Efficient hospital discharge planning is crucial for maintaining patient flow. At St. Antonius Hospital in the Netherlands, delays in aftercare planning lead to prolonged hospital stays, placing unnecessary burden on resources. This study investigates the use of machine learning to improve predictions of both remaining length of stay and aftercare needs, aiming to align care and transfer processes. The key contribution of this paper is a new method to train prediction models for the remaining length of stay and aftercare needs in such a way that they perform better on clinical impact rather than standard metrics, when evaluated jointly. These predictions are essential for efficient discharge planning. Our findings reveal that models trained using conventional metrics partly fail to reduce these inefficiencies, while those optimized on cost and clinical outcomes reduce prolonged stays.