You have it backwards. First you get a problem then you get some way to answer it. David Booth

Let me clarify the situation:

Training: you find a good PCA A to reduce 10 features to 3 , then build a good lung cancer predict model B with these 3 features.

Testing: For the patient, you could use your PCA A to tranform his 10 features to 3 , then use model B to predict lung cancer result.

@Chris Tsai thank you for your answer however, could you please describe little more on how could we apply PCA on a single patient's feature?

For instance, when patient with (1×10) features comes to the deployed model how PCA could reduce it to (1×3) for prediction??

In Python, after fitting PCA with components = 3 for training data (nx10) , call PCA's tranform function to transform test data (mx10) into 3 features.

@Chris Tsai I wanted to ask just about one thing that how could we apply this for only 1 patient who has (1×10) features not (m×10) features. just imagine when a patient comes to know his lung cancer probability from your system, do we consider m or just 1?

N.B. I am not asking about training testing the model. Imagine it’s done perfectly and you want to deploy the trained model in real life hospital.