Development and Validation of a Machine Learning-Based Prediction Model for Illness Uncertainty in Patients with...

Area of Science:

• Oncology
• Medical Informatics
• Psychological Oncology

Background:

• Illness uncertainty is a significant concern for patients with malignant tumors.
• Accurate risk assessment models are needed to identify patients requiring support.
• Machine learning offers potential for developing individualized prediction tools.

Purpose of the Study:

• To develop and validate an individualized prediction model for illness uncertainty risk in cancer patients.
• To identify key predictors of illness uncertainty using machine learning algorithms.
• To construct and evaluate a nomogram for clinical application.

Main Methods:

• Cross-sectional study involving 966 patients with malignant tumors.
• Utilized Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) for predictor selection.
• Developed a nomogram based on logistic regression and validated it internally and through time-stratified analysis.

Main Results:

• Seven predictors were identified: age, education, diagnosis, depression, anxiety, coping modes, and social support.
• The nomogram demonstrated good internal discrimination (AUC training=0.763, internal validation=0.724) and calibration.
• Time-stratified validation showed acceptable discrimination (AUC=0.663) but revealed risk underestimation in certain groups.

Conclusions:

• The developed nomogram shows promise as an exploratory classifier but is not yet ready for clinical prediction.
• Calibration bias and modest temporal discrimination highlight limitations for real-world application.
• Further multi-center prospective validation and model optimization are essential.