• HOME
  • NEWS
  • EXPLORE
    • CAREER
      • Companies
      • Jobs
    • EVENTS
    • iGEM
      • News
      • Team
    • PHOTOS
    • VIDEO
    • WIKI
  • BLOG
  • COMMUNITY
    • FACEBOOK
    • INSTAGRAM
    • TWITTER
Wednesday, November 18, 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

Cancer Malnutrition Risk Models Show High Discrimination but Inconsistent Validation

by
November 11, 2026
in Health
Reading Time: 6 mins read
0
Cancer Malnutrition Risk Models Show High Discrimination but Inconsistent Validation

Cancer Malnutrition Risk Models Show High Discrimination but Inconsistent Validation

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

A comprehensive scoping review published in Frontiers in Nutrition has evaluated the emerging landscape of machine learning prediction models designed to identify malnutrition and nutritional risk among adult patients with cancer. Led by researchers Ruibin Xing, Mengting Zhu, Lengjianghai Zheng, Xiaoqin Wang, Sisi Guo, and Dan Zhou from institutions including the People’s Hospital of Deyang City, Chengdu University, and Xi’an Jiaotong University, the investigation systematically mapped nineteen studies published between 2021 and 2026. While the surveyed predictive models frequently exhibited impressive discriminative accuracy, with reported area under the receiver operating characteristic curve values spanning from 0.770 to 1.000, the authors emphasized that current evidence remains insufficient to establish the clear superiority of advanced machine learning algorithms over conventional logistic regression. The review highlighted that despite substantial technical enthusiasm, widespread clinical implementation in oncology wards is currently premature due to pervasive methodological vulnerabilities, heterogeneous definitions of nutritional endpoints, infrequent calibration testing, and a pronounced deficiency of independent external validation across diverse patient populations.

Malnutrition represents a profound clinical challenge in oncology, driven by complex disease-related metabolic alterations, systemic inflammatory cascades, reduced oral intake, and treatment-induced toxicities. Previous literature indicates that thirty to eighty percent of patients with cancer experience malnutrition across varying phases of their disease trajectory, with vulnerability notably elevated in gastrointestinal, head and neck, and other malignancies that directly impede ingestion, nutrient absorption, or metabolic processing. Nutritional depletion accelerates muscle wasting, induces functional deterioration, compromises tolerance to antineoplastic therapies, prolongs hospital stays, and increases postoperative complications and overall mortality. Standard clinical screening and assessment tools, including the Nutritional Risk Screening 2002, the Patient-Generated Subjective Global Assessment, and the Global Leadership Initiative on Malnutrition criteria, remain foundational in routine practice. Nevertheless, the investigators noted that these tools were conceived primarily to identify established, concurrent nutritional deficits rather than to generate forward-looking risk forecasts. Furthermore, manual administration and reliance on patient recollection of weight history introduce subjectivity and potential underreporting, prompting growing interest in algorithmic risk prediction capable of detecting impending malnutrition before overt clinical deterioration occurs.

To rigorously examine whether artificial intelligence fulfills this preventive objective, the review team adhered to the updated Joanna Briggs Institute scoping review framework and registered their protocol under the identifier 10.17605/OSF.IO/2V3UA. The research questions were structured around the Participants, Concept, and Context framework, focusing specifically on adult oncology patients aged eighteen years or older, machine learning models targeting malnutrition or nutritional risk, and clinical oncology practice settings. The systematic search encompassed nine major international and Chinese electronic literature databases: PubMed, Embase, the Web of Science Core Collection, Scopus, the Cochrane Library, the China National Knowledge Infrastructure, WanFang Data, the VIP Database, and the Chinese Biomedical Literature Database. Screening records from database inception through March 16, 2026, without preliminary language restrictions during initial searches, two independent reviewers identified eligible original investigations that developed, validated, or compared machine learning architectures, resolving discrepancies through structured consensus or adjudication with a third reviewer.

The nineteen eligible studies demonstrated a rapid acceleration in research activity, with sixteen publications, representing 84.2 percent of the corpus, appearing between 2024 and 2026. Geographically, eighteen studies originated from medical institutions in China, alongside one multicenter investigation conducted across hospital centers in Spain by Vilchez-López and colleagues. Methodologically, retrospective cohort designs predominated, accounting for ten studies or 52.6 percent of the total sample, followed by five cross-sectional investigations representing 26.3 percent, three prospective cohort studies comprising 15.8 percent, and one combined retrospective-prospective study representing 5.3 percent. Study sample sizes varied substantially, ranging from a focused cohort of 164 surgical patients in a cross-sectional study by Yang and colleagues to large retrospective datasets encompassing 4,487 individuals analyzed by Wu and colleagues. The underlying clinical populations spanned diverse oncological specialties, prominently featuring cohorts treated for colorectal cancer, gastric cancer, lung cancer, oral cancer, ovarian malignancies, nasopharyngeal carcinoma, as well as heterogeneous mixed cancer cohorts receiving inpatient oncology care.

