Scholarlink

Background <p>Acute kidney injury (AKI) is a common and serious complication among hospitalized patients, and early risk stratification remains challenging due to heterogeneous clinical data and limited interpretability of existing prediction models. Foundation models designed for tabular data may enable accurate prediction while preserving interpretability, but their application in early AKI risk assessment has not been fully explored.</p> Objective <p>To develop and externally validate an interpretable early prediction model for AKI using the Tabular Prior-data Fitted Network (TabPFN) based on routinely available admission data.</p> Methods <p>In this retrospective cohort study, predictors were restricted to clinical variables recorded within the 24&#xa0;h prior to hospital admission, and this temporal definition was applied consistently across both the internal cohort and the external MIMIC-IV validation dataset. Time zero was defined as the admission time. Baseline serum creatinine (SCr) was defined as the first creatinine measurement at admission. The primary outcome was in-hospital AKI, defined according to KDIGO SCr criteria as a subsequent rise in creatinine relative to this admission baseline at any time during the index hospitalization. A total of 44,324 patients were included in the development cohort. TabPFN was trained on a stratified subsample and evaluated on a held-out internal test set, and benchmarked against seven conventional machine-learning models. Missing data were handled using multivariate imputation, and model interpretation was performed using SHAP-based attribution analyses. External validation was conducted in the MIMIC-IV database following predefined inclusion criteria and feature harmonization.</p> Results <p>In the internal test set, TabPFN achieved an AUROC of 0.953, outperforming comparator models. External validation demonstrated robust discrimination with an AUROC of 0.859. Calibration analyses indicated good agreement between predicted and observed risks. Attribution analyses identified baseline renal function and acute illness markers as major contributors to model-attributed AKI risk, with heterogeneous association patterns across patient subgroups.</p> Conclusions <p>Using routinely available pre-admission data, TabPFN enabled accurate early prediction of in-hospital AKI and provided interpretable risk attribution patterns. These findings suggest potential utility for early risk stratification; however, results are observational and hypothesis-generating, and prospective validation is required before clinical deployment.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Development and external validation of an interpretable early prediction model for acute kidney injury using TabPFN and routine admission data: a retrospective cohort study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Zheng Xu,
  • Chenyu Li,
  • Chen Guan,
  • Lingyu Xu,
  • Xinyuan Wang,
  • Siqi Jiang,
  • Ningxin Zhang,
  • Minghao Gu,
  • Yanlu Xin,
  • Yan Xu

摘要

Background

Acute kidney injury (AKI) is a common and serious complication among hospitalized patients, and early risk stratification remains challenging due to heterogeneous clinical data and limited interpretability of existing prediction models. Foundation models designed for tabular data may enable accurate prediction while preserving interpretability, but their application in early AKI risk assessment has not been fully explored.

Objective

To develop and externally validate an interpretable early prediction model for AKI using the Tabular Prior-data Fitted Network (TabPFN) based on routinely available admission data.

Methods

In this retrospective cohort study, predictors were restricted to clinical variables recorded within the 24 h prior to hospital admission, and this temporal definition was applied consistently across both the internal cohort and the external MIMIC-IV validation dataset. Time zero was defined as the admission time. Baseline serum creatinine (SCr) was defined as the first creatinine measurement at admission. The primary outcome was in-hospital AKI, defined according to KDIGO SCr criteria as a subsequent rise in creatinine relative to this admission baseline at any time during the index hospitalization. A total of 44,324 patients were included in the development cohort. TabPFN was trained on a stratified subsample and evaluated on a held-out internal test set, and benchmarked against seven conventional machine-learning models. Missing data were handled using multivariate imputation, and model interpretation was performed using SHAP-based attribution analyses. External validation was conducted in the MIMIC-IV database following predefined inclusion criteria and feature harmonization.

Results

In the internal test set, TabPFN achieved an AUROC of 0.953, outperforming comparator models. External validation demonstrated robust discrimination with an AUROC of 0.859. Calibration analyses indicated good agreement between predicted and observed risks. Attribution analyses identified baseline renal function and acute illness markers as major contributors to model-attributed AKI risk, with heterogeneous association patterns across patient subgroups.

Conclusions

Using routinely available pre-admission data, TabPFN enabled accurate early prediction of in-hospital AKI and provided interpretable risk attribution patterns. These findings suggest potential utility for early risk stratification; however, results are observational and hypothesis-generating, and prospective validation is required before clinical deployment.