Probabilistic modeling personalized treatment pathways using electronic health records

Area of Science:

• Medical Informatics
• Health Services Research
• Machine Learning in Healthcare

Background:

• Personalized treatment pathway modeling is crucial for improving healthcare delivery.
• Modeling patient pathways is complex due to case-specific variations and poorly understood latent semantics.
• Existing methods lack robust extraction of semantic transitions within patient pathways.

Purpose of the Study:

• To propose an extension of the Hidden Markov Model (HMM) for mining and modeling personalized treatment pathways.
• To extract latent treatment topics and their sequential dependencies from Electronic Health Record (EHR) data.
• To represent pathway semantics using probabilistic distributions and transitions.

Main Methods:

• Utilized an extended Hidden Markov Model (HMM) approach.
• Extracted latent treatment topics and sequential dependencies from raw EHR data.
• Modeled pathways using probabilistic distributions and transitions of treatment events.

Main Results:

• Discovered 15 distinct treatment topics and their transition routes from 1.39 million treatment events across 48,024 cardiovascular disease patients.
• The proposed model demonstrated superior performance compared to Latent Dirichlet Allocation (LDA) and Sequent Naïve Bayes (SNB) models (p<0.01).
• Manual evaluation by clinicians confirmed the model's effectiveness in understanding personalized treatment pathways.

Conclusions:

• The developed model efficiently mines and models personalized treatment pathways using EHR data.
• Discovered treatment topics and transitions offer actionable insights for clinical practice improvement.
• This approach aids physicians in understanding specialties and learning from past patient treatment experiences.