01
The question
An AI tool can flag a pulmonary embolism on CT and page the response team when the clot is central and the right ventricle is enlarged. The ratio of right to left ventricle (RV/LV) at which that page fires is a choice. Set it low and the team hears about almost every central clot; set it higher and the alert becomes selective but occasionally stays quiet on a patient who is later escalated. This study asks where that line should sit.
The cohort is 111 consecutive PERT-positive patients drawn from 14,300 CT studies (4,190 dedicated CT pulmonary angiograms and 10,110 studies screened for incidental PE) over seven weeks in 2023. Every patient carried a central embolism and met an RV/LV floor of 1.0 at intake.
02
How the cohort was built
| Detection path | CT studies | Intake screen | PERT-positive |
|---|---|---|---|
| All studies | 14,300 | Central PE · RV/LV ≥ 1.0 | 111 |
| Dedicated CTPAordered for suspected PE | 4,190 | → | 103by the PE algorithm |
| Incidental screennot ordered for PE | 10,110 | → | 8found without looking |
Two arms, one screen. The dedicated arm behaves as expected: someone suspected PE, ordered the scan, and the algorithm confirmed it. The incidental arm is the AI-specific catch — eight PERT-positive patients on scans nobody ordered for PE. Every patient met an RV/LV floor of 1.0 at intake, so the cohort holds no screen-negative patients: the threshold analysis re-selects within an already PERT-positive group, it does not catch new cases.
03
The trade-off, row by row
As the threshold climbs from 1.0 to 1.3, alert volume falls and agreement with the escalation clinicians actually delivered rises. The two lines cross just below 1.2, the point Northwell is adopting.
| RV/LV threshold | Alerts | % of 111 | Concordance | Escalated, no alert |
|---|---|---|---|---|
| 1.0 PE-RADS / Yale floor | 105 | 95% | 40% | 2 |
| 1.1 | 75 | 68% | 59% | 6 |
| 1.2 Operating point | 64 | 58% | 66% | 8 |
| 1.3 | 54 | 49% | 71% | 10 |
Concordance is agreement between the alert and a composite escalation outcome (thrombectomy candidacy or performance, systemic thrombolysis, or ICU admission). Moving to 1.2 removes 41 of the alerts seen at the 1.0 floor while still capturing 34 of the 42 escalated patients.
04
Sensitivity and specificity by threshold
The four rows above fix the line at tenth-point stops. The control below moves the same line in increments of one hundredth, from the PE-RADS floor of 1.0 to a ratio of 2.0, and reports the alert's sensitivity and specificity against the recorded clinical outcome at each position. Sensitivity is the share of patients receiving the reference outcome on whom the alert would fire; specificity is the share not receiving it on whom the alert stays quiet.
Two-by-two at the selected threshold
| Escalated | Not escalated | |
|---|---|---|
| Alert fired | 34true positive | 30false positive |
| Alert quiet | 8false negative | 39true negative |
The curves below trace both measures across the full range of the threshold. The marker follows the slider above, so its position and the two points read off alert volume and concordance at the selected value.
What these figures measure
These values describe the alert's agreement with a recorded clinical outcome, not its accuracy at detecting pulmonary embolism. Every patient in the cohort carries a central embolism and entered above an RV/LV ratio of 1.0, so the sample holds no PE-negative and no screen-negative patients. Sensitivity and specificity therefore express how the threshold sorts an already PERT-positive group by the care its patients received, and they are not transferable to the wider population a prospective screen would encounter.
Recorded right-ventricular strain was read with the algorithm value visible, so agreement against it may be inflated pending the blinded re-read. Observed escalation is the more independent reference of the two.
05
Where PE-RADS v2026 fits
The new Pulmonary Embolism Reporting and Data System grades the most proximal clot from 0 to 4 and reserves category 4 for central emboli. It adds an RV+ modifier for right ventricular enlargement, set at an RV/LV ratio of 1.0 or greater, and routes a central embolism carrying RV+ toward specialist consultation and reperfusion assessment. Every patient in this cohort is a PE-RADS 4, so the alert is exactly the 4/RV+ combination, and the threshold under study is the RV+ line itself.
The working group chose 1.0 to label an enlarged ventricle, valuing specificity over sensitivity for that description. As an alerting line within a central-PE population, 1.0 fires on nearly everyone, which is why a local operating point above the floor makes the alert useful for triage.
06
Hub and spoke
Spoke patients
Escalated 45.7% of the time; reached thrombectomy at a median of 28 hours.
Hub patients
Escalated 35.6% of the time; reached thrombectomy at a median of 24 hours.
Spoke patients were escalated at least as often as hub patients yet waited longer for intervention, a gap that reflects the added steps of recognition, mobilisation, and transfer. This is the part of the pathway a faster alert could compress, before any transfer clock begins.
07
What the numbers do and do not show
Read as a plan, not a proof
Aidoc ran retrospectively and never paged anyone during the study, so every alert-versus-care figure describes how the alert would have behaved. Faster consultations in higher-ratio patients reflect clinicians prioritising sicker patients, not an effect produced by the alert. Establishing an outcome benefit needs a prospective study with the alert active.
Other limits: the cohort has no screen-negative patients, so alert specificity against the wider PE population cannot be measured here; thrombectomy numbers are small; consultation times were coarsely recorded; and the recorded RV strain read may be anchored to the algorithm value pending a blinded re-read.