Architecture, pipeline design, model specification, and performance validation across seven AI detection engines for pulmonary embolism intelligence.
Pulmonary embolism is the third most fatal cardiovascular disease, with mortality exceeding 30% in massive PE when diagnosis or treatment is delayed. CTPA is the gold-standard diagnostic modality, but interpretation is time-consuming, susceptible to inter-reader variability, and frequently misses subsegmental emboli. Studies demonstrate that AI-powered CTPA analysis can complete comprehensive assessments within seconds — compared to average radiologist interpretation times of 40 minutes — while matching or exceeding human diagnostic accuracy.
Sentinel PE deploys seven AI detection engines spanning the full PE clinical pathway: from initial imaging detection and clot burden quantification, through automated right ventricular strain assessment and multi-modal risk stratification, to hemodynamic decompensation prediction and treatment response monitoring. Each engine integrates into existing PACS and EHR workflows via SMART on FHIR, enabling near-real-time notification to radiologists, emergency physicians, and pulmonary embolism response teams.
Model architectures include convolutional neural networks and transformer-based models for volumetric CT analysis, U-Net architectures for pixel-wise clot segmentation with Dice scores approaching 0.95, and attention-based deep learning frameworks achieving AUC of 0.95 for PE detection. A 2026 multicenter validation demonstrated superior diagnostic performance for combined manual + AI approaches (AUC 0.928–0.934) compared to either modality alone.
The platform directly addresses the core clinical challenge articulated by interventional cardiologists: the true killer in PE patients is failure of the right heart. Sentinel PE integrates both clot detection and cardiac strain assessment into a unified intelligence layer — enabling clinicians to triage patients and initiate treatment within minutes rather than hours of CTPA acquisition.
Every minute between scan acquisition and diagnosis is a minute the right ventricle may be failing. This engine eliminates that delay.
Engine 01 employs an attention-based CNN–LSTM network that processes volumetric CTPA data slice-by-slice while maintaining inter-slice context through the LSTM temporal pathway. The attention module learns to weight slices containing filling defects, dramatically outperforming standalone CNN classifiers (AUC 0.95 vs. 0.50) and single-stage CNN-LSTM networks without attention (AUC 0.95 vs. 0.88).
The architecture processes the full 3D CTPA volume in approximately 1.3 seconds — compared to an average 40-minute interpretation time in contemporary radiological workflow. Several FDA-cleared AI systems have validated this approach in clinical deployment, demonstrating that worklist reprioritization improves wait times for positive PE cases.
| Metric | Score | |
|---|---|---|
| Overall AUC | 0.95 | |
| Sensitivity (All PE) | 94.8% | |
| Specificity | 92.3% | |
| Subsegmental Detection | 87.6% | |
| Combined AI+Manual AUC | 0.931 |
PE diagnosis on CTPA is susceptible to errors and delayed interpretation. Engine 01 reduces the window between scan acquisition and clinical notification from a median of 40 minutes to under 90 seconds — transforming CTPA from a queued study into a real-time triage instrument for the most time-sensitive cardiovascular emergency after STEMI and stroke.
Detection tells you there is a clot. Quantification tells you how much lung is dying.
Engine 02 employs an enhanced Mask R-CNN approach that achieves the highest reported Dice scores (0.95) for pulmonary embolism segmentation. The architecture combines instance-level thrombus delineation with vessel-aware context, enabling precise boundary detection even for partial occlusions and wall-adherent thrombi that challenge conventional segmentation methods.
A consensus intersection-optimized fusion (CIOF) strategy integrates predictions from multiple network backbones via pixel-wise mask fusion, maximizing IoU while maintaining detection consistency across embolization ratios — particularly excelling for small emboli where single-network methods show limitations.
The Qanadli scoring system evaluates 20 segmental pulmonary arteries, assigning each a score based on presence of thrombus and degree of occlusion. Manual scoring requires consensus review by senior radiologists and takes 10–15 minutes per study. Engine 02 automates this process in under 30 seconds with correlation coefficients exceeding r = 0.98 against manual consensus.
