EjectionPoint: what AI point-of-care LVEF means for 1000 patients

Set a local prevalence of left ventricular systolic dysfunction, the proportion of scans that fail to return an AI value, and how failed scans are handled. The page applies the pooled sensitivity and specificity from studies and patients, the threshold-specific estimates and the Bayesian prediction interval to show expected outcomes, post-test probabilities and net benefit.

Inputs

Prevalence of systolic dysfunction

Emergency department and heart failure clinics: often 20 to 40%. Community screening and primary care: often 5 to 15%.

Scans without an AI value

Accuracy estimate

The primary estimate uses each study's prespecified cut-off. The 50%, 40% and 30% rows come from the hierarchical model in which study random effects are shared across cut-offs.

Failed scans are

Referral treats a failed scan as test positive; "treated as negative" means no further action, the pathway in which failures become missed disease.

Threshold probability for referral

The probability of dysfunction at which a clinician would consider referral for comprehensive echocardiography worthwhile (used for net benefit).

Correctly identified (true positives)

per 1000 examined

Missed (false negatives)

per 1000 examined

Unnecessary referrals (false positives)

per 1000 examined

Comprehensive echos not prompted

true negatives plus missed cases

1000 patients, one square each

True positiveFalse negative (missed)False positiveTrue negativeFailed scan

Post-test probabilities

Probability of systolic dysfunction after a positive and after a negative AI result, from the selected estimate at the chosen prevalence. The 95% confidence interval reflects uncertainty in the summary; the prediction interval (primary estimate only) reflects what a new setting might observe.

Result	Probability	95% CI	95% prediction interval
AI value below cut-off (positive), or failed scan if referred
AI value above cut-off (negative)

Accuracy applied, including failed scans

The selected summary estimate (dark bar), its 95% confidence interval (mid tone) and, for the primary estimate, the Bayesian 95% prediction interval for a new study (light tone), after adjustment for the failure rate and the failure handling you chose.

Sensitivity

0255075100%

Specificity

0255075100%

Net benefit of test-and-refer

Net benefit = (true positives - false positives x pt / (1 - pt)) / n, per 100 patients. The strategy with the highest value at your threshold probability is marked.

AI test-and-refer

Refer every patient

Refer no patient

0.00

Compare the three pathways

The same prevalence and failure rate under the per-protocol assumption (failures ignored, as in conventional 2x2 tables) and the two intention-to-diagnose pathways.

Pathway	Sensitivity	Specificity	Missed / 1000	Referred / 1000	Echos not prompted / 1000

What happens to 1000 patients examined with AI-automated point-of-care LVEF?

1000 patients, one square each

Post-test probabilities

Accuracy applied, including failed scans

Net benefit of test-and-refer

Compare the three pathways