Medical Insurance Fraud Detection: Text analysis I'm trying to analyse a dataset to detect fraudulent insurance claims. Unfortunately, other than basic demographics the rest of the claim is a free format OCR scanned text file made from documents submitted by the patient. These documents include lab tests, hospital bills etc., patient summaries. 
Right now, manual coding is one option but expensive. 
Is there any chance that applying textual analysis / text-mining might yield some predictive strength? It need not be perfect but even if it can reliably identify high fraud potential claims, that's a plus. 
Any ideas how I could model this?
 A: This a tough one without revealing more about the data. I would say your best bet is to create features out of the plain text file based on subjective data. For example some features could be 
1) How likely is fraud from a given zip code?
2) Spell errors and other typos in a given claim
3) Income level if it is revealed somewhere on those claims
4) Others?
Once you have all of this, you want to break up the cases into a training set and a test set and use some basic machine learning methods. The Naive Bayes approach is a good one for such cases and works quite well. Hope this helps
A: Text mining is often used for fraud detection, so there is definitely predictive potential in text mining.
If I were you I would try to mine the text using n-grams in order to see if there are word phrases that yields any predictive power. Maybe one could imagine that fraudulent claims are more precise or formal, as some people are really good at knowing how to phrase the claims in order to go through.
So to try providing you with an answer, I would do traditional text processing procedures like stop word removal, convert upper to lower case, prune etc. Having done that I would use TF-IDF or binary weighting scheme and extract up to 3-grams depending on your computing power and see if that yields any predictive accuracy. It should...
One can use several algorithms, but dependending on how many training cases you have I would go for support vector machines, Naive Bayes and or decision trees. Decision trees are good if you want to understand which features/variables that are descriminative.
