Abstract

The performance of machine learning (ML) models like the Multilayer Perceptron (MLP) is critical for medical applications such as thyroid disease diagnosis but is highly dependent on hyperparameter selection. Suboptimal configurations can degrade diagnostic accuracy and reliability, making hyperparameter tuning essential. This study evaluates the impact of the Optuna tuning framework on an MLP model for automated thyroid diagnosis. The methodology involved using a public UCI dataset, data preprocessing, training a baseline MLP, and performing extensive hyperparameter optimization with Optuna to maximize the F1-score. This research provides an empirical evaluation of Optuna's effectiveness, showing that classification accuracy increased by 11.6 percentage points (from 0.837 to 0.954) and the macro F1-score rose by 21.3 percentage points (from 0.721 to 0.933) compared to the baseline model, with a particularly noteworthy improvement in recall for the minority classes. Optuna successfully identified a more complex and effective two-hidden-layer architecture. In conclusion, hyperparameter optimization using Optuna is highly effective for enhancing the accuracy and reliability of MLP models in thyroid diagnosis. This approach overcomes the limitations of suboptimal configurations, affirming that meticulous tuning is a critical step in developing robust clinical decision-support systems.

Concepts :

Access to Document

Citations by Year