A New Natural Language Processing-Inspired Methodology (Detection, Initial Characterization, and Semantic...

Area of Science:

• Health Informatics
• Data Science
• Machine Learning

Background:

• Analyzing health care data is vital for improving patient outcomes and understanding treatment impacts.
• Temporal shifts in health data, like those seen during the COVID-19 pandemic, significantly affect patient characteristics and outcomes.
• Understanding these temporal dynamics is crucial for accurate health care analysis.

Purpose of the Study:

• To propose a novel methodology, Detection, Initial Characterization, and Semantic Characterization (DIS), for analyzing temporal changes in health outcomes and variables.
• To discover contextual changes within large health care datasets over time.
• To identify factors driving changes in patient outcomes by analyzing temporal data shifts.

Main Methods:

• The DIS methodology integrates three steps: detection of data drifts, initial characterization of data distribution changes, and semantic characterization using NLP-inspired techniques.
• Jensen-Shannon divergence is employed for effective detection of significant data drifts.
• The approach combines machine learning and natural language processing to analyze temporal shifts in health care data.

Main Results:

• The DIS methodology successfully identified data drifts and uncovered reasons for mortality declines in COVID-19 and MIMIC-IV datasets.
• Key factors influencing outcomes, such as vaccination and reduced iatrogenic events, were highlighted.
• Shifts in patient demographics and disease patterns were pinpointed, offering insights into the evolving health care landscape.

Conclusions:

• A novel methodology combining machine learning and NLP was developed to detect, characterize, and understand temporal shifts in health care data.
• This understanding enhances predictive algorithms, improves patient outcomes, and optimizes health care resource allocation.
• The DIS methodology is applicable to various health care data analysis scenarios beyond the studied datasets.