基于j2ee数据挖掘的健康预测系统

Question

I want to develop a health prediction system using data mining. Can anyone give me some tips on how to develop it?

The requirement is that when user enter first symptoms then system will check that in how many disease have this same symptoms after that system will give some options related to symptoms so it can differentiate between those disease and can infer what the disease is.

The added component here (and this is where the data mining and prediction portions comes in) is that, when the user enters a symptom, it should also suggest other symptoms that they might also be experiencing. For example, if they have a fever, there is a high probability that they also have chills, so when they enter "fever" it should suggest "chills" as an additional symptom.

Answer 1

Here's one possible database design:

Here's an example of what the data would look like: So, basically, this is saying that the symptoms of the flu are fever, chills, and fatigue, but the symptoms of a cold is just "fatigue." (Obviously this isn't all that complete but it's good for illustration).

One person searched for a disease whose symptoms were "fever" and "chills." A second person searched for a disease whose symptoms were just "chills." A third person searched for a disease whose symptoms were "fever" and "fatigue."

Searches would be done with a stored procedure: you input one or more symptoms and it inserts a record of the search into the SearchHistory table and retrieves a list of diseases associated with those.

If you want to calculate the odds of two symptoms going together, you can have a User-Defined Function that calculates the percent of searches that contain both symptoms. For example, if searches 1, 2, and 3 have both "fever" and "chills" as symptoms, but search 4 has "fever" and "fatigue" and search 5 has just "chills," there's obviously a 60% chance that someone who searches for "fever" will also search for "chills."

There's also a 20% chance that someone who searches for "fever" will also search for "fatigue," but in this case the sample is obviously too small to know for sure if that's actually representative. (That's one of the downsides to this design - your predictions will get better over time, but the early predictions might not be too accurate).

The advantage of this is that you don't have to manually enter any data on the odds of any symptoms occurring together, and the system'll "automatically" adapt and improve over time (ie your predictions will keep improving as you get more data) depending on what users of your system search on.

The downside, of course, is that you'd only have data on likely co-occurring symptoms once people started using the system, so early users wouldn't get the advantage of having the predictions, and it would take a little while before it was all that accurate at predicting what symptoms are likely to "go together." (Think of the case above where you were predicting a 20% chance of "fever" and "fatigue" going together based only on a single search).

Hope this helps some.