Availability-Aware Prediction of CCTA Findings Using Opportunistic Low-Dose CT and Routine Health-Check Data.
Authors
Affiliations (6)
Affiliations (6)
- Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei, 112, Taiwan.
- Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, Taipei, 112, Taiwan.
- Heart Center, Cheng Hsin General Hospital, Taipei, 112, Taiwan.
- Health Management Center, Cheng Hsin General Hospital, Taipei, 112, Taiwan. [email protected].
- Institute of Biophotonics, National Yang Ming Chiao Tung University, Taipei, 112, Taiwan. [email protected].
- Center for Smart Health and Medicine, Taipei City Hospital, Taipei, Taiwan. [email protected].
Abstract
Laboratory and history data are inconsistently available in health-check records, complicating prediction from combined clinical and imaging inputs. We evaluated an availability-aware ensemble for predicting coronary CT angiography (CCTA) findings in 864 participants, including 221 with Coronary Artery Disease Reporting and Data System (CAD-RADS) category ≥ 3. Models used eight clinical and 31 low-dose CT (LDCT) inputs, with optional laboratory and history variables. Six learner families underwent fivefold outer and fivefold inner validation; preprocessing, model selection, and cross-fitted sigmoid calibration were confined to outer training data. Calibrated adaptive top-3 aggregation achieved an area under the receiver operating characteristic curve (AUROC) of 0.808, average precision of 0.622, and Brier score of 0.14152. Compared with selection of one compatible model, the AUROC difference was + 0.0193 (conditional 95% confidence interval 0.0029-0.0358) and the Brier difference was - 0.01174 (- 0.01723 to - 0.00621). Comparisons with the expanded-pool imputation baseline and common-feature ensembles were inconclusive. Matched analyses supported adding laboratory variables to single-model selection; history increments were uncertain. At threshold 0.15, 29/221 positive cases were missed (13.1%; 95% Wilson interval 9.3-18.2%). Human reviewers considered the proxy measurements acceptable in all 864 analyzed participants. Aggregation improved AUROC and Brier score relative to compatible single-model selection in this cohort. Its additional value over simpler ensembles and its suitability for clinical triage require external and prospective evaluation.