Objective: Recent advancements in artificial intelligence have created new opportunities in social insurance, enabling the provision of tailored services based on novel knowledge to the insured. However, the social insurance industry faces significant challenges in utilizing AI, including heterogeneous data, imbalanced data distribution across prediction classes, low assignment rates of data points to specific categories, and the presence of numerous features in a process. This study aims to present an efficient method for predicting social insurance processes that addresses the challenges of heterogeneous, imbalanced, and high-dimensional data to deliver more accurate and optimized prediction results.
Method: This research proposes a novel approach based on recurrent deep learning networks with a Long Short-Term Memory (LSTM) architecture. The method incorporates data preprocessing stages and segmentation of data into unit categories (A) to overcome existing challenges. The main objective is to improve prediction accuracy of social insurance processes with reduced computational cost, relying on historical process records.
Findings: The proposed method was simulated using real data from the Iranian Social Security Organization. Results indicate that while the approach slightly increases resource consumption, including memory usage and CPU utilization, it significantly reduces prediction errors compared to two benchmark methods. Moreover, error reduction was achieved even with fewer data samples during the deep learning training phase, highlighting another advantage of the proposed method.
Conclusion: The presented method substantially enhances prediction accuracy and accelerates error reduction while maintaining relative savings in computational resources. Therefore, this model represents an effective and suitable option for predicting social insurance processes, with strong potential for deployment in insurance organizations to enable intelligent automation and improve operational performance.