• 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

Fisher Information Decomposition Calculates Sample Size for Fair Clinical Prediction Models

by
November 13, 2026
in Health
Reading Time: 7 mins read
0
Fisher Information Decomposition Calculates Sample Size for Fair Clinical Prediction Models

Fisher Information Decomposition Calculates Sample Size for Fair Clinical Prediction Models

Share on FacebookShare on TwitterShare on LinkedinShare on RedditShare on Telegram

Medical researchers developing or updating statistical equations to predict clinical risks face a persistent methodological dilemma when deciding how many study participants are needed. While existing statistical guidelines provide formulas to ensure overall population-level calibration and prevent statistical overfitting, meeting those minimum criteria frequently leaves substantial uncertainty in the risk predictions calculated for specific individual patients. Their analytical approach allows investigators to evaluate before data collection whether a planned study size or an existing observational database will yield sufficiently narrow uncertainty intervals around individual risk estimates. By connecting parameter uncertainty directly to clinical decision thresholds and demographic subgroups, the methodology offers an explicit path to evaluate predictive precision and algorithmic fairness in regression-based clinical prediction models for binary health outcomes.

Clinical prediction models are mathematical algorithms that combine multiple patient characteristics, such as demographic variables, physiological signs, and laboratory biomarkers, to calculate the probability of a health event occurring within a defined timeframe. Common clinical applications range from estimating thirty-day hospital mortality and in-hospital deterioration among individuals diagnosed with infectious illnesses to predicting the development of severe chronic complications. Despite explicit reporting standards established in the TRIPOD statement and recent TRIPOD plus AI guidelines, a substantial proportion of published prediction model studies fail to include an adequate statistical justification for their enrolled sample sizes. Prior guidance established by the authors had defined sample size formulas to estimate overall event risk precisely, minimise predictor overfitting, and limit statistical optimism in apparent model fit. Those baseline requirements correspond to the initial levels of model stability, ensuring dependable estimates across an entire patient population, but they were never intended to guarantee precise estimates for single individuals.

When clinical prediction models are constructed using small sample sizes, parameter estimates for regression coefficients and baseline intercepts inevitably carry substantial variance, leading to wide confidence or credible intervals around estimated individual probabilities. Furthermore, models developed with limited data often demonstrate predictive instability, where minor fluctuations in the underlying development sample cause substantial swings in calculated risk values. The new framework demonstrates that addressing individual-level precision is not merely a mathematical refinement but an essential ethical requirement for healthcare algorithms.

To resolve this problem, the authors formulated a five-step analytical workflow that can be applied prior to gathering prospective data or when auditing an existing electronic health record archive. The first step requires investigators to define a core set of predictive clinical variables, which represent established prognostic indicators identified from prior literature, systematic reviews, or expert consensus. This initial core predictor set must also incorporate key demographic variables, such as age, sex, and ethnicity, whenever algorithmic fairness audits across protected characteristics are required. In the second step, the joint distribution of these selected predictors is characterized, utilizing either directly observed records from pilot databases, synthetic datasets generated through simulation routines, or summary statistics extracted from published baseline tables. In situations where correlations between continuous markers or cross-tabulations of categorical factors are unrecorded, analysts can test multiple plausible correlation scenarios to measure the sensitivity of subsequent sample size estimates.

The third step of the methodology establishes an assumed core prediction model that mathematically describes how individual probabilities depend on core predictor values. In studies aimed at updating or extending an established regression equation, developers can directly adopt the intercept and relative regression coefficients reported by previous publications. When prior coefficients are unavailable or derived from black-box algorithms, the authors outline pragmatic mathematical approximations based on iterative algorithms. By fixing the anticipated overall population outcome risk and setting an expected discrimination C-statistic, researchers can identify suitable regression intercept and scaling terms while assuming equal or specified relative weights across standardized predictors. The fourth step decomposes the inverse of Fisher’s information matrix, which defines the variance-covariance matrix of estimated logistic regression coefficients, into the total study sample size multiplied by the inverse of Fisher’s unit information matrix. Calculating the mathematical components of this unit information matrix across simulated or observed participants provides an expected matrix value that dictates prediction variance.

