AI-derived right ventricular ejection fraction for mortality prediction and subclinical dysfunction

Introduction

AI-based segmentation enables rapid and reproducible quantification of cardiac function. Although RVEF is an important prognostic marker across a range of cardiovascular diseases, its assessment has been limited by time-intensive manual clinician measurement. We hypothesised that with a novel AI tool, RVEF (RVEFAI) can be performed rapidly and routinely, and would improve prognostic value, reflecting higher measurement accuracy.

Methods

Consecutive adult patients who underwent CMR between 2014–2019 (no complex congenital heart disease) and with clinician-recorded RVEF (RVEFDoc) were studied. A subgroup of patients with normal CMR or minor abnormalities (e.g. mild valve lesions, mild chamber/aortic dilation/ LVH) was identified. Patients were followed for all-cause mortality via a national database. AI segmentation used a validated U-Net model trained on >2,000 subjects; no manual correction was performed. Prediction of mortality was evaluated using Cox proportional hazards models. Incremental prognostic value was tested using likelihood ratio tests of nested models (AI added to clinician models and vice versa), AIC, and C-indices.

Results

RVEFAI segmentation required 13 s per patient. Of 33,396 patients scanned, 15,317 (46%) had clinician-measured RVEF and were included in survival analysis; AI segmentation was performed in all cases with a 3% contour-error rate. Median age was 55 years (IQR 42-66)years, 9,776 (64%) were male. Over a median follow-up of 7.3(IQR 6.1–8.7)years, 2,259 patients (15%) died.

Among the subgroup with normal CMR or minor abnormalities (n=6,115), median age was 50(IQR: 37–62)years, 3,180(52%) were male, and 427 patients died (7%). In both cohorts, mortality was associated with increasing age and male sex.

Measurement variability was greater for RVEF than LVEF (Figure).

Addition of RVEFAI to RVEFDoc resulted in a six-fold greater improvement in model fit than addition of RVEFDoc to RVEFAI (Δχ² 147 vs 24; both p<0.0001) with similar findings for LVEF (Δχ² 217 vs 36; both p<0.0001). Findings were replicated in the normal/minor CMR subgroup (RVEF: Δχ² 58 vs 14; LVEF: Δχ² 59 vs 28; all p<0.0001).

Hazard ratios were higher for RVEFAI than RVEFDoc in both cohorts (Allcomers: HRperIQR 1.6 [95% CI 1.5–1.6] vs 1.5 [95% CI 1.4–1.5]; Normal/minor: HRperIQR1.5 [95%CI: 1.3–1.7] vs 1.1 [95% CI: 1.0–1.3]).

In multivariable models adjusted for age and sex, RVEF provided incremental prognostic value beyond LVEF; AI-based models demonstrated consistently better prediction to clinician models (Table).

Conclusion

The AI-based RVEF quantification method evaluated is rapid, robust, and feasible in all patients undergoing CMR. AI-derived RVEF outperforms clinician measurements for prognostic value, including in patients with normal or minor findings. Age- and sex-adjusted models incorporating both RVEF and LVEF improved risk stratification, with superior performance using AI-derived measurements.FigureTable