Prediction Of Daily Patient Visits At A Community Health Center Using Long Short-Term Memory (LSTM) With Laplace Differential Privacy

Azzahra Azzahra (Universitas Nusa Putra)
Sudin Saepudin (Department of Information Systems, Faculty of Engineering, Computers and Design, Nusa Putra University, Sukabumi, Indonesi)
Gina Syabani Yuda (Department of Information Systems, Faculty of Engineering, Computers and Design, Nusa Putra University, Sukabumi, Indonesi)

International Journal of Health, Engineering and Technology
Vol. 5 No. 3 (2026): Vol. 5 No. 3 SEPTEMBER 2026

This study aims to analyze the performance of Long Short-Term Memory (LSTM) in predicting daily patient visits using limited-scale Electronic Medical Record (EMR) data and to evaluate the effect of applying Laplace Differential Privacy on prediction accuracy. The data were obtained from the EMR of Puskesmas in Sukabumi Regency, covering the period from January to December 2025. The data were preprocessed through cleaning, aggregation of daily patient visits, exclusion of holidays, and the construction of 14 days sequences. The LSTM model was evaluated using walk-forward validation with an expanding-window scheme consisting of five folds, while privacy protection was implemented using the Laplace mechanism with . The results showed that the baseline model achieved an RMSE of 25.08 and an MAE of 22.19. The application of Differential Privacy increased prediction errors, with higher errors levels observed at smaller  Increasing  from 1.0 to 2.0 resulted in the largest reduction in prediction error, while further increas provided relatively small improvements. Based on these results,  provided the most acceptable balance between privacy protection and prediction performance for the dataset used in this study.