

to do this. But figuring out and analyzing this in real time is very computationally

intensive. So that's what we're trying to do. We're trying to create a low-cost hardware

solution that will enable us to monitor for this condition in real time.

The problem that we're looking at is a general term called time series monitoring or in the

database community, similarity search. We are looking for things that are similar. So

things that are similar. We have template patterns that will allow us to detect this

pulses paradoxes condition and we need to be able to look at an ongoing signal being

read from a person and you can say how similar is this to a syndrome that we're looking for.

Of course, finding similar things is a much more general problem in data mining. So many

techniques that have been designed for things like classification of language or similarity

of images, different aspects of these algorithms can be applied in order to help find similarities

in time series data which is what's shown up top. So I'm going to be talking primarily

today about looking at time series similarity but many of the techniques can actually be

generalized and extended to do different types of similarity searches.

I'm going to talk primarily about two similarity measures. The first is very simple. It is

called Euclidean distance. You take your two curves and you simply align all of the points

in time and you take the difference in measuring them and then you actually sum the squares

of the distances and take a square root at the end. So this is very simple, linear to

compute. The problem with Euclidean distance is that it cannot handle situations where

the curves are in fact quite similar to one another but have small shifts in time. So

you could argue that there is two forms of similarity here. First, each of these curves

has four bumps. Second, if you look at the two bumps together are spaced very similarly