The Use of AI-Driven Algorithms in Detecting Pulmonary Embolism

Pulmonary embolism (PE) remains one of the most dangerous cardiovascular emergencies in clinical practice. Each year, PE affects an estimated 300,000 to 600,000 patients in the United States alone, contributing to a significant number of preventable deaths. The condition occurs when a blood clot—most often from a deep leg vein—travels through the bloodstream and lodges in the pulmonary arteries, obstructing blood flow to the lungs. Without rapid intervention, this blockage can lead to respiratory failure, cardiac arrest, and death. Early detection is the single most powerful lever to improve outcomes, yet PE is notoriously difficult to diagnose because its symptoms—chest pain, dyspnea, tachycardia—overlap with many other conditions.

In this context, artificial intelligence (AI) has emerged as a transformative tool. AI-driven algorithms, especially those using deep learning, can analyze computed tomography pulmonary angiography (CTPA) images with a speed and consistency that complement human expertise. By flagging suspicious findings and quantifying clot burden, these systems help radiologists make faster, more accurate diagnoses. The technology is now being deployed in major hospitals worldwide, offering a glimpse of a future in which AI serves as a tireless second reader for every emergency CT scan.

Understanding Pulmonary Embolism

Pathophysiology and Risk Factors

Pulmonary embolism is a complication of venous thromboembolism (VTE), a spectrum that includes deep vein thrombosis (DVT). Most emboli originate from proximal leg veins; after detaching, they travel through the venous system into the right atrium, right ventricle, and then into the pulmonary circulation. The size and location of the clot determine the clinical presentation. Small peripheral emboli may cause only mild chest pain, while large central saddle emboli can cause hemodynamic collapse.

Risk factors for PE include prolonged immobility (surgery, long-haul flights, hospitalization), active cancer, pregnancy, hormone therapy, obesity, and inherited clotting disorders such as factor V Leiden. According to the CDC, four out of five DVT cases go undetected until a PE occurs, underscoring the need for effective screening tools.

Clinical Presentation and Diagnostic Challenge

PE presents with a wide range of symptoms: sharp chest pain, shortness of breath, hemoptysis, syncope, and signs of right heart strain. Clinicians often use pre-test probability scores like Wells criteria or Revised Geneva score to guide testing. However, these scores have limited accuracy, and many patients undergo unnecessary CTPA—a test that exposes them to ionizing radiation and contrast nephropathy. Conversely, some high-risk patients are missed due to atypical presentations.

Traditional Diagnostic Methods and Their Limitations

The diagnostic workup for PE typically begins with a D-dimer blood test. A negative D‑dimer rules out PE in low‑probability patients, but a positive result is non‑specific. This leads to many false positives, especially in hospitalized or pregnant patients, forcing clinicians to rely heavily on imaging.

CT pulmonary angiography (CTPA) is the current gold standard. Modern multi‑detector CT scanners produce high‑resolution images of the pulmonary arteries, allowing identification of emboli as filling defects. However, interpreting CTPA requires specialized training and can be time‑consuming. Factors such as motion artifacts, breathing misregistration, and dense contrast in the superior vena cava can obscure subtle clots. Moreover, studies show that radiologists may miss up to 15% of emboli, particularly in smaller subsegmental arteries. These missed diagnoses can lead to recurrent thromboembolism or death.

Other limitations include:

  • Turnaround time: In busy emergency departments, a CTPA scan report may take 30–60 minutes or longer, delaying anticoagulation therapy.
  • Human variability: Reader agreement for subsegmental PE is only moderate, with inter‑observer kappa values around 0.6.
  • Resource burden: High-volume centers process dozens of CTPA studies daily, contributing to radiologist burnout.

These limitations highlight the urgent need for automated tools that can supplement human reading.

How AI‑Driven Algorithms Work in PE Detection

Deep Learning and Convolutional Neural Networks

Most modern AI systems for PE detection are built on convolutional neural networks (CNNs)—a class of deep learning models designed to process grid‑like data such as images. These networks are trained on large datasets of annotated CTPA exams. Typically, thousands of scans are labeled by expert radiologists, marking the location and extent of emboli. The AI learns to recognize patterns of Hounsfield units, vessel morphology, and contrast distribution that correspond to clots.

During inference, the algorithm processes the full CTPA volume—often hundreds of axial slices—and outputs a probability map indicating suspicious regions. Many systems perform instantaneous segmentation of the pulmonary vasculature and highlight filling defects directly on the images. A threshold is set to balance sensitivity and specificity. Some advanced models also quantify the right‑to‑left ventricular diameter ratio, a key biomarker for right heart strain, providing actionable information beyond clot detection.

One well‑studied algorithm is from the Radiology journal, which demonstrated a per‑patient sensitivity of 92% and specificity of 91% for detecting acute PE on CTPA. Such performance approaches that of expert radiologists while reducing reading time by nearly 40%.

