Comparative Analysis of LSTM Architectures for BPJS Drug Expenditure Forecasting Using Walk-Forward Validation

Rizkinawati . (Universitas Nahdlatul Ulama Sunan Giri)
Ifnu Wisma Dwi Prasetya (Universitas Nahdlatul Ulama Sunan Giri Bojonegoro)
Nur Mahmudah (Universitas Nahdlatul Ulama Sunan Giri Bojonegoro)

JOURNAL OF APPLIED INFORMATICS AND COMPUTING
Vol. 10 No. 3 (2026): June 2026

Drug demand planning in health facilities collaborating with BPJS Kesehatan is an important aspect in maintaining drug availability and improving service efficiency. However, drug expenditure data generally forms complex time series patterns that are fluctuating, nonlinear, and influenced by trend and seasonal components, making them difficult to model using conventional forecasting methods. Therefore, this study aims to compare the performance of Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM), and Stacked LSTM models in forecasting BPJS patient drug expenditure by medication type at Basuki Rahmat Pharmacy using Walk-Forward Validation. The dataset used consists of monthly drug expenditure transaction data from January 2023 to May 2025 covering 41 types of drugs. The data preprocessing stages include data cleaning, transformation into time series format, Min-Max normalization, and windowing for input-output sequence generation. Time series characteristic analysis was conducted using the Augmented Dickey-Fuller (ADF) test, trend analysis, and seasonality analysis. The results showed that most drug data were stationary with p-values below 0.05, although several drugs still exhibited non-stationary patterns requiring additional transformation. Trend analysis indicated both increasing and decreasing consumption patterns, while seasonality analysis showed that all drug data exhibited seasonal patterns. The forecasting models were evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). Experimental results using Walk-Forward Validation showed that the Bidirectional LSTM model achieved the best forecasting performance with MAE of 51.99, RMSE of 72.06, and MAPE of 402.90, outperforming Single LSTM and Stacked LSTM models. These findings indicate that Bidirectional LSTM is more effective in capturing complex temporal dependencies in BPJS drug expenditure data and has potential to support decision-making in drug inventory management within healthcare facilities.