Towards Unsupervised Detection of Process Models in Healthcare

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Process Mining

Background:

• Healthcare processes are complex and variable, making analysis difficult.
• Traditional process mining often results in overly complex models ('spaghetti models').
• Significant domain expert input is typically required to refine these models.

Purpose of the Study:

• To explore the use of Hidden Markov Models (HMMs) for unsupervised detection of hidden healthcare sub-processes.
• To refine complex process mining models by reducing their complexity.
• To enrich event logs with HMM-derived states for improved process analysis.

Main Methods:

• Utilized Hidden Markov Models (HMMs), specifically the Viterbi algorithm, for unsupervised sub-process detection.
• Enriched electronic health record (EHR) event data with HMM-derived states.
• Remodeled healthcare processes as state transitions using process mining tools.
• Applied the method to 'Altered Mental Status' patient data from the MIMIC-III database.

Main Results:

• Successfully reduced the complexity of healthcare process models.
• Detected several hidden healthcare sub-processes without domain expert supervision.
• Demonstrated the effectiveness of HMMs in enhancing process mining for patient pathways.
• The approach yielded promising results for analyzing complex patient data.

Conclusions:

• Hidden Markov Models offer a powerful unsupervised method for refining healthcare process mining.
• This approach can significantly reduce model complexity and uncover hidden processes.
• The method shows potential for improving the analysis of patient pathways in electronic health records.