The fifth step translates these mathematical matrices into practical evaluations of sample size across two study design scenarios designated as Option A and Option B. Under Option A, researchers assess an existing dataset or a planned cohort size to derive anticipated ninety-five percent uncertainty intervals around logit-risk estimates and absolute probabilities for each individual in the target population. Under Option B, researchers working prospectively invert the variance formula to calculate the exact sample size needed to keep prediction variance or uncertainty interval widths below a predetermined clinical threshold. To make these calculations accessible to investigators, the researchers implemented the complete five-step framework in an open-source software program called pmstabilityss, available for both Stata and R statistical environments. This tool generates prediction instability plots displaying individual confidence intervals alongside smooth LOWESS curves that illustrate typical uncertainty across the entire range of absolute risks from zero to one.

The authors demonstrated the practical application of their framework through two clinical case studies, beginning with a planned prospective cohort study to update a prediction model for diabetic foot ulcers over a two-year horizon. The model incorporated three binary predictors: foot insensitivity to monofilament testing, absence of a palpable foot pulse, and prior history of ulceration or amputation. Using previous population-level criteria implemented in the pmsampsize module, an assumed outcome incidence of five point nine percent and a C-statistic of zero point seven seven dictated a minimum development sample of four hundred fifty-three participants, encompassing twenty-seven clinical events. However, applying Option A in pmstabilityss revealed that this baseline sample produced an average ninety-five percent uncertainty interval width of zero point zero eight, with extremes reaching zero point forty-two. Individuals exhibiting the rarest predictor combination had a true risk of forty-nine percent but suffered an uncertainty interval spanning from twenty-eight percent to seventy percent.

Evaluating the diabetic foot ulcer study through Option B demonstrated that guaranteeing an uncertainty interval width of zero point one zero or less across all possible predictor combinations would demand a sample size of nine thousand one hundred twenty-six individuals. The authors explained that such a massive enrollment target is driven entirely by the rarest patient subgroups and is clinically unnecessary when evaluated through formal decision analysis. Conversely, if clinicians adopted a higher intervention threshold of twenty percent due to medication side effects, four hundred fifty-three patients left considerable classification overlap, requiring an intermediate sample of one thousand two hundred twenty-four participants to achieve low classification instability.

The second case study examined whether an existing retrospective intensive care database could support the development of a prediction model for acute kidney injury occurring within forty-eight hours of admission. The investigators extracted a cohort of twenty thousand four hundred thirteen adult intensive care admissions from the Medical Information Mart for Intensive Care III database, in which the observed risk of acute kidney injury was seventeen percent. The model incorporated nine core clinical variables, including age, gender, bicarbonate, serum creatinine, haemoglobin, blood urea nitrogen, potassium, systolic blood pressure, and oxygen saturation. Evaluating this massive cohort demonstrated remarkable individual precision, yielding a mean uncertainty interval width of zero point zero two seven and an average probability of misclassification of only zero point zero two two relative to a ten percent clinical action threshold, confirming the existing dataset was fully sufficient for model construction.

To illustrate the risks of relying strictly on traditional minimum sample sizes, the researchers modeled what would have occurred had the intensive care dataset contained only five hundred eleven patients, the minimum enrollment recommended by population-level criteria for nine predictors. At that reduced sample size, the mean uncertainty interval width expanded more than sixfold to zero point seventeen, with individual intervals extending up to zero point eighty-one, while the average probability of misclassification rose to thirteen percent. A patient with an actual kidney injury risk of twenty percent faced a typical uncertainty interval ranging from ten percent to forty percent, severely undermining clinical utility. When enrollment was hypothetically restricted to two hundred twenty-one participants, the minimum needed solely to estimate overall event risk precisely, prediction intervals became broad and individual classification unstable, demonstrating that standard sample size heuristics can produce clinically unreliable prediction tools.

The acute kidney injury dataset also highlighted the critical relationship between sample size, predictor selection, and algorithmic fairness across demographic subgroups. In the full cohort of twenty thousand four hundred thirteen patients, seventy-two percent were White, eight percent Black, two percent Asian, four percent Hispanic, three percent other, and eleven percent of unknown ethnicity. When ethnicity was omitted from the regression equation, risk prediction precision depended solely on the joint distribution of physiological markers, resulting in identical average uncertainty interval widths of zero point seventeen for White and Asian patients at a sample size of five hundred eleven. However, when investigators added ethnicity as five indicator variables to examine demographic risk differences, the minimum required sample size rose to seven hundred ninety-five participants. At that enrollment, data sparsity among minority groups produced stark disparities: Asian patients experienced uncertainty interval widths and misclassification rates approximately twice as high as White patients, demonstrating that adding demographic predictors without expanding sample sizes can inadvertently compromise predictive precision for minority groups.