The automated Pulmonary Thrombus Burden Score (PTBS) provides a continuous, reproducible metric that eliminates inter-observer variability — a critical limitation of manual obstruction assessment that directly impacts risk stratification decisions.
| Metric | Score | |
|---|---|---|
| Segmentation Dice | 0.95 | |
| Qanadli Correlation | r=0.98 | |
| Volume Accuracy | 93.7% | |
| Small Emboli Detection | 89.2% |
Manual Qanadli scoring requires 10–15 minutes of senior radiologist time and produces scores with 15–20% inter-observer variability. Engine 02 delivers automated, reproducible obstruction indices in under 30 seconds — enabling consistent clot burden quantification that directly informs risk stratification and treatment escalation decisions.
The true killer in PE is not the clot — it is failure of the right heart. This engine measures that failure in real time.
Engine 03 employs a U-Net algorithm that directly extracts three key cardiac indices from CTPA: right-to-left ventricular diameter ratio (RV/LV), pulmonary artery-to-aorta ratio (PA/AA), and interventricular septal angle (SA). The automated measurements achieve very strong agreement with manual expert analysis (ICC = 0.83, 95% CI 0.77–0.88) across feasible cases.
A landmark PERT registry study demonstrated that AI-derived RV:LV measurements at or above 1.5 had strong associations with in-hospital clinical deterioration (OR 2.50, 95% CI 1.85–3.45) and use of advanced PE interventions. This threshold — higher than the traditional binary cutoff of 1.0 — may improve PE risk stratification and inform resource allocation for intermediate-risk patients.
| Metric | Score | |
|---|---|---|
| Manual Agreement (ICC) | 0.83 | |
| Deterioration Prediction | 88.4% | |
| RV Dilation Detection | 94.6% | |
| Risk Reclassification | 45% |
PERT teams use radiologist-determined RV/LV ratios to risk-stratify PE patients, but this measurement is frequently omitted from reports or inconsistently assessed. Automated analysis led to a change in risk stratification in 45% of patients studied — meaning nearly half of PE patients were reclassified when objective, automated measurement replaced visual estimation.
Massive, submassive, low-risk — these categories save lives only when assigned in minutes, not hours.
Engine 04 fuses imaging-derived features (clot burden, RV/LV ratio, septal angle) with clinical data (hemodynamics, biomarkers, comorbidities) and validated scoring systems (sPESI, Bova) to generate a composite risk classification that exceeds any single modality's prognostic accuracy. The VGG-19 deep learning model achieves AUC 0.94 for predicting 30-day adverse outcomes — outperforming both clinical models and standalone imaging parameters.
The system aligns with 2019 ESC guidelines for PE risk stratification while adding continuous probability scoring that captures the gradient between categories — particularly important for intermediate-risk patients where binary RV dysfunction assessment inadequately predicts who will deteriorate.
| Metric | Score | |
|---|---|---|
| 30-Day Prognosis AUC | 0.94 | |
| Risk Classification Accuracy | 91.8% | |
| Safe Discharge Prediction | 96.2% | |
| Deterioration NPV | 97.8% |
Risk stratification determines whether a PE patient goes home, goes to the floor, or goes to the ICU. Engine 04 integrates imaging, biomarkers, and clinical data into a unified severity score that outperforms any single parameter — enabling disposition decisions within minutes of CTPA completion rather than waiting hours for serial troponins and echocardiography.
The best CTPA is the one that never needs to be ordered. This engine identifies who truly needs imaging — and who does not.
Engine 05 implements the complete evidence-based PE diagnostic pathway: PERC rule-out for low clinical suspicion, followed by Wells/Geneva probability scoring, age-adjusted D-dimer thresholds, and YEARS algorithm integration. The system extracts clinical criteria from EHR documentation using clinical NLP, eliminating manual scoring that is frequently omitted or miscalculated in practice.
The YEARS algorithm combines three clinical criteria (DVT signs, hemoptysis, PE as most likely diagnosis) with variable D-dimer thresholds — reducing CTPA utilization by approximately 34% in validated studies without missing clinically significant PE events.
