Transport-resolved inverse modelling of intra-procedural angiographic time-density curves predicts first-pass effect and futile recanalization after thrombectomy

Overview Intra-procedural digital subtraction angiography acquired during mechanical thrombectomy records how contrast passes through the parenchyma, but outcome models usually reduce each territorial time-density curve to peak timing or deconvolved flow. We asked whether inverting a two-branch transport model of each curve recovers a transport state that predicts the first-pass effect and futile recanalization. In 760 patients, every curve was encoded in a single forward pass by an amortised Bayesian inverse trained on simulated curves from a stated population prior and acquisition model, giving a posterior over arrival delay, bolus width, convection and retention time constants and the convection fraction. Posteriors were summarised within each territory into a patient-level fingerprint: the pre-pass summary for the first-pass model, and the pre-to-final change added for futile recanalization. A gradient-boosted model combined the fingerprint with clinical variables. On the 182 held-out patients the fingerprint reached an AUROC of 0.819 for the first-pass effect and, among the 165 held-out patients with successful recanalization, 0.775 for futile recanalization, compared with 0.790 and 0.758 for scalar curve descriptors, 0.676 and 0.707 for deconvolved perfusion maps, 0.646 and 0.628 for clinical variables, and 0.802 and 0.808 for the prespecified end-to-end comparator on each endpoint. No score-by-stratum interaction reached significance across occlusion site, onset-to-puncture time, collateral score, device, pre-pass acquisition truncation and pre-pass signal-to-noise tertile (subgroup discrimination in Supplementary Table S3, interaction tests in Supplementary Table S10). The transport fingerprint is interpretable and computable from the runs already acquired during thrombectomy. Contents This record contains 34 files. The table lists every data table in the archive with its row count. FileRowsSize (KB)data/derived/ablation.csv212data/derived/cohort_boundary.csv50data/derived/comparisons.csv121data/derived/comparisons_matched.csv40data/derived/curve_posterior.csv9,1202017data/derived/final_grade_comparator.csv40data/derived/final_grade_deltas.csv40data/derived/importance_groups.csv684data/derived/inference_time.csv10data/derived/main_table.csv185data/derived/metrics.csv367data/derived/predictions.csv5,211298data/derived/prior_sensitivity.csv51data/derived/recovery.csv52data/derived/recovery_boundary.csv60data/derived/recovery_calibration.csv454data/derived/recovery_curves.csv4,8001380data/derived/recovery_example.csv18010data/derived/recovery_truncation.csv316data/derived/sensitivity.csv181data/derived/subgroups.csv25225data/derived/subgroup_interactions.csv121data/derived/svd_rectification.csv82data/derived/transport_fit.csv12,0061266data/derived/transport_fit_summary.csv61data/raw/cohort.csv76053 Reuse Released under cc-by-4.0. If you use these data, please cite this record together with the associated article.