AI derived RV:LV ratio >1.18 significantly improved prediction of one year survival in acute PE patients compared to radiologist assessment (p = 0.04).
Does AI automated RV:LV quantification software reliably measure right heart strain and predict survival in patients with acute pulmonary embolism compared to radiologist assessment?
AI-automated RV:LV ratio quantification on CTPA is reliable and improves prognostic risk stratification for one-year survival in acute pulmonary embolism compared to initial radiologist reports.
Absolute Event Rate: 0% vs 0%
Abstract Background Increased right ventricle to left ventricle (RV:LV) ratio in acute pulmonary embolism (PE) indicates right heart strain (RHS), a marker of poor prognosis.1,2 RV:LV ratio is unreliably reported in CT pulmonary angiography (CTPA) despite its importance in risk stratification.3 Automated measurement could improve reporting consistency and influence patient care. Purpose To retrospectively compare the performance of an artificial intelligence (AI) automated RV:LV quantification software with routine and expert thoracic radiologist assessment. To assess its potential clinical impact on acute PE risk stratification and patient management in line with European Society of Cardiology guidelines.2 Methods Retrospective, single-centre analysis of 114 consecutive CTPAs (Dec 2022-Nov 2023) confirmed PE-positive by an expert thoracic radiologist. RV:LV ratios were measured retrospectively by the expert thoracic radiologist and AI using contrast-enhanced axial slices. RHS was defined as an expert RV:LV ratio 1.0 and an AI derived RV:LV ratio 1.18, in keeping with previous research.4 Original CT reports were interrogated for mention of RHS by two senior clinicians. Cohen’s kappa (κ) evaluated their agreement. Clinical outcome data was extracted from electronic health records. Sensitivity, intraclass correlation coefficient (ICC), and Kaplan-Meier analyses evaluated performance and one year survival. Results RV:LV ratios were measurable for 108/114 (95%) CTPAs. Mean RV:LV ratios for expert and AI were 1.09 (range 0.58-2.12) and 1.17 (0.56-2.97) respectively. ICC was 0.78 (p0.01) indicating good agreement between methods although systematic bias in AI measurements was present (mean difference = 0.09 -0.28, 0.46). RHS was mentioned in 92/114 (81%) initial reports (RHS-positive 27/92 29%, RHS-negative 65/92 71%; κ = 0.99). AI would have reclassified RHS in 18 patients: upgrading 16 to RHS-positive and downgrading 2 to RHS-negative (Figure 1). The expert review reclassified 17 additional scans as RHS-positive deemed RHS-negative by AI. The sensitivity and specificity of AI RV:LV ratio 1.18 for detecting RHS compared to expert radiologist were 0.70 (95% CIs 0.56-0.81) and 0.92 (95% CIs 0.81-0.98). Risk stratification using an AI RV:LV ratio 1.18 upgraded 3 patients from low to high risk and downgraded 1 patient from high to low risk compared to the initial CT reports. Complete risk stratification was limited by missing troponin data. This was only available in 73/114 (64%) cases. AI derived RV:LV ratio 1.18 significantly improved prediction of one year survival (p = 0.04) compared to initial radiologist assessment of RHS (p = 0.31) (Figure 2). Conclusions AI derived RV:LV ratio calculation is reliable, offering better evaluation of RHS compared to initial radiologist report. AI derived RV:LV ratio 1.18 is a poor prognostic marker of all-cause mortality and should be used with other means of risk stratification to inform patient care.
Gunning et al. (Thu,) reported a other. AI derived RV:LV ratio >1.18 significantly improved prediction of one year survival in acute PE patients compared to radiologist assessment (p = 0.04).