Diagnosing thyroid disorders: Comparison of logistic regression and neural network models

Area of Science:

• Medical informatics
• Computational diagnostics
• Endocrinology

Background:

• Thyroid disorders, including hyperthyroidism and hypothyroidism, are prevalent.
• Accurate diagnosis is crucial for effective patient management.
• Existing diagnostic methods require evaluation for improved predictive capabilities.

Purpose of the Study:

• To diagnose hyperthyroidism and hypothyroidism using multinomial logistic regression and neural network models.
• To compare the diagnostic predictive ability of clinical symptoms versus laboratory tests.
• To evaluate the performance of different predictive models for thyroid dysfunction.

Main Methods:

• Data from 310 patients with euthyroid, hyperthyroid, or hypothyroid states were collected.
• Demographics, symptoms, and laboratory test results were analyzed.
• Multinomial logistic regression and neural network models were applied and compared using accuracy and AUC metrics.

Main Results:

• Neural network models outperformed multinomial logistic regression in diagnosing thyroid disorders.
• The highest predictive accuracy was achieved when all variables were included (91.4% for logistic regression, 96.3% for neural networks).
• Models based on symptomatic variables demonstrated superior predictive performance compared to those using laboratory variables.

Conclusions:

• Both logistic regression and neural network models show high diagnostic accuracy for thyroid disorders, with neural networks being superior.
• Nonparametric predictive techniques like neural networks offer promising avenues for enhancing diagnostic accuracy in medical research.
• Clinical symptom data appears more predictive than laboratory test data for these models.