A cost-minimizing diagnostic methodology for discrimination between neoplastic and non-neoplastic brain lesions:...

Area of Science:

• Medical imaging analysis
• Biostatistics
• Machine learning in healthcare

Background:

• Magnetic resonance spectroscopy (MRS) is used to detect brain neoplasms.
• Logistic regression models are common for diagnostic prediction.
• Conventional coefficient estimation methods may have limitations in accuracy and cost.

Purpose of the Study:

• To enhance logistic regression model performance for brain neoplasia prediction using MRS data.
• To introduce and evaluate a novel cost-minimizing (C-min) approach for logistic regression coefficient estimation.

Main Methods:

• Developed a C-min approach using a genetic algorithm to estimate logistic regression coefficients and decision rules.
• Constructed logistic regression models with conventional and C-min estimates using MRS data from patients with suspected brain neoplasms.
• Compared sensitivity and specificity outcomes at various probability thresholds.

Main Results:

• The C-min approach outperformed the conventional approach in 14 out of 18 trials.
• C-min demonstrated reduced false negatives and false positives, or a reduction in one error type with equal or fewer of the other.
• The C-min method consistently resulted in lower costs.

Conclusions:

• The C-min approach shows promise for reducing errors and costs in diagnostic and treatment processes.
• Further validation on larger, diverse datasets and determination of statistical performance characteristics are necessary before clinical implementation.