Kindly mention the step wise approach for the applicability of HMM for the classification of a disease. Lets suppose the dataset is having 200 patients and 150 healthy people.
Hello Aman,
Before create your HMM, you need to specify your goal and your variables.
e.g., Model aims to... "Represent the evolution of the disease in time" "Get the probability of acquire this disease in some conditions" Etc.
Variables... How many variables have influence in this disease? Diet, environment, etc.
Dear Joaquim.
Goal is to classify a disease. We have 50 attributes. How to get the probability is the main concern to.
Ok, if I understand your data, it is something like
attr1, attr2, ... attrN, disease
... , .... , ... , ...
Unless you are more interested in HMMs than the classification itself, I do not recommend the use of HMM in this case.
First, to achieve a minimum precision, your model will have a huge state space; second, your precision will not be as good as some data mining algorithms, e.g., Random Forests.
I suggest you to explore some classification algorithms. You can use Weka, it is user friendly and they have a good documentation.
http://www.cs.waikato.ac.nz/ml/weka/
...
However If you need a stochastic model, try use something more robust, like SANs.
You can directly generate models using this tool: ( http://joaquim.pro.br/Software_acad.htm#SANGE )
The main paper, about the tool, is here ( http://joaquim.pro.br/pdf/SANGE.pdf )
Do not hesitate to ask me any questions in this concern.
Hi,
First, I recommend to try to classify the patients based on only one attribute in order to determine which attribute are the most relevant. Then, pick up a subset of the attribute that better specify the presence of the disease.
I don't a Markov method will be very useful unless you specify what you use Markov for? What are the cirrelation you'd like to model here? I would rather think about a multimodal classification problem. For this purpose, please refer to Thierry Desnoeux works on the subject.
Regards,
idea is to train one HMM for each class (what ever the class is). Here is an example for phoneme classification :
https://www.researchgate.net/publication/283889476_A_Doubly_Hierarchical_Dirichlet_Process_Hidden_Markov_Model_with_a_Non-Ergodic_Structure
Article A Doubly Hierarchical Dirichlet Process Hidden Markov Model ...
Hello. I share the view of many of the respondents. I am not sure that HMM is the right framework for classification since your system is not really "dynamical", i.e. you do not really need to model time dependence and therefore don't need a Markov model. On the other hand, the EM framework of the HMM can be used to learn a mixture model of your data that can supply classification probabilities. Or you can try any one of a number of other established classification methods, like artificial neural networks, SVM, decision trees/random forests, etc. Given your quite small data set (200 patients), you might want to try to extract the essential features and minimise the dimensionality of the learning problem. You also need to be careful about the weight assigned to each patient class as it seems you have 3 times as many healthy patients as unhealthy ones.
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis