Machine learning on thyroid disease: a review.

Area of Science:

• Medical Informatics
• Computational Biology
• Oncology

Background:

• Early diagnosis of thyroid disease is crucial for effective treatment and improved patient outcomes.
• Machine learning (ML) offers promising tools for analyzing complex medical data to identify disease markers.
• Recent advancements necessitate a review of ML applications in thyroid disease detection.

Purpose of the Study:

• To review recent progress in applying machine learning for the early diagnosis of thyroid disease.
• To identify appropriate ML methods for different data types used in thyroid disease diagnosis.
• To highlight key variables and features important for early thyroid disease detection.

Main Methods:

• Systematic review of machine learning techniques applied to thyroid disease diagnosis.
• Analysis of ML model performance across various data types (numeric, genomic, radiomic, ultrasound).
• Identification and categorization of significant diagnostic attributes, including clinical, genomic, and imaging features.

Main Results:

• Random forest and gradient boosting are effective for numeric data; random forest excels with genomic, radiomic, and ultrasound data.
• ML model performance metrics show high accuracy (64.3-99.5%), sensitivity (66.8-90.1%), specificity (61.8-85.5%), and AUC (64.0-96.9%).
• Key diagnostic variables include clinical factors, RNA features (e.g., ADD3-AS1, MIR100HG), gut microbiota (e.g., veillonella), and ultrasound characteristics (e.g., micro-calcification).

Conclusions:

• Machine learning, particularly random forest, demonstrates significant potential for the early diagnosis of various thyroid conditions.
• The selection of ML algorithms should be tailored to the specific data type being analyzed.
• Identifying critical clinical, molecular, and imaging biomarkers is essential for enhancing the accuracy of ML-based thyroid disease detection.