Early Detection of Pulmonary Embolism in a General Patient Population Immediately Upon Hospital Admission Using...

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Pulmonary Medicine

Background:

• Pulmonary embolism (PE) poses a significant threat due to challenges in timely identification.
• Late diagnosis of PE remains a critical issue in modern medicine.

Purpose of the Study:

• To develop accurate machine learning (ML) models for early identification of high-risk pulmonary embolism (PE) patients.
• To utilize only pre-admission medical record data for PE risk assessment.

Main Methods:

• A random forest ML algorithm was trained on demographics, comorbidities, and medications data.
• The study included 2568 PE patients and 52,598 controls, focusing on data available before emergency department admission.
• Specialized ML methods were employed to address data imbalance between PE and non-PE cases.

Main Results:

• Models predicted PE with 80% geometric mean accuracy using factors like age, BMI, past PE events, and anticoagulants.
• Clustering identified subgroups with over 61% PE prevalence, revealing associations with deep vein thrombosis and pneumonia.
• New risk factors, including previous pulmonary disease, were identified in general populations.

Conclusions:

• An ML tool enables early PE diagnosis upon hospital admission using historical medical data.
• The models accurately identified high-risk PE patients pre-clinically, even in imbalanced datasets.
• This approach allows for the discovery of novel PE risk factors beyond established scoring systems.