Pneumonia and Pulmonary Thromboembolism Classification Using Electronic Health Records.

Area of Science:

• Medical Informatics
• Computational Medicine
• Artificial Intelligence in Healthcare

Background:

• Pneumonia and pulmonary thromboembolism (PTE) share similar symptoms, complicating diagnosis.
• Accurate diagnosis requires integrating diverse clinical data, posing challenges for healthcare professionals.
• Existing research lacks analysis of clinical data for the Mexican population in Spanish.

Purpose of the Study:

• To develop and evaluate machine learning models for the automatic diagnosis of pneumonia and PTE.
• To compare the performance of models using structured (laboratory) and unstructured (clinical text) data.
• To identify key diagnostic attributes from clinical data for these respiratory diseases.

Main Methods:

• Utilized a cohort of 173 clinical records from the Mexican Social Security Institute.
• Applied machine learning algorithms (Naïve Bayes, SVM, Decision Trees, etc.) to structured and unstructured electronic health record data.
• Employed cross-validation techniques (K-fold, Leave-one-out) and hold-out for model evaluation.

Main Results:

• Naïve Bayes achieved the highest AUC-ROC for structured data, particularly in differentiating PTE vs. pneumonia (87.0%).
• Models using structured data outperformed those using unstructured data (BiLSTM).
• Decision trees identified important attributes for diagnosing PTE vs. pneumonia.

Conclusions:

• Machine learning, especially Naïve Bayes with structured data, shows promise for automated diagnosis of pneumonia and PTE.
• Structured clinical data is more effective than unstructured text for differentiating these conditions.
• Decision trees offer insights into diagnostic criteria for medical staff.