• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Tuesday, November 17, 2026
BIOENGINEER.ORG
No Result
View All Result
  • Login
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
  • HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
        • Lecturer
        • PhD Studentship
        • Postdoc
        • Research Assistant
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
No Result
View All Result
Bioengineer.org
No Result
View All Result
Home NEWS Science News Health

Machine Learning Predicts Post-Cardiac Surgery Kidney Injury Using Routine Perioperative Data

by
November 17, 2026
in Health
Reading Time: 6 mins read
0
Machine Learning Predicts Post-Cardiac Surgery Kidney Injury Using Routine Perioperative Data

Machine Learning Predicts Post-Cardiac Surgery Kidney Injury Using Routine Perioperative Data

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A retrospective benchmark investigation evaluating 210 distinct machine learning combinations has demonstrated that routine preoperative and intraoperative variables collected up to the conclusion of cardiac surgery can predict the development of cardiac surgery-associated acute kidney injury with modest discrimination. Published in Frontiers in Medicine by a collaborative research team based in China, the study analyzed clinical records from 1,352 consecutive adult patients who underwent open-chest operations at Linyi People’s Hospital between January 2016 and December 2024. Cardiac surgery-associated acute kidney injury represents a frequent and severe complication that independently increases patient mortality, prolongs intensive care unit stays, and accelerates the transition toward chronic kidney disease. A significant clinical hurdle to intervening effectively has been timing, as standard diagnostic definitions rely on serum creatinine elevations or oliguria that often take forty-eight to seventy-two hours to manifest clinically, after substantial renal tissue damage has already occurred.

To establish a practical predictive window before patients depart the operating room, the researchers focused exclusively on clinical variables recorded up to the precise landmark where surgery concluded. While previous machine learning models in this domain have reported higher discriminatory metrics, the investigators noted that those existing algorithms frequently relied on postoperative monitoring data such as serial creatinine measurements, fluid balances, and hemodynamic shifts captured inside the intensive care unit. Such models identify kidney damage as it is actively evolving rather than anticipating risk prior to injury. Additionally, the authors highlighted that many published clinical algorithms suffer from methodological fragility, relying upon unvalidated feature selection routines or single, arbitrary development and testing data splits without pre-specified evaluation protocols. In response, this investigation established a transparent, locked analysis protocol conforming to the TRIPOD plus AI reporting guidelines to examine predictive capacity rigorously.

The study cohort comprised 1,352 adult patients, of whom 211 individuals, or 15.6 percent, met the primary clinical endpoint of acute kidney injury within seven days following surgery. The outcome was defined strictly according to the 2012 Kidney Disease: Improving Global Outcomes creatinine criteria as a rise in serum creatinine of at least 26.5 micromoles per liter within forty-eight hours or an elevation reaching at least 1.5 times the preoperative baseline within one week. Because urine output volumes were not available in the electronic medical records and staging was unrecorded, creatinine served as the sole diagnostic marker. The total cohort exhibited a median age of 64 years, with an interquartile range of 57 to 69 years, and included 803 women, representing 59.4 percent of the population. Valve surgery was the most frequent surgical category, accounting for 648 patients or 47.9 percent, followed by isolated coronary artery bypass grafting in 480 patients or 35.5 percent, and combined procedures in 224 patients or 16.6 percent.

From an initial pool of 44 electronic health record variables, the investigators systematically removed seven postoperative medications that lacked uniform administration timeframes, four categorical variables that merely duplicated continuous laboratory measurements, and estimated creatinine clearance, which was calculated directly from existing physiological fields. This curation left 32 routine candidate predictors spanning 17 continuous and 15 categorical variables measured before or during the operation. Continuous measurements included hematological and metabolic panels such as baseline creatinine, serum albumin, uric acid, hemoglobin, and total homocysteine, alongside intraoperative records including surgical duration, estimated blood loss, red blood cell transfusions, and fluid infusion volumes. Categorical features documented procedural classifications including on-pump cardiopulmonary bypass versus off-pump approaches, surgical types, intra-aortic balloon pump support, biological sex, and established cardiovascular comorbidities including hypertension, diabetes mellitus, and heart failure.

The patient records were divided via random stratified sampling into a fixed development dataset containing 1,081 individuals and an independent test dataset holding 271 patients. The researchers systematically constructed and benchmarked 210 model combinations exclusively within the development cohort, crossing five distinct feature-selection strategies with seven learning algorithms and six class-weighting schemes. Feature selection included utilizing all 32 predictors, applying a LASSO penalty under the one-standard-error rule, selecting features with out-of-fold SHAP values exceeding 0.01, and retaining variables demonstrating nominal significance in univariable or multivariable logistic regressions. The algorithmic architectures evaluated were L2-regularized logistic regression, elastic net, random forest, support vector machines utilizing radial basis function kernels, XGBoost, CatBoost, and multilayer perceptrons. Class weighting spanned unweighted baselines, positive-class weights of 1.5, 2, 3, and 4, and balanced weighting. Internal performance was generated across ten repeats of stratified five-fold cross-validation.

