Forecasting Inpatient Visits of Diabetes Mellitus Using ARIMA

Abstract

Background: Diabetes Mellitus (DM) is a primary non‑communicable disease with a rising global burden. In Indonesia, inpatient admissions for diabetic complications are increasing, placing pressure on hospital resources. Reliable forecasting of patient visits is crucial for planning manpower, bed capacity, and supply chains.

Objective: To model and forecast monthly inpatient visits for diabetes mellitus at RSUD Banjar for 2025–2027 using Autoregressive Integrated Moving‑Average (ARIMA) time‑series models.

Methods: A retrospective time‑series study analyzed monthly inpatient visits for diabetes mellitus, both without and with complications, recorded from January 2020 to December 2024 (60 observations per series). Time‑series plots were inspected to identify trends or seasonality. The augmented Dickey–Fuller test was used to assess stationarity. The Auto Correlation Function (ACF) and Partial Autocorrelation Function (PACF) were reviewed to specify candidate ARIMA models. Missing observations were not present; potential outliers were screened using boxplots and retained because they represented actual clinical events. Candidate models were compared using the Akaike Information Criterion (AIC), the Bayesian Information Criterion (BIC), and the Root Mean Squared Error (RMSE). Residuals were examined using the Ljung–Box Q test to verify independence. Forecasts with 95% confidence intervals were generated for 2025–2027.

Results: The uncomplicated DM series comprised ≈ 240 inpatient visits, whereas the complicated series comprised ≈ 320 visits. An ARIMA (1,0,1) model was selected for uncomplicated DM, and an ARIMA (1,1,0) model for complicated DM, based on AIC/BIC and residual diagnostics. Both models produced white‑noise residuals (Ljung–Box p>0.05). Forecasts suggested gradual increases for both series; uncomplicated visits were expected to rise from about 85 in 2025 to 95 in 2027, and complicated visits from 105 to 115. The upward trends were small and not statistically significant. Forecast confidence intervals indicated a margin of error of approximately ± 10% of the predicted values.

Conclusion: ARIMA models provided reasonable short‑term forecasts of inpatient visits for diabetes at RSUD Banjar. The predicted increases underscore the need for proactive planning of staffing, bed capacity and procurement. Future work should explore ARIMAX or hybrid ARIMA–machine‑learning models to improve predictive accuracy.