In a retrospective study of 131 post-Fontan patients, abdominal MRI radiomics predicted portal hypertension with AUROC 0.85±0.01, while adding EHR clinical variables failed to improve performance (best combined AUROC 0.77±0.05). Across tasks, clinical-only models lagged (best AUROC 0.70±0.08) and ensemble radiomics varied widely (AUROC 0.33–0.72).

Why It Matters To Your Practice

• This is a real-world test of whether “more data” (EHR + imaging AI features) actually improves risk prediction in complex congenital heart disease follow-up.

• For Fontan-associated liver disease (FALD) surveillance, imaging-derived signals may add more actionable discrimination than routine chart variables—but not necessarily in combination.

• The wide performance spread in ensemble models (AUROC 0.33–0.72) flags potential instability when stacking multiple MRI sequences/features without clear gains.

Clinical Implications

• If you’re considering AI tools for FALD/Fontan surveillance, prioritize models with strong imaging-only validation rather than assuming EHR augmentation will help.

• Don’t equate “moderate AUROC” with readiness for clinical deployment: the best result here was for portal hypertension (AUROC 0.85), while other targets (including Fontan failure) were less consistent.

• When evaluating vendor claims, ask what the comparator was (radiomics-only vs clinical-only vs combined) and whether combining inputs was actually additive.

Insights

• Radiomics from T2-weighted liver + spleen performed best for portal hypertension, suggesting tissue/vascular congestion patterns may be captured better on that sequence.

• Clinical-only EHR models underperformed (max AUROC 0.70), which may reflect noisy, incomplete, or temporally misaligned variables in retrospective records.

• More complex isn’t always better: ensembling across T1/T2/DWI did not reliably improve discrimination and sometimes degraded it substantially.

The Bottom Line

• In this 131-patient Fontan MRI study, radiomics beat EHR-only prediction—and adding EHR data did not boost performance over radiomics alone (0.77 vs 0.85 AUROC for portal hypertension).

• For clinicians assessing AI in practice, insist on evidence that multimodal models improve outcomes, not just model complexity.