Deep learning versus conventional risk scores for postoperative mortality prediction in adult cardiac surgery: a systematic review

Background Earlier reviews of machine learning for mortality after cardiac surgery pooled deep learning (DL) with shallow and non-neural methods. We asked whether models meeting an explicit definition of DL discriminate short-term mortality better than EuroSCORE II, the Society of Thoracic Surgeons Predicted Risk of Mortality (STS-PROM) or logistic regression in the same patients. Methods This re-conducted review (PROSPERO CRD420261354546) supersedes an earlier preprint. MEDLINE, Europe PMC, Lens.org and IEEE Xplore were searched on 27–28 September 2026, with backward citation searching; Embase, CINAHL and CENTRAL were not searched, and forward citation searching was not performed. DL required at least two hidden layers, or a named architecture of established depth, shown in the source. A large language model pre-screened records and extracted data; two reviewers verified every decision and value. Risk of bias was assessed with PROBAST + AI, and results were synthesised without meta-analysis as within-cohort differences in the area under the curve (AUC). Results Of 1,845 records identified, 515 duplicates were removed and 1,330 were screened; of 76 reports sought, 54 were assessed in full and 22 could not be retrieved. Four studies (five cohort-level comparisons) met all criteria. A chest-radiograph convolutional network had lower AUCs than STS-PROM in temporal (0.829 vs. 0.884) and external (0.738 vs. 0.803) cohorts with about two dozen deaths each; two-hidden-layer perceptrons did not outperform logistic regression (0.761 vs. 0.762; 0.77 vs. 0.80); one study reported AUCs only graphically. All four studies had high or unclear risk of bias. Fifteen reports described networks of undetermined depth, including a tabular transformer that outperformed EuroSCORE II internally (0.834 vs. 0.807); in the 12 with extractable comparisons, the network had a higher AUC than the best comparator in 4 and lower in 8 (median difference − 0.004). Conclusions Evidence that DL improves on established scores for short-term mortality after adult cardiac surgery is sparse, inconsistent and at high or unclear risk of bias, and the search was incomplete relative to the registered protocol. Differences in discrimination were small or absent, and calibration and clinical utility were rarely evaluated. Superiority cannot be established. PROSPERO registration This review was registered on PROSPERO, the International Prospective Register of Systematic Reviews, on 29 March 2026 (CRD420261354546; https://www.crd.york.ac.uk/prospero/). At registration, the record stated that searching and screening for an earlier version of the review had been completed and that data extraction had not started. That earlier version was posted as a preprint on 31 March 2026 (Preprints.org, doi:10.20944/preprints202603.2465.v1); it is superseded by this report. Its withdrawal was requested from Preprints.org on 18 October 2026 and had not been confirmed at the time of resubmission. The review reported here was re-conducted in full after registration: the searches were run on 27–28 September 2026, and screening, data extraction and risk-of-bias assessment followed. The registered eligibility criteria were not changed in substance; the registered deep-learning criterion was operationalised before full-text assessment and, after peer review, applied more strictly (a post hoc change reported in the manuscript). The search sources, risk-of-bias tool and synthesis method differ from the registered protocol. Every difference, with its timing and reason, is reported in Table 1 of the manuscript, and an amendment describing them has been submitted to PROSPERO.