Knowledge graph-enhanced deep learning for pharmaceutical demand forecasting

Area of Science:

• Health Informatics
• Artificial Intelligence
• Supply Chain Management

Background:

• Pharmaceutical demand forecasting is critical for healthcare supply chain efficiency but challenged by complex, dynamic demand patterns.
• Existing statistical and machine learning models struggle to capture nonlinearities from factors like drug substitutions and seasonal trends.

Purpose of the Study:

• To develop a novel hybrid model integrating pharmaceutical knowledge graphs with deep learning for improved demand forecasting.
• To enhance the accuracy and robustness of pharmaceutical demand prediction in complex healthcare environments.

Main Methods:

• Proposed KG-GCN-LSTM: a hybrid model combining a pharmaceutical knowledge graph (KG) with Graph Convolutional Networks (GCN) and Long Short-Term Memory (LSTM) networks.
• Utilized GCN to extract features from historical drug demand and related drugs in the KG.
• Employed LSTM to capture temporal dynamics in drug demand patterns.

Main Results:

• KG-GCN-LSTM demonstrated superior performance over established benchmarks including ARIMA, SVR, XGBoost, RNN, CNN-LSTM, TimeMixer, and NBEATS.
• Achieved a 3.62% reduction in Symmetric Mean Absolute Percentage Error (SMAPE) compared to NBEATS.
• Performance was comparable to the state-of-the-art TimeMixer model.

Conclusions:

• Knowledge graph-enhanced deep learning, specifically KG-GCN-LSTM, significantly improves pharmaceutical demand forecasting accuracy and robustness.
• The model effectively captures complex, nonlinear demand patterns, offering valuable support for data-driven healthcare supply chain management.