Deep ECG survival prediction for all-cause mortality under case-mix shift: discrimination, risk underestimation, and recalibration

Background Accurate mortality ranking does not establish reliable absolute-risk prediction when the clinical population changes. We examined discrimination, calibration, demographic increment, and recalibration of deep ECG survival prediction under case-mix shift. Methods A five-seed ECG-demographic discrete-time survival ensemble was developed using CODE-15% training (183,689 participants; 6,603 deaths), validation (22,974; 828), and internal-test cohorts (22,959; 825). Cross-release waveform hashing and lagged cross-correlation identified four duplicates in the Chagas-focused SaMi-Trop cohort; their exclusion left 1,627 participants and 104 deaths. The fixed models were assessed at 1 and 2 years using censoring-aware metrics, paired patient bootstrap intervals, training-only Cox references, and fivefold out-of-fold recalibration. Input-sensitivity analyses included patient-level uncertainty and numerically verified integrated gradients. Results The concordance index (C-index) decreased from 0.842 (95% CI 0.829–0.854) internally to 0.797 (0.757–0.836) externally, and 1-year area under the receiver operating characteristic curve (AUROC) decreased from 0.862 to 0.818. External incremental discrimination over ECG-only was unresolved (C-index difference 0.011, − 0.025 to 0.049). At 2 years, mean external predicted mortality was 3.21% vs. 6.27% observed, with calibration-in-the-large 0.703 (0.494–0.899). Out-of-fold intercept-only and intercept-plus-slope recalibration reduced the Brier score from 0.05558 to 0.05337 and 0.05324, respectively. These improvements quantify within-cohort probability accuracy rather than independently validated clinical benefit. Conclusions This study shows that informative ECG-based mortality ranking can coexist with substantial risk underestimation under case-mix shift. Recalibration improved probability accuracy within SaMi-Trop but shared Brazilian acquisition infrastructure, Chagas-related selection, and unresolved cross-release identity linkage limit claims about transportability. Independent validation is needed before calibrated risks can support clinical decisions.

Authors

Institutions

Publication Details

Journal
Frontiers in Cardiovascular Medicine
Published
2026-11-19
DOI
https://doi.org/10.3389/fcvm.2026.1993391
Primary Topic
ECG Monitoring and Analysis
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Deep ECG survival prediction for all-cause mortality under case-mix shift: discrimination, risk underestimation, and recalibration

Mehmet Burukanlı, Shahid Mohammad Ganie, İshak Paçal, Hamidullah Turkmen
Frontiers in Cardiovascular Medicine
ECG Monitoring and Analysis
article

Deep ECG survival prediction for all-cause mortality under case-mix shift: discrimination, risk underestimation, and recalibration

Mehmet Burukanlı, Shahid Mohammad Ganie, İshak Paçal, Hamidullah Turkmen
article en

Abstract

Background Accurate mortality ranking does not establish reliable absolute-risk prediction when the clinical population changes. We examined discrimination, calibration, demographic increment, and recalibration of deep ECG survival prediction under case-mix shift. Methods A five-seed ECG-demographic discrete-time survival ensemble was developed using CODE-15% training (183,689 participants; 6,603 deaths), validation (22,974; 828), and internal-test cohorts (22,959; 825). Cross-release waveform hashing and lagged cross-correlation identified four duplicates in the Chagas-focused SaMi-Trop cohort; their exclusion left 1,627 participants and 104 deaths. The fixed models were assessed at 1 and 2 years using censoring-aware metrics, paired patient bootstrap intervals, training-only Cox references, and fivefold out-of-fold recalibration. Input-sensitivity analyses included patient-level uncertainty and numerically verified integrated gradients. Results The concordance index (C-index) decreased from 0.842 (95% CI 0.829–0.854) internally to 0.797 (0.757–0.836) externally, and 1-year area under the receiver operating characteristic curve (AUROC) decreased from 0.862 to 0.818. External incremental discrimination over ECG-only was unresolved (C-index difference 0.011, − 0.025 to 0.049). At 2 years, mean external predicted mortality was 3.21% vs. 6.27% observed, with calibration-in-the-large 0.703 (0.494–0.899). Out-of-fold intercept-only and intercept-plus-slope recalibration reduced the Brier score from 0.05558 to 0.05337 and 0.05324, respectively. These improvements quantify within-cohort probability accuracy rather than independently validated clinical benefit. Conclusions This study shows that informative ECG-based mortality ranking can coexist with substantial risk underestimation under case-mix shift. Recalibration improved probability accuracy within SaMi-Trop but shared Brazilian acquisition infrastructure, Chagas-related selection, and unresolved cross-release identity linkage limit claims about transportability. Independent validation is needed before calibrated risks can support clinical decisions.

Frontiers in Cardiovascular MedicineVol. 13
Fenerbahçe University (TR), Iğdır Üniversitesi (TR), Bitlis Eren University (TR), İstanbul Gelişim Üniversitesi (TR), Nakhchivan University (AZ), King Faisal University (SA), Nakhchivan State University (AZ)
ECG Monitoring and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon