A Bayesian method for the automatic extraction of meaningful clinical sequences from large clinical databases

Area of Science:

• Medical Informatics
• Computational Linguistics
• Health Data Science

Background:

• Clinical event recognition aids in analyzing negative hospital outcomes and supports medical education by helping students identify frequent clinical events.
• Automated extraction of clinical events is crucial for large-scale health data analysis.

Purpose of the Study:

• To develop a non-annotated, Bayes-based algorithm for extracting meaningful clinical events from electronic health records.
• To enable automated identification of clinical event sequences without requiring manual data labeling.

Main Methods:

• Utilized subsets of MIMIC and CMS LDS datasets with respiratory diagnoses.
• Calculated two-itemset rules to build clinical event sequences based on increasing conditional probability and positive certainty factor.
• Defined clinical events as collections of event blocks meeting specific probability criteria; sequences validated by physicians.

Main Results:

• The developed algorithm's rules were rated higher by medical experts compared to random Apriori rules.
• A graphical user interface (GUI) was created to explore associations between clinical events and outcomes like length of stay, mortality, and charges.
• The algorithm demonstrated success in identifying rule blocks that construct accurate clinical event narratives.

Conclusions:

• Presents a novel approach for automated clinical event sequence extraction, eliminating the need for user annotation.
• The algorithm effectively identifies clinically relevant event sequences, enhancing the understanding of patient journeys and outcomes.
• This method offers a scalable solution for uncovering patterns in complex medical data.