Double-Rule Causal Feature Selection in Risk Factor Mining for Disease Diagnosis

In healthcare, high-dimensional medical observation data brings challenges to disease diagnosis. In order to reduce the data dimension, feature selection becomes a key step in processing medical observation data. Causal feature selection, i.e., Markov Blanket (MB) discovering, discovers local causal relationships that exist around the target feature, and studies have shown that causal feature selection improves the interpretability and robustness of disease diagnosis models. However, in practical applications, the causal feature selection algorithm based on conditional independence (CI) tests always suffers from two types of errors. These errors include false positives where some incorrect MB features are selected, and false negatives where some true MB features are discarded. Existing algorithms cannot efficiently handle both types of errors at the same time. To address this issue, this paper presented a simultaneous-fusion-of-double-rules MB learning (SF-DRMB) algorithm for risk factor mining of disease. The algorithm is structured into three phases: in the first phase, SF-DRMB identifies both parent-child and spouse sets of the target feature; in the second phase, OR rule is used to solve false negatives error; in the third phase, an improved AND rule is used to solve false positives error. Through the three phases of SF-DRMB, the MB of the target feature is discovered. Finally, SF-DRMB was compared with existing causal feature selection algorithms with benchmark BN datasets and medical datasets, and the experimental results demonstrated that SF-DRMB is competitive with the fastest MB discovery algorithm in terms of computational efficiency and achieves good accuracy among the compared algorithms.