• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Thursday, November 19, 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

Explainable Machine Learning Predicts Unplanned Hospital Admissions After Day-Case Surgery

by
November 19, 2026
in Health
Reading Time: 6 mins read
0
Explainable Machine Learning Predicts Unplanned Hospital Admissions After Day-Case Surgery

Explainable Machine Learning Predicts Unplanned Hospital Admissions After Day-Case Surgery

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A retrospective evaluation of more than 51,000 outpatient surgical procedures in Belgium demonstrates that an intrinsically interpretable machine-learning architecture can forecast unplanned hospital admissions using only routine preoperative data. In modern perioperative systems, ambulatory surgery relies on discharging patients on the date of their operation, an approach that optimizes institutional bed allocation and shortens elective waiting lists. When unexpected complications or clinical needs force an outpatient to be admitted overnight, the abrupt disruption creates substantial logistical and medical friction. Hospital beds intended for other elective cases become occupied, while anaesthesia regimens and oral analgesia protocols originally intended for rapid domestic recovery must be hastily converted into inpatient orders. Researchers Rémi Florquin and Philippe Dony examined whether predictive algorithms could identify these vulnerable surgical candidates prior to anaesthetic induction, comparing transparent predictive structures directly against widely used black-box ensemble methods.

To investigate these care pathway deviations, the research team analyzed an extensive clinical repository derived from anaesthesia information management systems linked to electronic health records across two hospital sites within the Humani network in Belgium, comprising 27 operating rooms. The underlying database had recorded procedural metrics since 2015, but investigators restricted their investigation to consecutive surgical interventions performed between January 2020 and April 2026. This selected chronological window ensured consistency across surgical staffing and anaesthetic practices while compiling an expansive cohort of 142,704 anaesthesia records. The primary clinical outcome was unplanned hospitalisation, defined specifically as the failure of a scheduled same-day discharge that necessitated hospital admission, as well as the reverse occurrence of scheduled inpatients being safely discharged on the day of surgery. Because elective inpatient-to-outpatient pathway conversions were reserved for subsequent inquiry, this analysis focused exclusively on patients scheduled for day hospitalisation.

Preparing this complex clinical database required a sequential, multi-stage data curation pipeline to safeguard model integrity and eliminate data leakage. The raw clinical database initially contained 266 variables spanning demographic details, procedural timings, physiological observations, and categorical intervention codes. Manual clinical screening removed redundant or non-informative parameters, such as the volumetric consumption and financial cost of volatile anaesthetics, as well as identifiable personal details. Crucially, all intraoperative and postoperative variables were systematically excised, including total length of stay, which would act as leaky features by providing algorithms with retrospective cues regarding a patient’s ultimate destination. Highly specialized procedures involving cardiac surgery, neurosurgery, and intensive care were removed, whereas urgent cases permitting same-day recovery, such as diagnostic finger wound explorations, were maintained alongside standard elective cases.

Rigorous automated outlier detection was implemented to eliminate documentation anomalies while preserving genuine clinical variation across divergent age groups. Because physiological anthropometrics differ fundamentally between paediatric and mature populations, the researchers stratified patients by age before screening. The paediatric cohort was evaluated using a Local Outlier Factor algorithm set to a contamination threshold of 1 percent, accommodating the wide growth-dependent variability from infancy to adolescence where miscoding risks are elevated. For the adult subpopulation, an Isolation Forest algorithm was deployed using a conservative contamination factor of 0.1 percent across age, height, and weight. Records with missing core physiological metrics, including American Society of Anesthesiologists physical status scores, extreme body mass indices outside the 10 to 60 range, anaesthetic inductions exceeding 2 hours, or total operative durations exceeding 12 hours were discarded. Following these filtration steps, 100,667 cleaned surgical records remained, from which the final cohort of 51,112 scheduled ambulatory patients was drawn.

Feature engineering addressed categorical complexity and the cyclical structure of chronological metrics. The original dataset captured 2,867 distinct surgical procedure categories, a volume too vast for complete one-hot encoding without compromising computational tractability. The authors explicitly encoded only the 114 surgical procedures that appeared more than 200 times within the registry, aggregating the remaining 2,753 infrequent procedure types into a single consolidated classification. Temporal features presented distinct mathematical obstacles, as standard linear integer scales fail to represent the continuity of circadian schedules, falsely separating adjacent hours such as 23:00 and midnight. The team applied trigonometric cyclic transformations, converting the time of day into paired sine and cosine mathematical components. Days of the week were similarly organized into distinct operational groupings denoting Fridays, weekends, or standard weekdays to reflect procedural scheduling contexts.

