Graph autoencoder based causal discovery method for type 2 diabetes mellitus risk factors

The risk factors for Type 2 Diabetes Mellitus (T2DM) encompass biochemical indicators (such as glycated hemoglobin, high-density lipoprotein, etc.) and physiological indicators (such as age, weight, etc.). The interactions among these indicators influence the risk and progression of T2DM. Uncovering their causal relationships is a crucial means for T2DM prevention and intervention. Addressing the issues of severe coupling among T2DM risk factors, small sample sizes, and the erroneous causal relationships arising from existing nonlinear causal discovery methods struggling to satisfy the acyclicity assumption, this paper proposes a Graph Autoencoder (GAE)-based causal discovery method tailored for T2DM risk factors. Firstly, by integrating a data generation model with the Graph Autoencoder (GAE), a candidate space for the nonlinear causal structure of risk factors is constructed. Secondly, the problem of optimizing the causal structure is transformed into a task of minimizing the GAE reconstruction error. Utilizing the augmented Lagrangian method and innovatively introducing a bidirectional edge penalty term, a new scoring function is established to avoid generating cyclic causal graphs. Thirdly, the function parameters are iteratively optimized via gradient descent until converging to the optimal causal structure within the candidate space. Finally, validation experiments were designed using two types of datasets. The results show that on a dataset with a known true Directed Acyclic Graph (DAG), the proposed method achieves an accuracy of up to 80%, with optimal comprehensive metrics compared to other methods. On three diabetes datasets, the proposed method achieves an average plausibility rate of 82.8%, significantly outperforming comparative methods and enabling the discovery of more potential causal relationships. The method proposed in this paper can provide a reliable information-assisted tool for research into the pathological mechanisms of T2DM.