Scholarlink

Objective <p>To develop and validate a machine learning (ML) model based on the modified pancreatitis activity scoring system (mPASS) and imaging features for the early assessment of multiple clinically relevant outcomes in acute pancreatitis (AP).</p> Methods <p>A retrospective cohort of 420 AP patients was analyzed. Of these, 310 patients from the primary center were randomly divided into a training set and an internal validation set at a 3:1 ratio; an additional 110 patients from an external hospital served as the independent testing cohort. The Friedman test was used to compare mPASS scores at various time points. The Jonckheere‑Terpstra test assessed the ordinal trend between mPASS and 2012 Revised Atlanta Classification (RAC). The ML model was developed using the age‑adjusted Charlson Comorbidity Index (aCCI), mPASS and imaging features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and Brier scores for calibration.</p> Results <p>No significant difference was found between mPASS scores at admission and 24&#xa0;h post-admission (<i>P</i> = 1.000), but significant differences were observed at subsequent time points (all <i>P</i> &lt; 0.05). The mPASS score at 72&#xa0;h post-admission demonstrated the best predictive performance for the majority of clinically relevant outcomes, with only a marginal exception. A significant ordinal relationship existed between mPASS and 2012 RAC (<i>P</i> &lt; 0.001). In the testing dataset, AUCs for the ML model ranged from 0.704 to 0.935, exceeding 0.80 for most outcomes. Calibration was acceptable to excellent for most endpoints, though marginal for hospital stay and limited for time to oral refeeding. Overall, the ML model achieved performance generally comparable to that of mPASS and the 2012 RAC, with a significant improvement for complications.</p> Conclusion <p>Our ML model supports simultaneous multi-outcome risk assessment at this single time point. While showing promise in supporting individualized clinical decision-making, its performance warrants further refinement and prospective multicenter validation.</p> Clinical trial number <p>Not applicable (This study is a retrospective analysis and does not involve any prospective health care intervention on human participants).</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

A machine learning model based on mPASS and imaging features for early assessment of clinically relevant outcomes in acute pancreatitis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Di Tao,
  • Xing Hui Li,
  • You Qiang Hu,
  • Ting Ting Liu,
  • Zi Sheng Zhao,
  • Jun Hui Chen,
  • Hou Dong Zuo,
  • Yi Fan Ji,
  • Xiao Ming Zhang

摘要

Objective

To develop and validate a machine learning (ML) model based on the modified pancreatitis activity scoring system (mPASS) and imaging features for the early assessment of multiple clinically relevant outcomes in acute pancreatitis (AP).

Methods

A retrospective cohort of 420 AP patients was analyzed. Of these, 310 patients from the primary center were randomly divided into a training set and an internal validation set at a 3:1 ratio; an additional 110 patients from an external hospital served as the independent testing cohort. The Friedman test was used to compare mPASS scores at various time points. The Jonckheere‑Terpstra test assessed the ordinal trend between mPASS and 2012 Revised Atlanta Classification (RAC). The ML model was developed using the age‑adjusted Charlson Comorbidity Index (aCCI), mPASS and imaging features. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC) and Brier scores for calibration.

Results

No significant difference was found between mPASS scores at admission and 24 h post-admission (P = 1.000), but significant differences were observed at subsequent time points (all P < 0.05). The mPASS score at 72 h post-admission demonstrated the best predictive performance for the majority of clinically relevant outcomes, with only a marginal exception. A significant ordinal relationship existed between mPASS and 2012 RAC (P < 0.001). In the testing dataset, AUCs for the ML model ranged from 0.704 to 0.935, exceeding 0.80 for most outcomes. Calibration was acceptable to excellent for most endpoints, though marginal for hospital stay and limited for time to oral refeeding. Overall, the ML model achieved performance generally comparable to that of mPASS and the 2012 RAC, with a significant improvement for complications.

Conclusion

Our ML model supports simultaneous multi-outcome risk assessment at this single time point. While showing promise in supporting individualized clinical decision-making, its performance warrants further refinement and prospective multicenter validation.

Clinical trial number

Not applicable (This study is a retrospective analysis and does not involve any prospective health care intervention on human participants).