Under a pre-specified, two-stage selection protocol designed to prioritize screening utility, the primary model selected was a random forest algorithm with a positive-class weight of 1.5 and isotonic probability recalibration. Evaluated on the held-out test cohort of 271 patients containing 42 acute kidney injury events, this final model yielded an area under the receiver operating characteristic curve of 0.722, accompanied by a 95 percent confidence interval spanning 0.634 to 0.799. The precision-recall area under the curve reached 0.363, and the Brier score was 0.124. At the pre-specified screening decision threshold of 0.130, which was selected on the development set to maximize specificity while targeting a sensitivity of at least 0.80, the model achieved an observed test sensitivity of 0.762, a specificity of 0.515, a positive predictive value of 0.224, and a negative predictive value of 0.922.

A critical technical finding of the systematic benchmark was that no machine learning model family displayed decisive superiority over standard statistical techniques. Cross-validated areas under the curve across the 35 selected feature-algorithm combinations clustered tightly between 0.67 and 0.76. Simpler penalized linear formulations, specifically regularized logistic regression and elastic net models, achieved cross-validated performance within 0.02 of the leading tree-based models and matched or numerically exceeded the final random forest upon the held-out test set, where LASSO logistic regression attained a test area under the curve of 0.764. DeLong tests confirmed no statistically significant performance differences among these leading architectures. In contrast, multilayer perceptron neural networks exhibited the lowest discrimination, recording cross-validated values between 0.674 and 0.709, illustrating that high-capacity neural architectures remain prone to overfitting when applied to clinical tabular cohorts of modest dimension.

The researchers also explicitly investigated whether techniques tailored for small or moderate tabular datasets could enhance predictive capability. Evaluating TabPFN, a pre-trained synthetic tabular transformer applied in a few-shot capacity, yielded a cross-validated area under the curve of 0.752 and a test area under the curve of 0.730, failing to surpass the baseline random forest. Similarly, creating simple ensembled predictions by averaging output probabilities across top single models reached a test area under the curve of 0.743, remaining below regularized linear baselines. Crucially, bootstrap optimism correction evaluated over 500 resamples drastically reduced the apparent in-sample area under the curve of the random forest from an inflated 0.965 down to an optimism-corrected 0.724. This corrected internal estimate closely mirrored the actual independent test set discrimination of 0.722, highlighting how uncorrected training metrics generate severe overoptimism in complex non-linear models.

Interpretability assessments using SHAP values identified preoperative total homocysteine concentration as the leading individual predictor within the final random forest model, contributing an average absolute importance of 0.041 that doubled the contribution of the second-ranking feature, surgery duration at 0.022. Subsequent high-ranking variables included surgical approach, preoperative serum creatinine, transfusion volume, surgery category, red blood cell units, serum albumin, intra-aortic balloon pump use, and uric acid. Independent multivariable logistic regression across 1,338 complete patient profiles corroborated these patterns, confirming ten statistically independent factors associated with acute kidney injury. Preoperative homocysteine exhibited an adjusted odds ratio of 1.38 per 5 micromoles per liter increment, having attenuated only slightly from a crude odds ratio of 1.49, reinforcing experimental evidence linking hyperhomocysteinemia with endothelial injury, oxidative stress, and renal microvascular dysfunction.

The multivariable regression also highlighted prominent surgical and physiological risk indicators, including intraoperative intra-aortic balloon pump deployment with an adjusted odds ratio of 13.53, on-pump cardiopulmonary bypass with an odds ratio of 3.86, diabetes mellitus at 1.89, heart failure at 1.69, preoperative renin-angiotensin system inhibitor therapy at 1.65, hypertension at 1.54, intraoperative blood transfusion at 1.33 per 500 milliliters, and surgical duration at 1.18 per hour, while higher hemoglobin was protective with an odds ratio of 0.87 per 10 grams per liter. Because balloon pump placement was rare, involving 24 patients, the authors performed Firth penalized logistic regression to guard against sparse-data separation artifacts, obtaining a consistent adjusted odds ratio of 12.14. Pre-specified sensitivity analyses further demonstrated model stability, as removing balloon pump variables slightly lowered test discrimination to 0.699, while complete-case analysis maintained a test area under the curve of 0.723.