A critical finding of the scoping review was the pronounced heterogeneity surrounding nutritional outcome definitions, which precluded direct quantitative meta-analysis. The reported prevalence of primary study endpoints ranged from 11.3 percent to 60.0 percent across the nineteen investigations, reflecting fundamentally disparate clinical constructs rather than equivalent manifestations of malnutrition. Specifically, studies employing the Global Leadership Initiative on Malnutrition criteria assessed formal diagnostic malnutrition, whereas those applying the Nutritional Risk Screening 2002 or Malnutrition Screening Tool identified nutritional risk necessitating monitoring. Meanwhile, studies relying on the Patient-Generated Subjective Global Assessment utilized score thresholds of two or four to categorize assessment-defined nutritional impairment and intervention demand. Other investigations evaluated surrogate biochemical indices, such as Onodera’s prognostic nutritional index below forty-five or serum albumin below thirty-five grams per liter. The review authors underscored that hypoalbuminemia reflects systemic inflammation and hepatic functional status rather than a validated diagnosis of malnutrition, meaning models predicting low albumin predict biochemical abnormalities rather than comprehensive nutritional failure.

Across the investigated models, researchers incorporated diverse multidimensional predictor sets spanning demographic attributes, anthropometric measurements, routine laboratory parameters, oncological variables, therapeutic exposures, and functional indicators. The most frequently retained predictive variables across multiple studies included chronological age, body mass index, serum albumin concentration, hemoglobin, total serum protein, tumor clinical staging, and documented historical weight loss. Several investigations expanded beyond static blood biochemistry by incorporating functional markers, such as baseline handgrip strength, calf circumference, bioelectrical impedance phase angle, and scores from activities of daily living assessments. Notably, Huang and colleagues engineered an advanced multimodal architecture that merged computed tomography deep learning image features with conventional clinical parameters, illustrating an evolving trend toward integrating radiographic body composition with standard clinical data.

Methodological examination revealed that studies commonly evaluated multiple algorithmic frameworks concurrently, with conventional multivariable logistic regression appearing in sixteen of the nineteen publications (84.2 percent) primarily as a baseline performance comparator. Among distinct machine learning techniques, support vector machines were evaluated in twelve studies (63.2 percent), followed by random forests in ten studies (52.6 percent), extreme gradient boosting and decision trees in nine studies each (47.4 percent), k-nearest neighbors in seven studies (36.8 percent), naïve Bayes classifiers in six studies (31.6 percent), and artificial neural network or multilayer perceptron architectures in five studies (26.3 percent). Gradient boosting variants including LightGBM, AdaBoost, and traditional gradient boosting machines, alongside partial least squares algorithms, were also represented. Across the sixteen studies directly comparing machine learning against logistic regression within identical datasets, fifteen reported that at least one machine learning model achieved a numerically higher area under the curve. When identifying overall optimal models within individual reports, extreme gradient boosting and decision trees each led in five studies, random forest in four, and single studies favored logistic regression, artificial neural networks, support vector machines, LightGBM, and multimodal deep learning.

Despite encouraging discriminative metrics, systematic critical appraisal utilizing the Prediction model Risk of Bias Assessment Tool for Artificial Intelligence revealed substantial methodological weaknesses across the literature. In the model development phase, only three studies were judged to have achieved high overall methodological quality, while fourteen studies exhibited low quality and two remained unclear; similarly, only three studies demonstrated good clinical applicability, with fifteen rated low and one unclear. In model evaluation, four studies achieved low risk of bias, whereas thirteen carried high risk and two were categorized as unclear, with sixteen studies classified as having low clinical applicability. Pervasive deficiencies included reliance on single-center retrospective cohorts, small sample sizes yielding inadequate outcome event counts relative to candidate predictor complexity, and opaque data handling. The authors noted that these structural limitations substantially amplify the risks of algorithmic overfitting and statistical optimism, casting doubt on whether published model performance can be replicated across broader clinical environments.