To prevent algorithmic memorization and establish an unbiased evaluation framework, the researchers instituted a three-way data partitioning scheme comprising an 80 percent training partition, a 10 percent validation set, and a strictly held-out 10 percent test set. Because unplanned admissions occurred in only 9.08 percent of ambulatory operations, the dataset exhibited marked class imbalance that threatened to bias algorithms toward the majority class of uncomplicated discharges. To counteract this tendency without compromising downstream evaluations, the Synthetic Minority Over-sampling Technique for Nominal and Continuous data, known as SMOTENC, was applied exclusively to the scaled training partition. Prior to oversampling, the training set contained 37,177 standard discharges and 3,712 unplanned admissions; SMOTENC synthesized minority cases until both categories comprised 37,177 instances, totaling 74,354 training observations. Feature scaling via a standard scaler was fitted solely on the training data and applied agnostically to the validation and test sets, ensuring no distributional parameters leaked into the 5,112 held-out test patients.

Seven distinct classification algorithms representing two broad modeling philosophies were trained and evaluated: high-interpretability models consisting of logistic regression with elastic net regularization, a single decision tree, and the explainable boosting machine, alongside complex ensemble architectures including Random Forest, XGBoost, LightGBM, and Histogram-Based Gradient Boosting. Other candidate architectures were deliberately excluded: K-nearest neighbours was deemed impractical given feature dimensionality; support vector machines could not scale computationally to the large sample size; Gaussian naive Bayes violated conditional independence; and deep neural networks were avoided because tree-based methods historically equal or exceed deep networks on cross-sectional tabular clinical records lacking spatial structure.

Model evaluation on the sequestered test cohort revealed significant performance divergence across the selected algorithms. The explainable boosting machine attained the highest overall discrimination, registering an area under the receiver operating characteristic curve of 0.826, which proved statistically superior to all competitors under DeLong testing at adjusted p-values of 0.003 or lower. The ensemble group clustered closely behind the explainable boosting machine, with the Histogram-Based Gradient Boosting classifier achieving an area under the curve of 0.798, Random Forest recording 0.787, LightGBM reaching 0.779, and XGBoost achieving 0.762. The single decision tree achieved an intermediate discrimination of 0.724, whereas regularized logistic regression displayed the lowest discrimination at 0.595, performing only modestly better than random chance.

Because the clinical outcome was uncommon, precision-recall metrics provided the most informative perspective regarding real-world utility, as the area under the precision-recall curve avoids inflation from high true-negative counts. The explainable boosting machine established dominant precision across all recall levels, achieving the highest average precision of 0.472 alongside the top minority-class F1-score of 0.392. Pairwise paired bootstrap testing confirmed the explainable boosting machine’s average precision was significantly superior to all competing architectures at adjusted p-values below 0.001. Random Forest followed with an average precision of 0.303, while LightGBM and Histogram-Based Gradient Boosting tied at 0.268, forming a statistically indistinguishable intermediate tier. Conversely, the basic decision tree and XGBoost trailed at 0.185 and 0.202 respectively, lingering near the no-model random baseline established by the 9.08 percent natural outcome prevalence. Integrating these metrics into a multi-dimensional radar plot demonstrated that the explainable boosting machine enclosed the largest composite operational surface area at 41.5 percent.

At the optimal classification threshold established via Youden’s index on the validation partition, the models exhibited stark operational trade-offs across the 5,112 test patients, which encompassed 464 true unplanned hospitalisations. The Histogram-Based Gradient Boosting architecture delivered the highest sensitivity at 0.966, successfully capturing 448 unplanned admissions, yet it exhibited a specificity of only 0.496 by misclassifying 2,341 uncomplicated patients as deviations. Logistic regression similarly achieved high sensitivity of 0.856 but suffered a catastrophic specificity of 0.276, mislabeling 3,363 ambulatory patients. In contrast, the explainable boosting machine produced the most clinically balanced profile, securing the highest specificity at 0.827 alongside a sensitivity of 0.666 and an F1-score of 0.392, correctly isolating 309 unplanned hospitalisations while generating only 802 false alerts among uncomplicated outpatients.

