A multidimensional risk prediction framework for malignant intestinal obstruction based on machine learning:...

Area of Science:

• Oncology
• Medical Imaging
• Machine Learning

Background:

• Malignant intestinal obstruction (MIO) is a critical complication in advanced cancer patients.
• Current static assessment models fail to capture the dynamic nature of MIO, hindering effective clinical management.
• There is a need for advanced tools to dynamically assess and predict MIO risk.

Purpose of the Study:

• To develop and validate a multimodal machine learning framework for dynamic risk prediction of MIO.
• To identify core predictive features, including the tumor enhancement ratio (TER).
• To evaluate the clinical utility and impact of the developed model.

Main Methods:

• Utilized a multimodal machine learning approach incorporating dynamic imaging features.
• Employed Lasso-Boruta screening for core feature extraction, including TER.
• Developed a risk prediction model using an XGBoost-Random Survival Forest (RSF) cascade structure.
• Validated the model on internal and external patient cohorts (320 MIO patients total).

Main Results:

• Achieved an Area Under the Curve (AUC) of 0.84 ± 0.03 and a Brier score of 0.19 in internal validation.
• Demonstrated robust performance on external validation cohorts.
• Clinical implementation led to a 41% reduction in mechanical ventilation duration and decreased antibiotic resistance rates from 37% to 14%.

Conclusions:

• The developed machine learning framework provides dynamic and interpretable decision support for MIO.
• This approach enhances precise diagnosis and treatment strategies for MIO.
• The model shows significant potential for improving clinical outcomes in MIO patients.