Microsimulation embedding for the disease progression network

Abstract Background Recently, numerous disease network models have been developed using large-scale health utilization data. One widely used method for representing these networks as knowledge graphs is the random walk approach. However, random walks assume mutually exclusive transitions, making them inadequate for capturing multimorbidity and the inherent complexity of disease trajectories. To overcome this limitation, we propose a microsimulation-based method to extract disease trajectories and generate embeddings for disease network applications. Methods Using the previously developed disease progression network built on data from 50 million individuals in the Korean population, we used transition probabilities to model the progression from one disease to another. From each disease node, we simulated 100 trajectories, each consisting of 100 disease walks. These trajectories were embedded using Word2Vec. To evaluate the utility of the embeddings, we performed next-disease prediction using a Transformer model trained on Korea Health Panel Survey data from 12,746 individuals. For comparison, we also assessed embeddings generated directly from raw data and from standard random walks on the disease network. Results Embeddings derived directly from the data and via random walk both showed signs of overfitting, with minimum losses of 4.0521 and 3.8339, respectively. In contrast, the microsimulation-based embeddings achieved superior performance with a lower minimum loss of 3.8009. Furthermore, when visualized using the embedding projector, the microsimulation embeddings revealed more coherent and meaningful disease groupings. Conclusions These findings suggest that, from the perspective of pretrained foundation models, incorporating disease networks as knowledge graphs can enhance predictive performance in various prediction studies that utilize ICD-10 disease codes. Key messages • The disease networks can work like a knowledge graph to represent the relationship among diseases. • Microsimulation embedding can be effectively used when embedding disease networks.