Merging information of several events I'm working in a database related to endometrium ultrasound. My DB contains several columns that may describe one or more injuries (scar tissues) by dimensions and volume: Injury1Height, Injury1Width, Injury1Volume, Injury2Height, Injury3Width, etc...
The problem I'm running right now is that this columns tend to be sparse (f.e. most patients may have Injury-1 filled, but Injury-4 will be mostly empty). Therefore several of these columns tend to be ignored while trying to fit a model.
Now my question, what is a clever way of merging this information? (I will try right now just adding up 'volume' of all Injuries, but I think this is quite of a bad idea).
Is there some "keyterm" for what I'm trying to implement?
 A: You could try to represent the same information in a different way - as meta-information, similar to what you already suggested yourself. As the order of scars does not play any role, classification should probably more rely on the fact that e.g. a certain amount of scars is there at all (and on their aggregated, unordered properties). Here are some concrete things that came to my mind for your scenario:


*

*The amount of injuries per sample could become one simple numeric feature for that sample.

*The mean/min/max/SD/MAD of height/width/volume of all scars per sample could become numeric features of that sample.

*Maybe the relation between the individual widths/heights/volumes of scars of one sample could become features too - but this will depend on what the data represents/your domain specific knowledge about the data. Keep in mind that if there is a clear relation between widths/heights/volume for all scars already, you might also be able to reduce those features altogether (classic feature selection/dimensionality reduction). 

*[...]


All of those essentially give you the same amount of meta-information per sample - from the underlying, actual, hard information you have from each sample's scars. Of course there are many more ways of abstracting information like this, so you could employ some further, similar techniques too.
