CNN and LSTM issues

BUTRRRS

Hello, Rapidminer community,
I am a new RapidMiner learner.
Recently, I got some data and I also need to use those data to predict whether a patient has hypertension. 
Here are some files:
Patients_res.csv: It already has results for patients, based on signal.csv
signal.csv: it includes those resulted patient signal from time t=0 to t=2099 


NP_patient.csv: the patients need to predict.
SP.csv: need to predicted patients signal.


One question is that can I use the LSTM or CNN to predict hypertension based on signal features?
Thank you

Mate

Hey @BUTRRRS,

as a first step I would recommend you check our tutorial processes which might give you an idea how to provide data in an appropriate format (if I understand you correctly that's the essence of your question here: how to supply data appropriately right?).
E.g.:
The second tutorial process of the LSTM Layer uses more than 80 features/attributes for every single (time-)step (sequence analysis).
There you can see the "TimeSeries to Tensor" operator in action which takes care of tensor generation.
But if you'd prefer something else, I can also recommend our "Examples to Tensor" operator, more specifically its tutorial process (important attributes are: BATCH and ID which we use to identify sequences and their steps).


One more thing, please use the latest version(s) of the extension(s) [Deep Learning and ND4J, both 1.0.0].

Kind regards,
Mate

MartinLiebig

Hi @BUTRRRS ,

why exactly do you think that LSTMs are specifically good for this? This looks like 'normal classification' and not Time series forecast, the discipline LSTMs excel in.

Let me also reference @DocMusher here. He is a MD and thus our expert on those topics.

Cheers,

Martin

BUTRRRS

Hi @mschmitz,
I have used other models to predict hypertension. Such as K-NN, Naive Bayes, Decision Tree, Neural Net. They all work fine. 
The reason why I use LSTM or CNN to predict the patient's hypertension is I wanna know it can predict it or not?
I have tried LSTM by myself. Due to data structure, there are some problems. I checked other examples in the Community. For instance, data from online examples use 2 attributes to predicted , like time and price. In my example, it has more than 3 attributes.
This confused me, 

Cheers,
Frank

MartinLiebig

Hi @BUTRRRS ,

LSTMs use Tensors as input, thats there strength but make their handling also more complex. As I said LSTMS can be used for your use case, but are more frequently used for time series forecasting. I am not sure if we have an example for Time Series -> Binomial using LSTMs. Maybe @Mate can help?

Best,
Martin

Mate

Hey @BUTRRRS,

as a first step I would recommend you check our tutorial processes which might give you an idea how to provide data in an appropriate format (if I understand you correctly that's the essence of your question here: how to supply data appropriately right?).
E.g.:
The second tutorial process of the LSTM Layer uses more than 80 features/attributes for every single (time-)step (sequence analysis).
There you can see the "TimeSeries to Tensor" operator in action which takes care of tensor generation.
But if you'd prefer something else, I can also recommend our "Examples to Tensor" operator, more specifically its tutorial process (important attributes are: BATCH and ID which we use to identify sequences and their steps).


One more thing, please use the latest version(s) of the extension(s) [Deep Learning and ND4J, both 1.0.0].

Kind regards,
Mate

BUTRRRS

@Mate
Thank you.
Kind regards