An algorithm for direct causal learning of influences on patient outcomes

Abstract

Objective: This study aims at developing and introducing a new algorithm, called direct causal learner (DCL), for learning the direct causal influences of a single target. We applied it to both simulated and real clinical and genome wide association study (GWAS) datasets and compared its performance to classic causal learning algorithms.

Method: The DCL algorithm learns the causes of a single target from passive data using Bayesian-scoring, instead of using independence checks, and a novel deletion algorithm. We generate 14,400 simulated datasets and measure the number of datasets for which DCL correctly and partially predicts the direct causes. We then compare its performance with the constraint-based path consistency (PC) and conservative PC (CPC) algorithms, the Bayesian-score based fast greedy search (FGS) algorithm, and the partial ancestral graphs algorithm fast causal inference (FCI). In addition, we extend our comparison of all five algorithms to both a real GWAS dataset and real breast cancer datasets over various time-points in order to observe how effective they are at predicting the causal influences of Alzheimer's disease and breast cancer survival.

Results: DCL consistently outperforms FGS, PC, CPC, and FCI in discovering the parents of the target for the datasets simulated using a simple network. Overall, DCL predicts significantly more datasets correctly (McNemar's test significance: p<<0.0001) than any of the other algorithms for these network types. For example, when assessing overall performance (simple and complex network results combined), DCL correctly predicts approximately 1400 more datasets than the top FGS method, 1600 more datasets than the top CPC method, 4500 more datasets than the top PC method, and 5600 more datasets than the top FCI method. Although FGS did correctly predict more datasets than DCL for the complex networks, and DCL correctly predicted only a few more datasets than CPC for these networks, there is no significant difference in performance between these three algorithms for this network type. However, when we use a more continuous measure of accuracy, we find that all the DCL methods are able to better partially predict more direct causes than FGS and CPC for the complex networks. In addition, DCL consistently had faster runtimes than the other algorithms. In the application to the real datasets, DCL identified rs6784615, located on the NISCH gene, and rs10824310, located on the PRKG1 gene, as direct causes of late onset Alzheimer's disease (LOAD) development. In addition, DCL identified ER category as a direct predictor of breast cancer mortality within 5 years, and HER2 status as a direct predictor of 10-year breast cancer mortality. These predictors have been identified in previous studies to have a direct causal relationship with their respective phenotypes, supporting the predictive power of DCL. When the other algorithms discovered predictors from the real datasets, these predictors were either also found by DCL or could not be supported by previous studies.

Conclusion: Our results show that DCL outperforms FGS, PC, CPC, and FCI in almost every case, demonstrating its potential to advance causal learning. Furthermore, our DCL algorithm effectively identifies direct causes in the LOAD and Metabric GWAS datasets, which indicates its potential for clinical applications.

Keywords: Bayesian-score based learning; Causal discovery; Clinical decision support; Constraint-based learning; Predictive medicine; Simulated data.

Figure 1

A BN representing relationships among variables related to respiratory diseases.

Figure 2

Pictorial representation of the BN parameterizations and models. (A) Detailed outline of the 14,400 datasets. There are two different simple and complex networks, with each simple network having six different parameterizations based on the strong-weak schedule (see supplement S1 for more details) and each complex network having six random parameterizations. Furthermore, each model has 6 different case sizes, and each case size has 100 datasets. The simple and complex networks each have a total of 7200 datasets. (B) A sample diagram of a model of one of the simple networks. X and Y are the parents of T and have a strong causal relationship. P1 and P2 are indirect nodes to T and have a weak relationship to their children, X and Y. (C) A randomly parameterized model of one of the complex networks. V, W, X, Y, and Z are the parents of T, and P1-P12 are indirect nodes to T.