Improved Predictive Models for Acute Kidney Injury with IDEAs: Intraoperative Data Embedded Analytics

Acute kidney injury (AKI) is a common and serious complication after a surgery which is associated with morbidity and mortality. The majority of existing perioperative AKI risk score prediction models are limited in their generalizability and do not fully utilize the physiological intraoperative time-series data. Thus, there is a need for intelligent, accurate, and robust systems, able to leverage information from large-scale data to predict patient’s risk of developing postoperative AKI. A retrospective single-center cohort of 2,911 adult patients who underwent surgery at the University of Florida Health has been used for this study. We used machine learning and statistical analysis techniques to develop perioperative models to predict the risk of AKI (risk during the first 3 days, 7 days, and until the discharge day) before and after the surgery. In particular, we examined the improvement in risk prediction by incorporating three intraoperative physiologic time series data, i.e., mean arterial blood pressure, minimum alveolar concentration, and heart rate. For an individual patient, the preoperative model produces a probabilistic AKI risk score, which will be enriched by integrating intraoperative statistical features through a machine learning stacking approach inside a random forest classifier. We compared the performance of our model based on the area under the receiver operating characteristics curve (AUROC), accuracy and net reclassification improvement (NRI). The predictive performance of the proposed model is better than the preoperative data only model. For AKI-7day outcome: The AUC was 0.86 (accuracy was 0.78) in the proposed model, while the preoperative AUC was 0.84 (accuracy 0.76). Furthermore, with the integration of intraoperative features, we were able to classify patients who were misclassified in the preoperative model.

💡 Research Summary

Acute kidney injury (AKI) remains a frequent and serious postoperative complication, yet most existing peri‑operative risk scores suffer from limited generalizability and do not exploit the wealth of intra‑operative physiological time‑series data that modern electronic health records capture. To address this gap, the authors conducted a retrospective cohort study at the University of Florida Health, including 2,911 adult patients who underwent a variety of surgical procedures. AKI was defined according to KDIGO criteria and examined at three clinically relevant windows: within 3 days (AKI‑3day), within 7 days (AKI‑7day), and up to hospital discharge (AKI‑Discharge).

The dataset was divided into two major components. Pre‑operative information comprised demographics, comorbidities, baseline laboratory values (creatinine, eGFR, etc.), and operative details such as case duration and estimated blood loss. Intra‑operative data focused on three continuously recorded vital signs: mean arterial pressure (MAP), minimum alveolar concentration (MAC, a proxy for anesthetic depth), and heart rate (HR). Raw waveforms were resampled into 5‑minute epochs, and for each epoch a set of twelve statistical descriptors was computed (mean, standard deviation, min, max, coefficient of variation, and measures of abrupt change). These derived features were intended to capture both overall exposure (e.g., sustained hypotension) and dynamic instability (e.g., rapid MAP swings).

Model development followed a two‑stage stacking architecture. In the first stage, several “base” learners—logistic regression, XGBoost, and LightGBM—were trained exclusively on pre‑operative variables to generate an initial probability of AKI for each patient. Hyper‑parameters were tuned via nested cross‑validation to avoid over‑fitting. In the second stage, the output probabilities from the base models were concatenated with the intra‑operative statistical features and fed into a random‑forest classifier. This stacking approach allows the second‑stage model to correct systematic biases of the base learners while exploiting non‑linear interactions between pre‑operative risk factors and intra‑operative physiologic stressors.

Performance was evaluated using the area under the receiver operating characteristic curve (AUROC), overall accuracy, and net reclassification improvement (NRI). For the primary endpoint—AKI within 7 days—the pre‑operative‑only model achieved an AUROC of 0.84 and accuracy of 0.76. Incorporating intra‑operative features via the stacked random forest raised the AUROC to 0.86 and accuracy to 0.78. The NRI of 0.12 indicated that the integrated model correctly re‑classified a substantial proportion of patients who were mis‑identified by the pre‑operative model; specifically, 18 % of high‑risk patients previously labeled low‑risk were reassigned to the appropriate risk tier. Feature importance analysis highlighted that low MAP averages, high MAP variability, abrupt MAC fluctuations, and sudden HR accelerations were independent predictors of AKI, beyond the traditional demographic and comorbidity variables.

The authors acknowledge several limitations. First, the study is single‑center, raising concerns about external validity; a multi‑institutional validation cohort is needed before clinical deployment. Second, the reduction of raw waveforms to simple summary statistics may discard nuanced temporal patterns (e.g., brief but severe hypotensive episodes) that more sophisticated time‑series models could capture. Third, MAC reflects anesthetic concentration but does not differentiate among agents, dosing strategies, or patient‑specific pharmacokinetics, potentially blunting its predictive granularity.

Future directions proposed include: (1) external validation across diverse health systems; (2) application of deep learning architectures such as LSTM or Transformer networks to model raw intra‑operative waveforms directly; (3) integration of peri‑operative biomarkers (e.g., NGAL, cystatin‑C) to create a multimodal risk score; and (4) development of a real‑time decision support dashboard that alerts clinicians when a patient’s evolving intra‑operative profile crosses a predefined AKI risk threshold.

In summary, this work demonstrates that embedding intra‑operative physiological data into a machine‑learning stacking framework yields a modest but statistically and clinically meaningful improvement over traditional pre‑operative risk scores for postoperative AKI. The approach showcases the potential of data‑driven peri‑operative analytics to identify high‑risk patients earlier, thereby opening avenues for targeted preventive strategies and ultimately improving surgical outcomes.