Is the proposed neural network structure correct?

Question

Neural networks

The task is to recognize the type of diabetes by blood analysis and anthropometric parameters of a person.
Together with scientific The leader sketched out the following diagram of a neural network:

Neural network architecture

The essence is as follows: Red neuron - the results of the questionnaire are submitted to it
Green - the results of a blood test
Blue and Yellow - the results of checking risk factors (blue for Male, Yellow - for female)
The next layer for the upper black neuron is served with the results from the previous layer , only those that belong to the Male gender, respectively, to the lower one - for the female. The last layer - displays the result
I would like to ask for advice on how true this structure is and, maybe, advice on how to improve or rework it.
Thanks in advance!

Answer 1

I'll ask the question - why do you need a neural network? The fact is that your task is a CLASSICAL classification task, which reduces almost all tasks of medical diagnostics up to and including the famous IBM Watson.
In reality, a neural network is needed when classical methods do not work, for example, with very complex and intricate relationships between input parameters (as in visual image recognition or text meaning recognition), and another important condition is the presence of really very, very a very, very large amount of input data (from tens of thousands to millions). And only under these conditions, NNs are able to give at least some gain in accuracy, in comparison with the classics. The first is definitely not in your case, and the second is very likely not.
In addition, let me remind you that in medicine, no one will ever accept the result of your research if you are not able to interpret it (the term "evidence-based medicine" I hope you are familiar with, as well as its basic requirements). NNs do not have this property, classical algorithms do.
So bring at least one argument - except for fashion - in favor of the National Assembly.
If you still decide to use traditional machine learning algorithms, you can think and discuss which class of algorithms is better to use and how to implement it.
Good luck.

Answer 2

Determine the quality metric and experiment, science has not yet reached the architecture of neural networks, so the experiment is the main criterion :)
PS: I recommend using pytorch for experiments.

Answer 3

>only those that belong to the male sex, respectively, to the lower one - for the female.
the whole point of machine learning is that the machine itself looks for what signs to what and how they relate.
the fact that you already decide for her how your signs will work, this is no longer machine learning - this is heuristic programming.
Want machine learning - fully connect layers to each other

Answer 4

1. Two or four neurons in one layer is very sparingly. If the training will be on the first Pentium, then it may be
   
2. There is no need to restrict the network topology so explicitly. In the course of training, the weights of unnecessary connections will automatically go to 0.
   
3. See other machine learning algorithms like gradient boosting (classic implementation - xgboost library)