Fairness and bias correction in machine learning for depression prediction across four study populations