The need to protect data privacy is critical in industries like healthcare, finance, and banking. Federated Learning (FL) has attracted widespread attention due to its decentralized, distributed training and the potential to protect privacy. However, FL presents challenges like privacy concerns, network, and data heterogeneity. This motivated us to propose a two-phase, one-pass FL algorithm, where (I) In the first phase, the differentially private dataset is generated by incorporating the Laplacian mechanism, and (ii) during the second phase, we employed one-pass privacy-preserving federated extreme learning machine (OP-FedELM) for generating globally shared training dataset and build the global model. ELM generally tends to overfit training data at a client. Therefore, we employed a modified version of the Evolving clustering algorithm (ECM), an online one-pass clustering algorithm, to cluster training datasets at the clients. Further, it will reduce the computational training time of ELM and communication overhead. At the server, we employed a meta-clustering algorithm to cluster the updates from the clients. We also proposed another one-shot FL algorithm called privacy-preserving federated ELM based on minimum least square solution (FedELM-LS). The experimental analysis concludes that OP-FedELM is computationally less expensive yet achieves higher AUC in three of four datasets.