The authors extended their analysis to evaluate the models under two rigid clinical configurations reflecting opposing health system priorities, quantifying the associated operational burden or price to pay. In the patient safety configuration, decision thresholds were constrained on the validation data to detect 90 percent of unplanned admissions. To achieve this high-sensitivity target on the test set, the Histogram-Based Gradient Boosting model produced the lowest false-positive rate at 45.4 percent, generating 5.1 false alerts for every correctly predicted overnight admission, followed by LightGBM at 51.3 percent and the explainable boosting machine at 61.6 percent. Under the opposing resource preservation configuration, which demanded 90 percent specificity to guarantee same-day discharge certainty, the explainable boosting machine retained superior clinical sensitivity at 55.2 percent, corresponding to a false-negative rate of 44.8 percent and only 0.81 missed admissions per detected case, substantially outperforming logistic regression, which missed 81.2 percent of true hospitalisations.

Assessment of algorithmic stability confirmed that predictive performance remained robust between the validation and test datasets, ruling out severe overfitting. Absolute variations in the area under the receiver operating characteristic curve ranged between negative 0.019 for the explainable boosting machine and positive 0.024 for the decision tree, while F1-score fluctuations remained confined between negative 0.030 and positive 0.018 across all models. However, examination of calibration curves revealed an overarching systemic limitation: all trained models displayed an inclination to overestimate the absolute probability of unplanned admission, an expected consequence of training on artificially balanced SMOTENC distributions. While the models remain immediately effective for ranking patients by relative risk, the authors emphasized that prospective clinical deployment will require formal probability recalibration via isotonic regression or Platt scaling, coupled with external multicentre validation to verify transportability beyond the originating Belgian network.

Subject of Research: Preoperative machine-learning prediction of unplanned hospital admissions after ambulatory surgery

Article Title: Predicting patient care pathway deviations and their consequences in anaesthesia using machine learning

Article References: Florquin, R., & Dony, P. (2026). Predicting patient care pathway deviations and their consequences in anaesthesia using machine learning. Journal of Clinical Monitoring and Computing. https://doi.org/10.1007/s10877-026-01496-y

Image Credits: AI Generated

DOI: 10.1007/s10877-026-01496-y

Keywords: Ambulatory Surgery, Machine Learning, Explainable Boosting Machine, Unplanned Admission, Perioperative Medicine, Clinical Prediction Models, Precision-Recall Curve, Receiver Operating Characteristic, Class Imbalance, SMOTENC, Outlier Detection, Model Calibration, Anaesthesia Informatics, Day-Case Surgery, Health Informatics, Youden Index, Healthcare Resource Allocation

News Source: Ophelia Keating. (November 19, 2026). Explainable Machine Learning Predicts Unplanned Hospital Admissions After Day-Case Surgery. Scienmag.

Tags: ambulatory surgeryAnaesthesia Informaticsclass imbalanceclinical prediction modelsday-case surgeryExplainable Boosting Machinehealth informaticshealthcare resource allocationMachine Learningmodel calibrationoutlier detectionperioperative medicineprecision-recall curvereceiver operating characteristicSMOTENCUnplanned AdmissionYouden index
Share12Tweet7Share2ShareShareShare1

Related Posts

Japanese EMS Personnel Face Protocol and Decision Conflicts in Prehospital DNAR Cases

Japanese EMS Personnel Face Protocol and Decision Conflicts in Prehospital DNAR Cases

November 19, 2026
Study Reveals Outflow of Early-Career Japanese Pharmacists From Hospitals to Pharmacies

Study Reveals Outflow of Early-Career Japanese Pharmacists From Hospitals to Pharmacies

November 19, 2026

Study Identifies Factors Linked to Women’s Health Decision Autonomy in Zambia

November 19, 2026

Computed Tomography Quantifies Regional Lung Mechanics and Injury Patterns in ARDS

November 19, 2026

POPULAR NEWS

  • Study Reveals Outflow of Early-Career Japanese Pharmacists From Hospitals to Pharmacies

    Study Reveals Outflow of Early-Career Japanese Pharmacists From Hospitals to Pharmacies

    29 shares
    Share 12 Tweet 7
  • Mouse and Human Studies Show GATA2 Actively Maintains Müllerian Ducts

    29 shares
    Share 12 Tweet 7
  • Study Identifies Factors Linked to Women’s Health Decision Autonomy in Zambia

    29 shares
    Share 12 Tweet 7
  • Dual-Network Decomposition Model Improves Carbon Price Prediction Intervals Across Chinese Markets

    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

Japanese EMS Personnel Face Protocol and Decision Conflicts in Prehospital DNAR Cases

Study Reveals Outflow of Early-Career Japanese Pharmacists From Hospitals to Pharmacies

Mouse and Human Studies Show GATA2 Actively Maintains Müllerian Ducts

Subscribe to Blog via Email

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

Join 89 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.