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

Area of Science:

• Computational biology
• Medical informatics
• Health services research

Background:

• Understanding the temporal sequence of clinical events is crucial for analyzing healthcare processes and improving patient outcomes.
• Traditional methods for determining event order can be complex and data-intensive.

Purpose of the Study:

• To develop and validate a Bayesian method for detecting temporally ordered pairs of clinical events.
• To assess the efficacy of a novel metric, ConfDiff, in predicting the sequence of medical procedures.

Main Methods:

• Association mining rules were extracted from a database of medical procedures.
• Conditional probabilities P(A|B) and their inverses were calculated for procedure pairs (A,B).
• The ConfDiff metric was computed and correlated with the actual temporal sequence (%Tseq) of procedures.

Main Results:

• ConfDiff was identified as the strongest predictor of event sequence (%Tseq), with a correlation of r=0.278.
• P(B|A) also showed predictive power (r=0.129), but weaker than ConfDiff.
• These associations remained significant even after adjusting for other metrics like confidence and leverage.

Conclusions:

• The findings support the principle that asymmetric associations between attributes in structured domains, like clinical care, can imply temporal ordering.
• The developed Bayesian method and ConfDiff metric offer a novel approach to inferring temporality in clinical event data.
• This methodology has potential applications in healthcare process analysis, clinical trial design, and predictive modeling.