A Multilayer Adaptive Neural Network with Eigenspectrum Derived Features for Predicting Postoperative Mortality Using the VitalDB Intraoperative Database (Preprint)

BACKGROUND Background. Postoperative death is a persistent global concern in perioperative medicine. Established risk indices are simple but their granularity is coarse. OBJECTIVE Objective. The aim of this study is the construction of a mortality prediction model in which the spectral structure of intraoperative signals is used as the primary representation. METHODS Methods. The VitalDB intraoperative database was used as the data source. A cohort of 6388 surgical cases was analysed and 83 deaths within 30 days were recorded. Twenty intraoperative channels were resampled and standardised against cohort statistics. A channel covariance matrix was formed for every case and twelve leading eigenvalues were retained after eigen decomposition. Spectral entropy and the leading eigengap and two volume terms were also retained. These sixteen descriptors were fused with 22 static clinical variables. A four layer network was trained in which neuron gating and layer specific learning rates are adaptive and gradient variance is the modulating statistic. RESULTS Results. The proposed model reached an AUROC of 0.938 and an AUPRC of 0.341 which is 27 times the outcome prevalence. Gradient boosting was the strongest baseline and its AUROC was 0.896. Every pairwise difference was significant under the DeLong test. At the selected threshold 123 alerts per 1000 operations were expected and one death was found for every 11.5 alerts. External temporal validation produced an AUROC of 0.912. CONCLUSIONS Conclusions. Spectral structure of intraoperative waveforms is informative for postoperative mortality prediction. The adaptive network exploits that structure reliably. Recalibration and multicentre validation are required before clinical use.