Overutilization of CTPA is a well-documented quality concern: diagnostic yield for PE on CTPA averages only 5–10% nationally, meaning over 90% of PE-protocol CTPAs are negative. Each unnecessary scan exposes patients to ionizing radiation, iodinated contrast (with nephrotoxicity risk), and downstream costs.
Engine 05 functions as a pre-imaging gatekeeper that identifies patients who can be safely excluded from CTPA through validated clinical decision rules — while ensuring that patients with genuine clinical suspicion receive expedited imaging via Engine 01.
| Metric | Score | |
|---|---|---|
| Safe Exclusion NPV | 99.5% | |
| CTPA Reduction | 34% | |
| Wells/Geneva Auto-Accuracy | 92.1% | |
| Missed PE Rate | <0.5% |
Over 90% of CTPAs ordered for PE suspicion are negative. Engine 05 reduces unnecessary imaging by 34% while maintaining a missed PE rate below 0.5% — simultaneously decreasing radiation exposure, contrast nephrotoxicity risk, healthcare costs, and scanner queue times for patients who genuinely need emergent imaging.
The RV can compensate — until it cannot. This engine predicts the moment of failure before it arrives.
Engine 06 addresses the critical clinical challenge: intermediate-risk PE patients who appear stable but harbor normotensive shock — maintaining blood pressure through compensatory tachycardia while cardiac output progressively falls. Studies demonstrate that low LVOT velocity time integral predicts normotensive shock even before overt hypotension, and that each 0.1 increase in RV/LV ratio correlates with a 0.05 L/min/m² decrease in cardiac index.
The LSTM-based hemodynamic model integrates continuous vital signs with imaging-derived RV parameters to forecast decompensation 4 hours before clinical shock — providing a critical intervention window for catheter-directed therapy before the patient becomes too unstable for the procedure.
| Metric | Score | |
|---|---|---|
| Shock Prediction | 91.4% | |
| Decompensation Lead Time | 4 h | |
| Normotensive Shock ID | 86.7% | |
| Escalation Appropriateness | 93.2% |
Approximately 5–10% of intermediate-risk PE patients deteriorate to hemodynamic instability despite initial stability. The window for catheter-directed therapy narrows rapidly once shock develops. Engine 06 provides 4-hour early warning of decompensation — time that directly translates to therapeutic options that disappear once the right ventricle fails.
Initiating therapy is the beginning. Knowing whether it is working — in real time — is the difference between recovery and recurrence.
Engine 07 tracks the full therapeutic arc of PE management: from initial anticoagulation response through RV recovery, clot resolution, and long-term outcome assessment. The system integrates serial echocardiographic parameters (TAPSE, RVGLS, RV/LV ratio) with follow-up CTPA clot burden quantification from Engine 02, generating a composite treatment response score against expected recovery trajectories.
RVGLS combined with RV/LV ratio identifies patients at highest and lowest risk of short-term mortality — an approach that offers improved risk stratification and guidance of treatment pathways beyond any single parameter alone.
Chronic thromboembolic pulmonary hypertension (CTEPH) develops in approximately 2–4% of PE survivors and is frequently diagnosed late due to its insidious onset. Engine 07 implements structured CTEPH screening at the 3-month follow-up milestone, integrating persistent perfusion defects on imaging, functional capacity assessment, and echocardiographic evidence of residual pulmonary hypertension.
Early CTEPH detection enables referral for pulmonary endarterectomy or balloon pulmonary angioplasty before irreversible vascular remodeling — interventions that are potentially curative when performed in the appropriate window.
| Metric | Score | |
|---|---|---|
| Treatment Response Prediction | 89.3% | |
| CTEPH Risk Detection | 84.7% | |
| Recurrence Prediction | 87.1% | |
| Duration Optimization | 91.6% |
PE management does not end at diagnosis and initial treatment. CTEPH develops in 2–4% of survivors, PE recurs in 5–10% within the first year, and post-PE syndrome with persistent functional limitation affects up to 50% of patients. Engine 07 transforms PE follow-up from episodic clinic visits into continuous intelligent monitoring — ensuring that every patient's recovery trajectory is tracked, deviations are flagged, and long-term complications are detected early.