Machine learning in medicine: a practical introduction.

Area of Science:

• Medical research
• Computational biology
• Biostatistics

Background:

• Machine learning (ML) shows promise in predictive tasks, increasing interest among medical professionals.
• A need exists for ML capacity development in medicine, requiring accessible training and tools.

Purpose of the Study:

• To provide a conceptual introduction to machine learning for medical researchers and clinicians.
• To offer a practical guide for developing and evaluating predictive algorithms using open-source software and public data.

Main Methods:

• Developed three predictive models for cancer diagnosis: regularized General Linear Models (GLMs), Support Vector Machines (SVMs), and Artificial Neural Networks.
• Utilized a public dataset of 683 breast mass nuclei samples, randomly split into evaluation (n=456) and validation (n=227) sets.
• Employed the R statistical programming environment for algorithm development and evaluated performance using accuracy, sensitivity, and specificity.

Main Results:

• Trained algorithms achieved high performance in classifying cell nuclei: accuracy (.94-.96), sensitivity (.97-.99), and specificity (.85-.94).
• Support Vector Machines (SVM) demonstrated maximum accuracy (.96) and area under the curve (.97).
• A voting ensemble marginally improved performance, reaching accuracy (.97), sensitivity (.99), and specificity (.95).

Conclusions:

• Demonstrated the practical application of machine learning for cancer diagnosis using a clear example.
• Highlighted the potential of ML techniques, including ensembles, for improving diagnostic accuracy in medicine.
• Emphasized the broad applicability of these ML principles to other complex medical tasks like image recognition and natural language processing.