This study presents a privacy-preserving protocol for liver disease prediction using a feedforward neural network (FNN) within a client-server architecture. The client, whether an individual or an organization, stores sensitive data in its raw, unencrypted form, while a cloud-based server hosts a deep learning model trained to predict liver disease outcomes based on the client’s input data. To secure sensitive data during the prediction process, we employed the CKKS homomorphic encryption scheme, carefully selecting encryption configurations to balance precision, noise management, and computational feasibility. The protocol ensures that both the client’s raw data and the server’s model remain confidential throughout the process. The server performs linear matrix operations on encrypted data sent by the client and delegates non-linear activation function calculations to the client, with added noise on the data sent to the client to prevent model inference. Further, to enhance model performance, we utilized a random search algorithm for hyperparameter tuning, optimizing parameters like the number of layers, activation functions, and the learning rate. Our protocol was validated through a series of experiments, achieving an average inference (prediction) time of 89.61 seconds. Despite an increase in computational time, the protocol maintained a prediction accuracy of 88.8%, only marginally lower than the 89.2% accuracy of the unencrypted model. These results demonstrate that the proposed protocol effectively balances security, accuracy, and computational efficiency, making it a suitable solution for non-time sensitive applications like privacy-preserving liver disease prediction.