Training Datasets and Generalization

The performance of an AI algorithm depends critically on the diversity of its training data. Algorithms trained on data from a single institution may fail when exposed to different CT scanners, reconstruction kernels, or patient populations. To build robust models, developers must curate multi‑center datasets that include varying body habitus, comorbidities, and image quality. Several large public databases, such as the RSNA PE Detection Challenge dataset, have been instrumental in advancing the field. Ongoing research focuses on domain adaptation techniques to improve generalization across institutions.

Clinical Benefits and Evidence from the Literature

Mounting evidence supports the integration of AI‑assisted reading into routine clinical workflows. Multiple retrospective and prospective studies have reported significant improvements in diagnostic performance.

  • Improved sensitivity for subsegmental PE: AI systems can detect clots as small as 1–2 mm in peripheral arteries, which are often overlooked by human readers. A 2022 study published in European Radiology found that AI assistance increased reader sensitivity for subsegmental PE from 74% to 89%.
  • Reduced reading time: In a prospective trial at an academic medical center, radiologists using AI as a triage tool reported a median reading time reduction of 38% (from 4.3 minutes to 2.7 minutes per study).
  • Fewer missed cases: A landmark study by the FDA highlighted that AI‑assisted reading reduced the false‑negative rate by more than 50% in a high‑volume emergency department setting.
  • Workflow prioritization: By flagging positive cases immediately, AI allows critical results to be communicated to the care team faster, reducing time‑to‑anticoagulation by an average of 22 minutes, according to a study from Radiology: Artificial Intelligence.

Beyond detection alone, some AI tools now incorporate automated quantification of clot burden and right ventricular dysfunction, providing prognostic information that guides treatment decisions—such as whether fibrinolysis is warranted.

Practical Implementation in Radiology Workflows

Integrating an AI algorithm into a hospital’s existing picture archiving and communication system (PACS) requires careful planning. Most vendors offer on‑premise or cloud‑based solutions that interface with DICOM protocols.

The AI works in the background, typically as a “virtual second reader.” When a CTPA study is completed, the algorithm processes the images and pushes an alert to the radiologist’s worklist. Studies flagged as positive (above a pre‑defined confidence threshold) appear with a priority marker. Some systems also display a heat‑map overlay showing the location of suspicious clots, allowing the radiologist to confirm or dismiss the finding with a glance.

Several commercial AI products have received FDA clearance for PE detection, including those from Aidoc, Viz.ai, and Vital Images (a Canon company). These products are designed to operate 24/7 without fatigue, maintaining consistent performance throughout the day. Many institutions have reported that using AI has enabled them to reduce unnecessary call‑backs and improve after‑hours coverage.

Challenges in Deployment

Despite promising results, deploying AI in the clinical setting is not without obstacles:

  • Regulatory hurdles: Algorithms must undergo rigorous validation to obtain FDA or CE mark approval. Post‑market surveillance is required to monitor real‑world performance shifts.
  • Data privacy: Cloud‑based algorithms raise concerns about patient data confidentiality. On‑premise installations avoid transmission risks but require significant IT resources.
  • Algorithm bias: Performance may vary by demographic group if the training data lacks diversity. Recent audits have shown that some PE detection models perform worse on female and Black patients, underscoring the need for inclusive datasets.
  • Explainability: Many deep learning models are “black boxes,” making it difficult for clinicians to trust their output. Explainable AI methods, such as attention maps, are being developed to improve interpretability.
  • Workflow friction: Alerts that generate too many false positives can lead to alert fatigue, causing radiologists to ignore or dismiss AI suggestions.

Future Directions

Real‑Time and Multimodal AI

The next frontier in PE detection involves moving beyond static CT analysis. Researchers are exploring AI that can integrate clinical data—such as vital signs, laboratory results, and electrocardiography—with imaging findings to provide a comprehensive risk assessment before the CT scan is even ordered. This could pre‑stratify patients, reducing unnecessary radiation exposure.

Real‑time AI during the CT acquisition itself is also being developed. If the algorithm detects a high probability of PE while the patient is still on the table, it could prompt the technologist to acquire additional delayed phase images for better clot characterization.

Longitudinal Monitoring

AI may also play a role in monitoring treatment response. By comparing serial CTPA exams, algorithms can quantify changes in clot burden over time, helping physicians decide when to stop anticoagulation therapy. Such tools are still in early research stages but hold promise for personalized management.

Integration with Electronic Health Records

Combining AI‑reported PE findings with structured data in EHRs could enable automated clinical decision support—for example, generating recommendations for anticoagulation dosing or follow‑up appointments. This would free clinicians from repetitive tasks and reduce errors in documentation.

Conclusion

AI‑driven algorithms are reshaping the detection of pulmonary embolism, offering real gains in speed, accuracy, and consistency. From flagging subtle subsegmental clots to prioritizing urgent cases, these tools complement radiologists and help deliver timely care that saves lives. The evidence base continues to grow, with regulatory bodies approving more products each year. However, careful attention must be paid to algorithm validation, bias mitigation, and seamless workflow integration to fully realize the potential of AI. As the technology matures, it will likely become an indispensable part of the diagnostic pathway for PE—and for many other acute conditions that depend on rapid, reliable image interpretation.