Direct Effects Multiplicative Inference (DEMI): causal prediction from pairwise clinical relationships, applied to...

This paper introduces Direct Effects Multiplicative Inference (DEMI), a causal algorithm that reasons through a medical knowledgebase and predicts clinical outcomes from direct, unmediated effects. This study compares performance of DEMI to logistic regression, a common inference mechanism available for analysis of patient-level tabular data.

We extracted 430 million relationships between pairs of medical concepts from the All of Us database. DEMI uses this knowledgebase to predict likelihood of outcomes, in three steps: (1) patient features are temporally ordered using pairwise precedence counts; (2) each feature's total effect on the outcome is recursively decomposed into direct and indirect (mediated) components by working backward through the temporal order; (3) direct effects are aggregated multiplicatively to estimate the odds of outcomes. DEMI uses the distribution of a pair of variables, as reported in the medical knowledgebase. Logistic regression uses patient-level tabular data, which includes the joint distribution of all variables. Performance was assessed using Area under the Receiver Operating Characteristic curve (AUC).

DEMI predicted response to 15 antidepressants with a mean AUC of 0.934 (range, 0.512-0.996). Logistic regression achieved similar discrimination (mean AUC, 0.935; range, 0.512-1.000). At α = 0.05, the two methods did not differ significantly. DEMI achieved comparable discrimination using precomputed pairwise statistics without access to the joint distribution of patient features. Under the same variable-selection rule, DEMI could use on average 113 more predictors per task than logistic regression.

DEMI predicted antidepressant response from a medical knowledgebase as well as logistic regression did from patient-level data. Under its structural assumptions (no unmeasured confounding and correct temporal ordering), it does so by isolating direct, unmediated predictors and excluding effects that operate through intermediate variables.