Integration of intraoperative data in interpretable machine learning models to predict postoperative AKI in noncardiac surgery patients.

Objectives: We aimed to (1) quantify changes in discrimination when adding intraoperative data to preoperative data and (2) compare tabular machine learning with feature engineering against a time-aware LSTM-based model. Materials and Methods: Retrospective cohort of 46 204 adults undergoing 57 055 eligible noncardiac surgery in the INSPIRE database. We extracted 38 preoperative and 49 intraoperative variables; acute kidney injury (AKI) was defined by KDIGO serum creatinine criteria and modeled as stage 2/3 postoperative AKI. Models were trained on preoperative-only and combined pre- and intraoperative data. Intraoperative series were summarized using eight statistical features for tabular models or integrated directly using an MLP+LSTM architecture. Results: GBT with combined features achieved the highest AUROC (0.896, 95% CI, 0.878-0.914), followed by combined AutoGluon (0.893, 95% CI, 0.877-0.909) and preoperative-only GBT (0.891, 95% CI, 0.871-0.910). ASA-PS ≥3 (AUROC 0.723, 95% CI, 0.700-0.746) and adapted GS-AKI (AUROC 0.719, 95% CI, 0.700-0.739) underperformed machine-learning models. The hybrid MLP+LSTM model did not outperform simpler tabular models (AUROC 0.870, 95% CI, 0.848-0.892). Discussion: The small gain from adding low-frequency intraoperative summaries suggests that most discriminative information for stage 2/3 postoperative AKI was available before surgery. Conclusion: Preoperative tabular ML models provided excellent prediction of postoperative AKI, and added limited incremental discrimination at the available sampling frequency. Future work should evaluate whether higher-frequency intraoperative signals better leverage time-aware architectures.