This paper proposes a novel online learning method for medical insurance fraud detection called Pretrained Online Contrastive Learning (POCL). The method combines contrastive learning pre-training with online updating strategies to address the challenge of evolving fraud patterns. In the pre-training stage, contrastive learning is used to learn deep features from historical data, enabling rich risk representations. In the online learning stage, a Temporal Memory Aware Synapses (MAS) strategy is adopted to allow the model to perform incremental learning and optimization based on new data, ensuring timely adaptation to fraud patterns and reducing forgetting of past knowledge. The model undergoes extensive experiments on real-world insurance fraud datasets, demonstrating significant advantages in accuracy compared to state-of-the-art baseline methods, while also exhibiting lower running time and space consumption. The model's contributions include being the first to introduce an online learning model in insurance fraud detection, combining structural features with continual adaptation to fraud patterns, and proposing a novel online learning GNN model based on contrastive learning pre-training that accurately identifies fraudulent claims and significantly reduces forgetting of previously learned knowledge. The model is evaluated on a real-world medical insurance fraud dataset and other fraud datasets, showing high accuracy and performance in long-term online learning scenarios. It outperforms other models in terms of average monthly accuracy and average accuracy decline rate, and demonstrates significant time and space efficiency. The model's effectiveness is further validated through ablation studies and case studies, showing its ability to adapt to evolving fraud patterns and reduce catastrophic forgetting. The model's pre-trained contrastive learning and Temporal MAS method enable it to maintain high performance while reducing the impact of forgetting, making it an effective solution for medical insurance fraud detection.This paper proposes a novel online learning method for medical insurance fraud detection called Pretrained Online Contrastive Learning (POCL). The method combines contrastive learning pre-training with online updating strategies to address the challenge of evolving fraud patterns. In the pre-training stage, contrastive learning is used to learn deep features from historical data, enabling rich risk representations. In the online learning stage, a Temporal Memory Aware Synapses (MAS) strategy is adopted to allow the model to perform incremental learning and optimization based on new data, ensuring timely adaptation to fraud patterns and reducing forgetting of past knowledge. The model undergoes extensive experiments on real-world insurance fraud datasets, demonstrating significant advantages in accuracy compared to state-of-the-art baseline methods, while also exhibiting lower running time and space consumption. The model's contributions include being the first to introduce an online learning model in insurance fraud detection, combining structural features with continual adaptation to fraud patterns, and proposing a novel online learning GNN model based on contrastive learning pre-training that accurately identifies fraudulent claims and significantly reduces forgetting of previously learned knowledge. The model is evaluated on a real-world medical insurance fraud dataset and other fraud datasets, showing high accuracy and performance in long-term online learning scenarios. It outperforms other models in terms of average monthly accuracy and average accuracy decline rate, and demonstrates significant time and space efficiency. The model's effectiveness is further validated through ablation studies and case studies, showing its ability to adapt to evolving fraud patterns and reduce catastrophic forgetting. The model's pre-trained contrastive learning and Temporal MAS method enable it to maintain high performance while reducing the impact of forgetting, making it an effective solution for medical insurance fraud detection.