Detection of Temporal Clinical Events in Non-Temporal, Non-Annotated Data

Zikos, Dimitrios; Eappen, Philip; Schmidt, Ryan N.

doi:10.3233/SHTI250038

Abstract

Introduction:

This study developed and validated a Bayesian method which detects pairs of clinical events likely to be temporally ordered.

Methods:

Association mining rules between medical procedures were extracted into a research-generated database and, for each pair of procedures A,B, the conditional probability P(A|B), its inverse, and the difference from its inverse (ConfDiff) were calculated. The study hypothesized that the higher the ConfDiff is, the more likely it is for A and B to be temporally ordered. The actual calendar date of each medical procedure served as ground truth.

Results:

ConfDiff is the strongest predictor of %Tseq (r=0.278), followed by P(B|A) (r=0.129). This association continued to be present after controlling for the confidence, leverage and conviction metrics.

Conclusion:

Findings substantiate the assumption that, in a structured process-based domain (e.g., clinical care) if an attribute is strongly associated with another one, but not the other way around, this could imply temporality.

Introduction:

Methods:

Results:

Conclusion:

This website uses cookies

This website uses cookies