Traditionally, machine learning models in healthcare rely on centralized strategies using raw data. This poses limitations due to the amount of available data, which becomes hard to aggregate due to privacy concerns. Federated learning has been emerging as a new paradigm to improve model performance. It exploits information on the parameters from other clients while never sharing personal data from the patients. We present a proof-of-concept of federated learning techniques in the case of an automated screening tool for frailty in the older population. We used a frailty-specific dataset called FRELSA, with patients from nine regions of the UK used to simulate a scenario with regional hospitals. We compared three different strategies: separate regional training with no communication; federated averaging, the most widely used strategy for healthcare; and finally, global training on the full dataset for comparison. All three strategies were validated with two architectures: logistic regression and a neural network. Results show that federated strategies outperform local training and achieve global-like performance while preserving patient privacy. For Logistic Regression, the global validation F-score was 0.737 and the federated aggregated score was 0.735, offering improvement in seven of the nine regions. For Multi Layer Perceptron, the global validation F-score was 0.843 and the federated aggregated score was 0.834, improving in all nine regional models. The federated strategy is equivalent to pooling all the data together while avoiding all complications related to data privacy and sharing. The results of this study show that the proposed strategy is a viable method for improving frailty screening in healthcare systems.