A RECENT study utilised AI to detect pulmonary embolism (PE) early using only patient records.
The study led by Gurion University researchers trained and tested two AI models on data gathered from 2,568 patients with PE and nearly 52,600 controls. The AI models evaluated medical record data including demographics, comorbidities, and medication use, focusing on how effectively these factors could predict PE. Notably, past diagnoses of PE, pneumonia, and deep vein thrombosis were among the key predictors.
The AI models showed impressive performance. The first model, which adjusted for data imbalance, achieved a high accuracy rate of 0.86, while the second model, which used a balanced classifier approach, showed an accuracy of 0.78. Both models had high true-negative and true-positive rates, indicating reliable performance in distinguishing patients with PE from those without.
Moreover, further analysis clustered patients based on various factors, including age, sex, BMI, and medical history, to identify groups with high prevalence of PE. One notable cluster, characterised by a history of PE, pneumonia, and deep vein thrombosis, showed that 63% of individuals within this group had PE. This suggests that patients with specific medical histories, particularly those with past PE or related conditions, may require extra attention upon hospital admission.
The researchers also highlighted the transformative potential of these findings, suggesting that they could eventually be integrated into broader predictive models aimed at foreseeing patient deterioration and assessing disease risk. This could revolutionise how clinicians anticipate and manage conditions such as PE, improving patient outcomes through earlier detection.
Aleksandra Zurowska, EMJ
Reference
Yehuda OB et al. Early Detection of Pulmonary Embolism in a General Patient Population Immediately Upon Hospital Admission Using Machine Learning to Identify New, Unidentified Risk Factors: Model Development Study. J Med Internet Res. 2024;26: e48595.
