Natural Language Processing of Unstructured Healthcare Data for Predicting Heart Failure in Individuals with Type 2...

Area of Science:

• Cardiology
• Endocrinology
• Health Informatics

Background:

• Type 2 diabetes mellitus (T2DM) is a complex disease impacting metabolic, renal, and cardiovascular systems.
• Predicting heart failure (HF) risk in T2DM patients is crucial for proactive management.
• The Diabetic@ project utilizes real-world data from electronic health records (EHRs) for T2DM characterization.

Purpose of the Study:

• To develop a predictive model for two-year heart failure (HF) risk in individuals with Type 2 diabetes mellitus (T2DM).
• To leverage unstructured data from EHRs for enhanced risk stratification.
• To compare the performance of different machine learning models for HF risk prediction.

Main Methods:

• A multicenter, retrospective study of T2DM patients across eight Spanish hospitals (2013-2018).
• Clinical Natural Language Processing (cNLP) was used to extract data from unstructured EHR free text, mapped to SNOMED CT.
• Logistic regression, decision trees, random forest, and XGBoost were evaluated for predictive accuracy and interpretability.

Main Results:

• The study analyzed data from 588,756 T2DM individuals; 14.3% had prevalent HF.
• Logistic regression achieved the best performance (AUC-ROC 0.73) with 27 predictors, and a refined 9-predictor model was developed.
• Reduced models demonstrated similar performance, with a 9-predictor model integrated into a web tool.

Conclusions:

• Unstructured EHR data, processed by cNLP, can effectively develop a two-year HF risk model for T2DM patients.
• This approach facilitates risk stratification across the cardiovascular-renal-metabolic spectrum.
• The developed model and tool offer potential for early HF detection and intervention in T2DM.