Finding falls in ambulatory care clinical documents using statistical text mining.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Clinical Data Mining

Background:

• Identifying patient falls in clinical text is crucial for improving healthcare safety and surveillance.
• Traditional methods for fall detection may not fully leverage the rich information within unstructured clinical notes.

Purpose of the Study:

• To evaluate the performance of statistical text mining (STM) models in detecting falls within clinical text from ambulatory encounters.
• To assess the generalizability of these models across different clinical sites and document types.

Main Methods:

• Utilized a dataset of 2241 patients with fall-related diagnosis codes from the Veterans Health Administration.
• Extracted and annotated clinical documents within a 48-hour window of the diagnosis code.
• Trained and tested three STM models: logistic regression, support vector machine (SVM), and cost-sensitive SVM (SVM-cost).

Main Results:

• All STM models achieved high performance, with Area Under the ROC Curve (AUC) scores exceeding 0.950 across test datasets.
• The SVM-cost model demonstrated superior performance, with AUC scores ranging from 0.953 to 0.978.
• The SVM-cost model achieved excellent sensitivity (0.890-0.931) and specificity (0.877-0.944).

Conclusions:

• STM models are highly effective in identifying falls within diverse clinical text, performing well across various document titles.
• The models demonstrated robust generalization capabilities, even across sites with different linguistic patterns.
• STM-based models hold significant potential for enhancing fall surveillance and can be applied to other clinical surveillance topics.