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

Objectives: To develop and validate an individualized prediction model for assessing the risk of illness uncertainty in patients with malignant tumors, using a cross-sectional design. Methods: Patients with malignant tumors treated at Peking University People’s Hospital from August 2024 to January 2025 were enrolled. The Mishel Uncertainty in Illness Scale (MUIS) was used to classify patients into high-risk and low-to-moderate risk groups. Patients were divided into a model development set and a time-stratified validation set. The development set was further randomly split into a training set and an internal validation set at a 7:3 ratio. Three machine learning algorithms Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) were employed to screen for common predictors, and a nomogram was constructed based on logistic regression. The model’s performance was evaluated using the Area Under the Receiver Operating Characteristic Curve, calibration curves, and the Hosmer–Lemeshow test. Results: A total of 966 patients were included, with 676 in the development set and 290 in the time-stratified validation set. Seven predictors were ultimately identified for the nomogram: age, education level, diagnosis, depression, anxiety, medical coping modes, and social support. Notably, anxiety was the only variable jointly identified by all three algorithms, while the other six were commonly selected by both Random Forest and XGBoost. The nomogram showed good discrimination in training (AUC = 0.763, 95% CI: 0.717–0.808) and internal validation (AUC = 0.724, 95% CI: 0.645–0.803), with well-calibrated probabilities (p > 0.05). In time-stratified validation, discrimination was acceptable (AUC = 0.663, 95% CI: 0.58–0.739), but calibration revealed risk underestimation for low-to-moderate risk group (p < 0.05). Conclusions: The nomogram shows acceptable internal performance as an exploratory concurrent classifier rather than a genuine predictor, but calibration bias and modest temporal discrimination (AUC = 0.663) indicate it is not ready for clinical use. Further optimization and multi-center prospective validation are required.