Patient-level temporal aggregation for text-based asthma status ascertainment.

Area of Science:

• Clinical informatics
• Natural Language Processing (NLP)

Background:

• Current NLP practices often evaluate data at granular levels (term, sentence, document).
• A patient-level temporal aggregation approach is needed for comprehensive clinical data analysis.

Purpose of the Study:

• To address the challenge of patient-level temporal aggregation from clinical text.
• To introduce and evaluate probabilistic methods for this aggregation task.
• To classify patients by asthma status in retrospective and clinical decision support scenarios.

Main Methods:

• Utilized a pediatric asthma cohort with manual annotations.
• Generated a feature set using standard clinical NLP.
• Introduced six methods for aggregating time-distributed features from document to patient level.
• Applied machine learning for patient classification.

Main Results:

• Achieved solid patient classification performance across settings.
• Sum aggregation yielded the highest F1 score (85.71%) in retrospective epidemiology.
• A probability density function-based method achieved the highest F1 score (74.63%) in clinical decision support.
• Multiple techniques accurately estimated asthma diagnosis dates.

Conclusions:

• Clinical decision support presents a more challenging problem for aggregation methods.
• The study identified strengths of various aggregation algorithms in different contexts.
• Limited manually annotated data influenced the scope of method determination.