Abstract Sat403: Automated Whole-Brain ADC Histogram Analysis for Neurological Prognostication in Post-Cardiac Arrest Patients: A Validation Study

Abstract

Background: Apparent diffusion coefficient (ADC) from diffusion-weighted MRI reflects cytotoxic edema and enables neurological prognostication after cardiac arrest. However, current methods often rely on subjective interpretation or manual region-of-interest analysis, leading to inter-observer variability. Even quantitative tools often require time-consuming manual steps. This study aimed to develop and validate a fully automated whole-brain ADC histogram analysis system that eliminates human bias and enables rapid, objective outcome prediction. Method: This single-center retrospective study included adult out-of-hospital cardiac arrest (OHCA) survivors receiving targeted temperature management. 3T MRI scans acquired 72–96 h after ROSC were included. Patients were randomly divided into derivation (70%) and validation (30%) cohorts. MRI data were processed using JLK-ADC, an AI-driven platform that automatically segments brain parenchyma and generates whole-brain ADC histograms. Neurological outcomes were assessed at 6 months using Cerebral Performance Category (CPC 3–5 defined as poor). Results: A total of 119 comatose OHCA patients were included: derivation cohort (n = 83), validation cohort (n = 36). Voxel-wise histogram analysis revealed that patients with poor neurological outcomes exhibited significantly higher proportions of voxels in low ADC ranges (<=600 x 10^-6 mm2/s), reflecting cytotoxic edema. The 550–600 interval demonstrated highest prognostic performance (AUC 0.792; 95% CI, 0.679–0.891). Cumulative analysis showed that several cutoffs — particularly <=500, <=550, <=600, <=650, and <=700 x 10^-6 mm2/s — were all associated with strong outcome discrimination. The <=600 threshold yielded best overall performance in the derivation cohort at a voxel proportion cutoff of 3.79%, with sensitivity 76.2%, specificity 80.5%, PPV 80.0%, and NPV 76.7%. Internal validation confirmed robust performance: the same <=600 threshold achieved an AUC of 0.840, sensitivity 66.7%, specificity 94.4%, and PPV 92.3%. Pairwise ROC comparisons showed no significant differences in AUCs (p = 0.61), supporting generalizability. Conclusion: This study presents the first fully automated whole-brain ADC histogram analysis for neurological prognostication in cardiac arrest survivors. This approach achieved robust performance across independent validation cohorts, offering clinicians an objective, rapid alternative to subjective manual analysis in post-cardiac arrest care.