Model Retirement Creates Reproducibility Risk in Biomedical AI Publications
研究针对生物医学AI论文因模型退役导致的可重复性风险问题，通过分析2022-2026年间的相关文献及模型生命周期数据提出解决方案。
0 citationsRead paper
Selected work
Representative Papers
Doubly robust estimation of while-alive estimands in individually-randomized and cluster-randomized trials
This study addresses the challenges of bias from death-truncated recurrent events and the lack of doubly robust estimation in chronic disease trials. We propose a doubly robust estimator for the exposure-weighted while-alive rate based on a local Nelson-Aalen representation. Applicable to both individual and cluster-randomized trials, this method integrates augmented estimating equations with clustered influence function inference, establishing component-wise double robustness and asymptotic normality. Simulation studies and re-analyses of two clinical trials demonstrate that the proposed approach effectively eliminates survival truncation bias, enabling unbiased causal inference regarding event burden during survival. This work fills a critical methodological gap in handling informative censoring due to death in recurrent event analysis.
0 citationsRead paper
Clinical Feasibility of Low-Magnification Fluorescence Imaging for Breast Cancer Margin Detection Using Texture Analysis and Deep Learning
This study addresses the clinical need for efficient and accurate intraoperative assessment of breast cancer margins, where high-magnification imaging is often limited by small field-of-view and slow acquisition. For the first time, it systematically compares 4× and 10× MUSE fluorescence imaging performance, integrating Local Binary Pattern (LBP) texture analysis with Vision Transformer (ViT)-based deep learning for tissue classification. Results demonstrate that 4× imaging achieves 96.30% sensitivity, 100% specificity, and 98.18% accuracy under the ViT model, while LBP yields consistent 96.67% accuracy across both magnifications. These findings indicate that low-magnification imaging can deliver diagnostic accuracy comparable to high-magnification approaches while substantially improving field-of-view coverage and imaging speed, thereby enhancing practicality in intraoperative settings.
0 citationsRead paper
Improving inverse probability of censoring weighting for win statistics with composite survival outcomes
This study addresses the inefficiency of existing inverse probability of censoring weighting (IPCW) methods for analyzing composite time-to-event endpoints with censoring, which arises from discarding patient pairs whose ordering cannot be definitively determined—a problem exacerbated under high censoring rates or long follow-up windows. The authors propose a novel weighted estimator that replaces hard win/loss assignments with conditional tie probabilities, thereby incorporating partially observable pairs as fractional contributions to recover lost information and enhance estimation efficiency. Built upon the asymptotic theory of two-sample U-statistics, the method integrates IPCW with nuisance-parameter-adjusted inference and provides closed-form sandwich variance estimators for the win ratio, net benefit, and odds ratio. Simulations demonstrate substantial efficiency gains across low to high censoring scenarios, and the approach is successfully applied in a reanalysis of a completed randomized clinical trial.
0 citationsRead paper
Application of Propensity Score Models and Causal Estimators in Observational Studies under Model Misspecification
This study systematically evaluates the robustness of response surface modeling (RSM), inverse probability weighting (IPW), and augmented inverse probability weighting (AIPW) under various misspecification scenarios of propensity score and outcome models in observational studies. Leveraging multiple methods—including logistic regression, random forests, support vector machines, and linear discriminant analysis—to estimate propensity scores, the authors assess performance through extensive simulations and real-world applications to the ACTG175 and ADNI datasets. Findings indicate that AIPW demonstrates consistent robustness across most settings, benefiting from its double-robustness property; IPW proves highly sensitive to propensity score misspecification, while RSM performs well only when the outcome model is correctly specified. The results underscore that integrating flexible machine learning techniques within a doubly robust framework substantially enhances the reliability of causal effect estimation.
0 citationsRead paper
Recent publications
Latest Papers
Model Retirement Creates Reproducibility Risk in Biomedical AI Publications
研究针对生物医学AI论文因模型退役导致的可重复性风险问题，通过分析2022-2026年间的相关文献及模型生命周期数据提出解决方案。
0 citationsRead paper
Doubly robust estimation of while-alive estimands in individually-randomized and cluster-randomized trials
This study addresses the challenges of bias from death-truncated recurrent events and the lack of doubly robust estimation in chronic disease trials. We propose a doubly robust estimator for the exposure-weighted while-alive rate based on a local Nelson-Aalen representation. Applicable to both individual and cluster-randomized trials, this method integrates augmented estimating equations with clustered influence function inference, establishing component-wise double robustness and asymptotic normality. Simulation studies and re-analyses of two clinical trials demonstrate that the proposed approach effectively eliminates survival truncation bias, enabling unbiased causal inference regarding event burden during survival. This work fills a critical methodological gap in handling informative censoring due to death in recurrent event analysis.
0 citationsRead paper
Clinical Feasibility of Low-Magnification Fluorescence Imaging for Breast Cancer Margin Detection Using Texture Analysis and Deep Learning
This study addresses the clinical need for efficient and accurate intraoperative assessment of breast cancer margins, where high-magnification imaging is often limited by small field-of-view and slow acquisition. For the first time, it systematically compares 4× and 10× MUSE fluorescence imaging performance, integrating Local Binary Pattern (LBP) texture analysis with Vision Transformer (ViT)-based deep learning for tissue classification. Results demonstrate that 4× imaging achieves 96.30% sensitivity, 100% specificity, and 98.18% accuracy under the ViT model, while LBP yields consistent 96.67% accuracy across both magnifications. These findings indicate that low-magnification imaging can deliver diagnostic accuracy comparable to high-magnification approaches while substantially improving field-of-view coverage and imaging speed, thereby enhancing practicality in intraoperative settings.
0 citationsRead paper
Improving inverse probability of censoring weighting for win statistics with composite survival outcomes
This study addresses the inefficiency of existing inverse probability of censoring weighting (IPCW) methods for analyzing composite time-to-event endpoints with censoring, which arises from discarding patient pairs whose ordering cannot be definitively determined—a problem exacerbated under high censoring rates or long follow-up windows. The authors propose a novel weighted estimator that replaces hard win/loss assignments with conditional tie probabilities, thereby incorporating partially observable pairs as fractional contributions to recover lost information and enhance estimation efficiency. Built upon the asymptotic theory of two-sample U-statistics, the method integrates IPCW with nuisance-parameter-adjusted inference and provides closed-form sandwich variance estimators for the win ratio, net benefit, and odds ratio. Simulations demonstrate substantial efficiency gains across low to high censoring scenarios, and the approach is successfully applied in a reanalysis of a completed randomized clinical trial.
0 citationsRead paper
Application of Propensity Score Models and Causal Estimators in Observational Studies under Model Misspecification
This study systematically evaluates the robustness of response surface modeling (RSM), inverse probability weighting (IPW), and augmented inverse probability weighting (AIPW) under various misspecification scenarios of propensity score and outcome models in observational studies. Leveraging multiple methods—including logistic regression, random forests, support vector machines, and linear discriminant analysis—to estimate propensity scores, the authors assess performance through extensive simulations and real-world applications to the ACTG175 and ADNI datasets. Findings indicate that AIPW demonstrates consistent robustness across most settings, benefiting from its double-robustness property; IPW proves highly sensitive to propensity score misspecification, while RSM performs well only when the outcome model is correctly specified. The results underscore that integrating flexible machine learning techniques within a doubly robust framework substantially enhances the reliability of causal effect estimation.
0 citationsRead paper