The authors stressed that their statistical derivation addresses epistemic uncertainty arising from parameter estimation in standard maximum likelihood logistic regression, rather than irreducible aleatoric variation inherent in biological systems. They characterized the framework as phase-three methodology for regression models expected to exhibit low overfitting, while noting that non-linear machine learning algorithms like random forests typically introduce greater instability and require larger sample sizes. While the decomposition method relies on explicit assumptions regarding core predictor distributions and relative effect sizes, it enables researchers to conduct sensitivity analyses across ranges of plausible C-statistics before committing research funds. By establishing mathematical criteria for individual precision and subgroup equity at the study design stage, the new framework provides clinical researchers and grant funding bodies with an objective quantitative mechanism to ensure prediction models are sufficiently precise, dependable, and fair for personalized healthcare decisions.

Subject of Research: Statistical methodology for determining sample sizes in clinical prediction models

Article Title: A decomposition of Fisher’s information to inform sample size for developing or updating fair and precise clinical prediction models for individual risk—part 1: binary outcomes

Article References: Riley, R. D., Collins, G. S., Whittle, R., Archer, L., Snell, K. I. E., Dhiman, P., Kirton, L., Legha, A., Liu, X., Denniston, A. K., Harrell, F. E., Jr, Wynants, L., Martin, G. P., & Ensor, J. (2025). A decomposition of Fisher’s information to inform sample size for developing or updating fair and precise clinical prediction models for individual risk—part 1: binary outcomes. Diagnostic and Prognostic Research, 9(1), Article 14. https://doi.org/10.1186/s41512-025-00193-9

Image Credits: AI Generated

DOI: 10.1186/s41512-025-00193-9

Keywords: clinical prediction models, sample size calculation, Fisher information matrix, binary outcomes, logistic regression, model stability, algorithmic fairness, individual-level risk, classification instability, pmstabilityss, pmsampsize, acute kidney injury, diabetic foot ulcer, MIMIC-III database, epistemic uncertainty, decision analysis, health equity

News Source: Ophelia Keating. (November 13, 2026). Fisher Information Decomposition Calculates Sample Size for Fair Clinical Prediction Models. Scienmag.

Tags: acute kidney injuryalgorithmic fairnessbinary outcomesclassification instabilityclinical prediction modelsdecision analysisdiabetic foot ulcerepistemic uncertaintyFisher Information Matrixhealth equityindividual-level risklogistic regressionMIMIC-III databasemodel stabilitypmsampsizepmstabilitysssample size calculation
Share12Tweet7Share2ShareShareShare1

Related Posts

Overweight Exceeds Thinness Among Secondary School Students in Sierra Leone

Overweight Exceeds Thinness Among Secondary School Students in Sierra Leone

November 18, 2026
MRI Features Identify Fracture-Dislocation Patterns in Schatzker Type IV Tibial Plateau Fractures

MRI Features Identify Fracture-Dislocation Patterns in Schatzker Type IV Tibial Plateau Fractures

November 18, 2026

Keloids and Hypertrophic Scars Share Matrix Programmes Despite Divergent Spatial Distributions

November 18, 2026

Shortened Frontal Assessment Battery Effectively Detects Executive Dysfunction in Neurodegenerative Disorders

November 18, 2026

POPULAR NEWS

  • Delayed Corticosteroids Linked to Improved Survival in Deteriorating Septic Shock Phenotype

    Delayed Corticosteroids Linked to Improved Survival in Deteriorating Septic Shock Phenotype

    29 shares
    Share 12 Tweet 7
  • NLRP1 Inhibition Reduces Inflammation in Familial Mediterranean Fever Cellular Models

    29 shares
    Share 12 Tweet 7
  • New Framework Refines Echocardiographic Guidance for Complex Transcatheter Mitral Valve Procedures

    29 shares
    Share 12 Tweet 7
  • Pied Kingfisher Optimization Enhances Dynamic Stability In Simulated Wave Energy Conversion Systems

    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

Lactobacillus Rhamnosus Promotes Abdominal Fat Deposition in Mice Without Changing Body Weight

Overweight Exceeds Thinness Among Secondary School Students in Sierra Leone

MRI Features Identify Fracture-Dislocation Patterns in Schatzker Type IV Tibial Plateau Fractures

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.