Improving calibration of logistic regression models by local estimates

Area of Science:

• Biostatistics
• Machine Learning in Healthcare
• Predictive Modeling

Background:

• Individualized risk assessment tools are crucial in modern healthcare.
• External validation frequently reveals poor calibration in existing risk models.
• Improving model calibration is essential for reliable clinical decision-making.

Purpose of the Study:

• To enhance the calibration of logistic regression (LR) estimates by incorporating local information.
• To develop a novel method for more accurate probability predictions in risk assessment.
• To improve the reliability of individualized risk prediction tools.

Main Methods:

• A clustering algorithm was integrated with a logistic regression (LR) model.
• The combined approach generates probability estimates tailored to individual cases.
• Performance was evaluated against standard LR using calibration (Sum of Absolute Differences - SAD) and discrimination (Area Under the ROC Curve - AUC).

Main Results:

• The new method demonstrated significantly lower SAD values (p < 0.0001) in synthetic datasets, indicating superior calibration.
• Area Under the ROC Curve (AUC) was significantly higher for the new method in one dataset (p < 0.01), suggesting improved discrimination.
• The proposed approach showed a marked improvement in the accuracy of probability estimates.

Conclusions:

• The integrated clustering and logistic regression method effectively improves model calibration.
• This approach offers a promising strategy for enhancing the accuracy of risk prediction tools.
• The findings suggest the method's utility in clinical settings requiring precise risk assessment.