The review underscored that discriminative capacity represents only a single facet of predictive performance and should not be conflated with clinical readiness. While high area under the curve scores indicate effective ranking of patients according to relative risk, they provide no assurance that calculated individual probabilities correspond accurately to actual clinical event rates. Calibration analysis, which evaluates concordance between observed and predicted risk probabilities across risk deciles, was evaluated inconsistently or omitted entirely across several included reports. Furthermore, decision curve analysis and net-benefit assessments measuring real-world clinical utility at actionable therapeutic thresholds were rarely documented. The authors highlighted that seemingly extraordinary statistical outputs, such as the support vector machine reported by Yang and colleagues achieving an area under the curve of 1.000 in a dataset of just 164 patients relying exclusively on internal validation, warrant extreme scientific caution. Such near-perfect performance frequently signals subtle data leakage during preprocessing or acute model overfitting rather than genuine prognostic discovery, highlighting the indispensable need for rigorous feature-selection boundaries.

Ultimately, the scoping review concludes that while machine learning algorithms offer promising computational capabilities for modeling complex, non-linear relationships among clinical predictors of malnutrition risk, existing evidence does not justify routine clinical adoption. The authors called for future research to prioritize methodological rigor over excessive algorithmic complexity. Future studies must recruit adequately sized cohorts, establish clear event-to-variable ratios, isolate hyperparameter tuning and feature selection within robust resampling frameworks, and adhere strictly to TRIPOD+AI reporting guidelines and PROBAST+AI quality standards. Crucially, researchers must transition away from single random train-test splits toward prospective validation and independent external cohorts across diverse international healthcare systems. By standardizing nutritional diagnostic endpoints and evaluating calibration alongside decision-analytic utility, future investigators can determine whether artificial intelligence prediction tools genuinely improve clinical decision-making, optimize timely nutritional support, and enhance therapeutic outcomes for patients undergoing cancer therapy.

Subject of Research: Machine learning prediction models for malnutrition risk in adult cancer patients

Article Title: Machine learning-based prediction models for malnutrition risk in patients with cancer: a scoping review

Article References: Xing, R., Zhu, M., Zheng, L., Wang, X., Guo, S., & Zhou, D. (2026). Machine learning-based prediction models for malnutrition risk in patients with cancer: a scoping review. Frontiers in Nutrition, 13, Article 1944563. https://doi.org/10.3389/fnut.2026.1944563

Image Credits: AI Generated

DOI: 10.3389/fnut.2026.1944563

Keywords: machine learning, malnutrition risk, cancer oncology, scoping review, PROBAST+AI, predictive performance, area under the curve, logistic regression, extreme gradient boosting, random forest, decision tree, support vector machine, GLIM criteria, PG-SGA, NRS-2002, external validation, model calibration, serum albumin

News Source: Nathaniel Bowman. (November 11, 2026). Cancer Malnutrition Risk Models Show High Discrimination but Inconsistent Validation. Scienmag.

Tags: area under the curvecancer oncologydecision treeexternal validationextreme gradient boostingGLIM criterialogistic regressionMachine Learningmalnutrition riskmodel calibrationNRS-2002PG-SGApredictive performancePROBAST+AIRandom Forestscoping reviewSerum albuminsupport vector machine
Share12Tweet7Share2ShareShareShare1

Related Posts

Lower Pre-PCI Serum NRF2 Levels Correlate With In-Hospital Events After STEMI

Lower Pre-PCI Serum NRF2 Levels Correlate With In-Hospital Events After STEMI

November 18, 2026
Brucella Abortus Detected Across Livestock and Humans in Tanzania's Mvomero District

Brucella Abortus Detected Across Livestock and Humans in Tanzania’s Mvomero District

November 18, 2026

Global Pain Consensus Standardizes Pulsed Radiofrequency Protocols and Stepwise Treatment Algorithms

November 18, 2026

Cystic Fibrosis Restricts Systemic Lipid Modulation During Lung Infection In Rats

November 18, 2026

POPULAR NEWS

  • Epilepsy Marks Lower Adaptive Functioning in Boys With MECP2 Duplication Syndrome

    Epilepsy Marks Lower Adaptive Functioning in Boys With MECP2 Duplication Syndrome

    29 shares
    Share 12 Tweet 7
  • Geno2pheno Overestimates CXCR4 Tropism in Chinese HIV-1 CRF01_AE Clinical Isolates

    29 shares
    Share 12 Tweet 7
  • Climate Variability Drives Growing Global Risks Across Foodborne Pathogen Pathways

    29 shares
    Share 12 Tweet 7
  • Catheter Drainage Reduces Surgical Burden While Endoscopy Accelerates Early Hemorrhage Clearance

    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

Systematic Review Finds Collaborative Multi-Agent Networks Dominate Research While Autonomy Remains Constrained

Lower Pre-PCI Serum NRF2 Levels Correlate With In-Hospital Events After STEMI

Brucella Abortus Detected Across Livestock and Humans in Tanzania’s Mvomero District

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.