AI Accelerates Pulmonary Embolism Diagnosis

Using clinical data and chest X-ray features, Italian researchers have developed a machine learning (ML) algorithm that accurately predicts the risk of pulmonary embolism (PE) in emergency patients. Nicoletta Raciti, MD, a second-year radiology resident at Niguarda Metropolitan Hospital in Milan, presented the results of the AI model’s performance Monday afternoon.

“Pulmonary embolism is one of the leading indications for emergency imaging referrals,” Dr. Raciti said. “Chest X-ray is widely available and inexpensive, yet it has low sensitivity and specificity for predicting PE.”

CT pulmonary angiography, the reference standard, directly visualizes thrombi, compared to the indirect signs provided by X-ray. Dr. Raciti’s team developed an ML algorithm to optimize referrals for CT imaging, which can be especially important in emergency settings with limited resources. “The ability for radiologists and emergency physicians to identify patients at higher risk for PE could streamline triage, accelerate the diagnostic workflow and prioritize urgent CT imaging,” she said.

For the study, Dr. Raciti and colleagues retrospectively reviewed the records of adult patients referred to Milan’s Niguarda Metropolitan Hospital in 2023 and 2024 who had undergone both a chest X-ray and a CT scan for suspected pulmonary embolism. The patient population included 255 patients, median age 75 years, of whom 59% were male. Eighty patients (31%) were diagnosed with PE by CT.

Patient data, including chest X-rays and clinical information, was split between training the ML model (85%) using a five-fold cross-validation scheme and testing (15%). Clinical data included age, sex, D-dimer test results, history of cancer, along with the presence of DVT, tachycardia, dyspnea and cough. Chest X-rays were labeled for specific findings including the Westermark sign, Hampton hump, Fleischner sign and pleural effusion.

Findings Expand Body of Evidence for Clinical AI Use

The ML algorithm was evaluated and correctly identified PE cases with an 85% accuracy. It was reliable at ruling out those individuals without the condition, demonstrating 93% specificity and strong overall performance with a 0.92 AUC score. While it caught just 67% of true cases, its balanced F1 score, which measures the performance of a classification model, highlights its effectiveness in real-world emergency settings.

“The advantage of using an AI-based predictive model over traditional risk assessment tools is that it’s free from a rule-based structure and can capture non-linear relationships between clinical data and chest X-ray data,” Dr. Raciti said.

The research team hopes to validate their AI model in a large, prospective, multi-center cohort using unsupervised imaging data.

“This prospective validation is essential to confirm the robustness and clinical applicability of our model,” Dr. Raciti said. “To facilitate adoption in clinical settings, future versions of the model must also integrate with electronic health record systems and provide a user-friendly interface.

“Chest X-ray findings combined with clinical information have a powerful predictive power that allows us to obtain an effective risk stratification and deliver more personalized care,” Dr. Raciti said. “It can streamline, triage, and get faster diagnoses with CT for patients who are at high risk for PE and avoid unnecessary scanning for patients who are at low risk.”

Access the session, “A Prediction Model for Pulmonary Embolism from Clinical and Chest X-Ray Data,” (M7-SSCH04-2) on demand at RSNA.org/MeetingCentral.