0

I've developed a neural network which predicts if someone is diagnosed with cardiovascular disease. Is it possible to find out which feature is the main cause of the prediction? For example, when a person is predicted to have CVD, it should show that his serum sodium level is too high.

Tim
  • 1
  • You seem to be looking for Principal Component Analysis (PCA) – tripleee Mar 03 '22 at 05:27
  • PCA is just a dimensionality reduction technique, it gives you no hints whatsoever about feature importance for a specific prediction of a trained model. At best you can use it to preprocess the features before training. – Edoardo Guerriero Mar 03 '22 at 08:11
  • what you're looking for are a bunch of techniques falling into the macro category of explainable AI, you can check a fair summary of these techniques in this [answer](https://ai.stackexchange.com/questions/12870/which-explainable-artificial-intelligence-techniques-are-there/24138#24138) – Edoardo Guerriero Mar 03 '22 at 08:13
  • In the context of statistics, what you want to do is often called inference, which is different from prediction. Inference is just another term for what many people in AI now call _explainable AI_. – nbro Mar 03 '22 at 10:44

0 Answers0