Technically, it is possible, by one doing the following:

1. constructing predictive models using all statistically significant GWAS hits for each condition / phenotype
2. cross-validating and refining the models on training and testing data
3. making model predictions on new data

Some extra points to consider:

• statistically significant GWAS hits may not necessarily result in disease or confer a particular phenotype; instead they may only increase / decrease risk (that is, to say, that many of these variants have incomplete penetrance)
• getting samples to do this work will be difficult
• 'polygenic risk score' is a generic term and there are many ways to construct these. Most are built from the beta coefficient from the regression model fit
• you should consider how you are going to build and fit the model. Perhaps something along the lines of elastic-net or ridge regression would be a start. Others have use lasso-penalised regression, in the past, to do something similar for breast cancer somatic variants.

Note, that, replacing 'predictive models' with 'AI' or 'machine learning algorithm' will likely increase your chance of funding for the work, if that is ultimately what you want.

Kevin