Ilmu Komputer & AI editorial
Open AccessOA2026
A Leakage-Aware Multimodal Evaluation Framework for Early Intraoperative Acute Kidney Injury Prediction
SynerT, SynerT-MM, and SynerT-Stack: waveform-only temporal backbones, late-fusion multimodal extensions, and leakage-safe stacked ensembles for AKI risk stratification within the first 60 intraoperative minutes
Quang Minh Nguyen; Duc Minh Le; Ho Nhat Minh Nguyen; Thuy Quynh Nguyen; Trong Nghia Nguyen· 2026· DOI 10.48550/arXiv.2609.26848
The core problem
Postoperative acute kidney injury (AKI) after major non-cardiac surgery carries substantial morbidity, yet early intraoperative risk stratification remains difficult. The authors address this gap by proposing a family of models for AKI risk prediction restricted to information available within the first 60 intraoperative minutes. The central hypothesis is that waveform-only temporal modeling is insufficient under strict early constraints, and that structured clinical context must be integrated in a leakage-aware manner. The study is a retrospective cohort analysis on VitalDB, a high-fidelity perioperative database, comprising 2,413 waveform-usable cases with 180 AKI-positive cases (7.46% prevalence). The work introduces three model variants: SynerT, a waveform-only hybrid temporal backbone; SynerT-MM, a late-fusion multimodal extension; and SynerT-Stack, a leakage-safe stacked ensemble. The evaluation framework is explicitly leakage-aware, ensuring that no future information contaminates predictions at any stage.
Innovation
Among 2,413 waveform-usable cases (180 AKI-positive; 7.46% prevalence), SynerT fell well below strong structured-data baselines, demonstrating that waveform-only temporal modeling is insufficient under strict early constraints. SynerT-MM recovered discrimination by incorporating hemodynamic burden summaries and preoperative covariates. SynerT-Stack achieved the best overall performance across AUROC, AUPRC, and -max. Cross-fitted Platt recalibration substantially corrected calibration defects in both multimodal variants. Decision-curve analysis confirmed the recalibrated stacked model delivered the strongest net clinical benefit across low-to-intermediate thresholds. These results indicate that leakage-aware multimodal integration and stacked ensembling are essential for robust early intraoperative AKI prediction.
Postoperative acute kidney injury (AKI) after major non-cardiac surgery carries substantial morbidity, yet early intraoperative risk stratification remains difficult. The authors address this gap by proposing a family of models for AKI risk prediction restricted to information available within the first 60 intraoperative minutes. The central hypothesis is that waveform-only temporal modeling is insufficient under strict early constraints, and that structured clinical context must be integrated in a leakage-aware manner. The study is a retrospective cohort analysis on VitalDB, a high-fidelity perioperative database, comprising 2,413 waveform-usable cases with 180 AKI-positive cases (7.46% prevalence). The work introduces three model variants: SynerT, a waveform-only hybrid temporal backbone; SynerT-MM, a late-fusion multimodal extension; and SynerT-Stack, a leakage-safe stacked ensemble. The evaluation framework is explicitly leakage-aware, ensuring that no future information contaminates predictions at any stage.
The SynerT backbone combines a causal dilated temporal convolutional network (TCN) with a hierarchy of dilated recurrent layers to encode early intraoperative physiologic trajectories. Formally, for an input waveform sequence
over the first 60 minutes, the causal dilated TCN computes hierarchical temporal features:
Why it matters
The findings underscore a critical methodological insight: under strict early constraints, waveform-only temporal modeling is insufficient for AKI risk stratification. The failure of SynerT relative to structured-data baselines highlights that physiologic waveforms alone, within the first 60 intraoperative minutes, do not capture the full predictive signal. The recovery of discrimination in SynerT-MM demonstrates the value of structured clinical context, specifically hemodynamic burden summaries and preoperative covariates. The superiority of SynerT-Stack across AUROC, AUPRC, and -max suggests that leakage-safe stacked ensembling effectively combines complementary information from multimodal and tabular sources. The substantial calibration correction via cross-fitted Platt recalibration is clinically important, as decision-curve analysis showed the recalibrated stacked model delivered the strongest net clinical benefit across low-to-intermediate thresholds. The leakage-aware evaluation framework is a key contribution, ensuring that reported performance reflects realistic early prediction conditions. Limitations include the retrospective design, the single-center VitalDB cohort, and the modest AKI prevalence of 7.46%, which may affect generalizability. Future work should validate these models prospectively and across diverse perioperative populations.
Who should read this
CS practitioners and researchers
Opening member content…