Statistical inference for natural language processing algorithms with a demonstration using type 2 diabetes...

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Health Informatics

Background:

• Pointwise Mutual Information (PMI) is crucial for natural language processing (NLP) algorithms like word2vec, uncovering semantic word relationships.
• Applications include document indexing, topic analysis, and categorization, highlighting the importance of understanding word co-occurrence.
• Existing methods for PMI estimation and its application in healthcare analytics require refinement, particularly concerning patient-specific data patterns.

Purpose of the Study:

• To theoretically demonstrate the relationship between PMI and word2vec using probability theory.
• To present a simplified method for modeling and estimating PMI.
• To develop standard error estimates accounting for within-patient clustering in health records and apply PMI for disease prediction.

Main Methods:

• Utilized probability theory to establish a theoretical link between PMI and word2vec.
• Developed a straightforward approach for modeling and estimating PMI.
• Incorporated methods for standard error estimation addressing within-patient clustering in electronic health records (EHRs).
• Applied the enhanced PMI methodology to predict Type 2 Diabetes Mellitus from free-text EHR notes.

Main Results:

• Established a clear theoretical connection between PMI and word2vec.
• Demonstrated a simplified and effective method for PMI estimation.
• Successfully applied the developed methods to predict Type 2 Diabetes Mellitus from over 400,000 clinical notes.
• Showcased the utility of PMI in identifying diseases from unstructured EHR data.

Conclusions:

• The study provides a robust theoretical and practical framework for PMI estimation and application in NLP.
• The developed methods offer a significant advancement in leveraging semantic relationships for disease prediction from EHRs.
• PMI, when appropriately modeled and estimated, is a valuable tool for clinical informatics and predictive health analytics.