To support potential clinical deployment where data collection burden and model transparency represent primary concerns, the investigators evaluated parsimonious alternatives derived through stability selection across 300 random sub-samples. A compact 23-variable regularized linear model achieved a test area under the curve of 0.751, a sensitivity of 0.857, and a specificity of 0.546, performing comparably to the full 32-variable random forest without requiring complex ensemble or tree structures. Conversely, a more aggressively restricted seven-variable model experienced a drop in discrimination to an area under the curve of 0.677. Decision curve analysis indicated positive net clinical benefit across practical decision thresholds for the primary model, indicating potential clinical utility as an early postoperative warning tool to guide timely renal protective strategies, such as hemodynamic optimization and intensified monitoring, rather than acting as a definitive diagnostic instrument.

The study authors explicitly detailed several key limitations that require caution prior to translating these findings into routine practice. The retrospective investigation was restricted to a single medical center and lacked external validation on independent hospital cohorts, which represents a mandatory prerequisite before clinical implementation. Furthermore, the test set contained only 42 acute kidney injury events, which falls below the standard threshold of approximately 100 outcome events recommended for definitive external validation. The investigators also acknowledged that the random stratified partition was less rigorous than a prospective temporal split, that post-hoc comparisons were conducted across the held-out sample, and that model calibration remained imperfect with a test calibration slope of 0.631. Finally, because baseline estimated glomerular filtration rates and urine outputs were unavailable, residual confounding could not be entirely excluded, underscoring that observed associations must not be interpreted as proven causal mechanisms.

Subject of Research: Machine learning prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables

Article Title: Prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables: a pre-specified benchmark of 210 machine learning combinations

Article References: Li, Z., Zhang, W., Wei, Y., Liu, W., Zhang, F., Han, Y., Wang, S., & Li, X. (2026). Prediction of acute kidney injury at the end of cardiac surgery using routine perioperative variables: a pre-specified benchmark of 210 machine learning combinations. Frontiers in Medicine, 13, Article 1992789. https://doi.org/10.3389/fmed.2026.1992789

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1992789

Keywords: acute kidney injury, cardiac surgery, machine learning, perioperative medicine, homocysteine, random forest, penalized logistic regression, risk stratification, TRIPOD-AI, SHAP analysis, stability selection, bootstrap optimism correction, intra-aortic balloon pump, cardiopulmonary bypass, clinical prediction model, TabPFN, internal validation, biomarkers

News Source: Ophelia Keating. (November 17, 2026). Machine Learning Predicts Post-Cardiac Surgery Kidney Injury Using Routine Perioperative Data. Scienmag.

Tags: acute kidney injurybiomarkersbootstrap optimism correctionCardiac Surgerycardiopulmonary bypassclinical prediction modelhomocysteineinternal validationintra-aortic balloon pumpMachine Learningpenalized logistic regressionperioperative medicineRandom Forestrisk stratificationSHAP analysisstability selectionTabPFNTRIPOD + AI
Share12Tweet7Share2ShareShareShare1

Related Posts

Meta-Analysis Finds Nearly One-Third of Ethiopian Newborns Receive Harmful Cord Care

Meta-Analysis Finds Nearly One-Third of Ethiopian Newborns Receive Harmful Cord Care

November 17, 2026
Biliopancreatic Limb Lengths Showed No Significant 36-Month Weight-Loss Differences in Cohort

Biliopancreatic Limb Lengths Showed No Significant 36-Month Weight-Loss Differences in Cohort

November 17, 2026

Professional Footballers Experience Sustained Career Declines After ACL Reconstruction

November 17, 2026

Postoperative Alignment Influences Long-Term Knee Revision Risk Following High Tibial Osteotomy

November 17, 2026

POPULAR NEWS

  • High Prevalence of Gender-Based Violence Follows Armed Conflict in Southern Tigray

    High Prevalence of Gender-Based Violence Follows Armed Conflict in Southern Tigray

    29 shares
    Share 12 Tweet 7
  • Air Pollutants Tied to Adverse Sleep Outcomes Amid Notable Evidence Uncertainty

    29 shares
    Share 12 Tweet 7
  • Network Analysis Identifies Core Positive Development Attributes in Rural Chinese Adolescents

    29 shares
    Share 12 Tweet 7
  • China Silver Economy Policies Show Good Overall Consistency but Uneven Regional Breadth

    29 shares
    Share 12 Tweet 7

About

We bring you the latest biotechnology news from best research centers and universities around the world. Check our website.

Follow us

Recent News

Meta-Analysis Finds Nearly One-Third of Ethiopian Newborns Receive Harmful Cord Care

Prunus Small Peptide Suppresses Anthocyanin Accumulation in Peach Fruit Flesh

Reprogrammed Bacterial Gene Cluster Creates Selective Production Platform for Potent Antipseudomonal Antibiotic

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 88 other subscribers
  • Contact Us

Bioengineer.org © Copyright 2023 All Rights Reserved.

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • Homepages
    • Home Page 1
    • Home Page 2
  • News
  • National
  • Business
  • Health
  • Lifestyle
  • Science

Bioengineer.org © Copyright 2023 All Rights Reserved.