GeNIe and SMILE

[pablrodr]

Hi, we've developed a system that generates quite an amount of data that I think that could be exploited using bayesian networks and I'd really appreciate if anyone could give me his/her opinion or just one hint in a word or two about it.

I'm a software engineer working for 5 years now in a neurorrehabilitation hospital in Barcelona.
We've developed a neuropsychological rehabilitation system for traumatic brain injury patients treatment, we've been testing it for 2 years in our hospital.

The system is designed for running in desktop PCs where the patients mostly have to click images using the mouse in a variety of situations to exercise cognitive funcions (Attention, Memory, Sequencing, Categorization,...). When the patient finishes each exercise he gets
a result (a number from 0 to 100, 0 the lowest puntuation and 100 the highest, for every exercise). The neuropsychologists schedule the exercises weekly for the patients.

Now after 2 years of exercises executions we've got a database of 300 patients and more than 50000 exercises executions.
When a patient starts the rehabilitation treatment he undergoes an
evaluation test assessing his cognitive functions (a 30 items test, each item result is a number from 0 to 4, 0 indicating normal functioning, 4 indicating the highest level of impairment). And when the patient finishes the rehabilitation process (about 6 months later) he undergoes the same evaluation test.

There are about 100 diferent exercises, each one with about 10 levels of difficulty and there's no predefined execution plan, the neuropsychologist programs the exercises session for his patients 3 or 4 times a week only based in the results of the previos sessions, there's not a predefined number of exercises for patient (it's about 8 to 14 exercises for session) it takes about half an hour for the patient to execute all the exercises of one session. So all patients did a different number of exercises during the rehabilitation process and there's no predefined sequence of executions in any way and each patient executed many times (more than 100 in many cases) the same exercise during the rehabilitation process.

Now we are planning to provide the neuropsychologists with some kind of decision support system (hopefuly based on bayesian networks) to assist them in the election of the exercises for each patient at any moment.

The hypothesis is that if an exercise is too difficult or too easy for the patient it is not useful, so we are trying to FIND THE RANGE OF EXERCISES RESULTS (from 0 to 100) that are most useful.
For measuring the usefulness of the exercises we've got the 30 items test above-mentioned. Maybe this range will be different for the different exercises types and at different moments of the rehabilitation process and for the different levels of impairment of the
patients.

Sorry for this so long explanation, it'd be really very useful for me to have an opinion about tackling this problem using bayesian networks, I have no experience in them but I think that could be what we are in need of.

Best regards

[marek]

[pablrodr]

Hi! the first problem that I have is that I don't see how to translate the clinical problem to a bayesian network and that's why I don't know if bayesian networks are suitable for the problem.

The clinical problem is:
Is there an optimal result range for the exercises? There should be a result range for the exercises that is better for a given patient in a given moment, for example if a patient gets a 0 result it means that the exercise is too dificult for him and if he gets a 100 the exercise is too easy. So maybe if he gets a result in the range 65-85 this means that it's not too easy or too dificult for him but we don't know if this is the optimal range, maybe it is for a given patient profile in a given moment and that's what we want to find: optimal results ranges for given patients profiles.

To find the patients profiles we've got a 30 items questionaire that assesses the patients before and after the patient executes all the exercises, but I don't see how to use the 50000 executions of the exercises Patients execute exercises 3 or 4 or 5 times a week during about 6 months and I should consider that it's not the same the first execution of an exercise when a patient is starting the rehabilitation that to execute that exercise 5 months later but I don't see how to model that using a bayesian network.
There's a whole story of previous executions that somehow are important for the result of the actual exercise but I don't see how to model that.. even if it is possible to model that..