A comparative study of machine learning and deep learning algorithms to classify cancer types based on microarray...

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Accurate cancer classification is crucial for effective medical treatment and diagnosis.
• Previous studies utilized RNA profiling and Machine Learning (ML) for tumor classification across various cancer types.
• The 11_tumor database, comprising diverse human tumor data, lacks comparative algorithm studies.

Purpose of the Study:

• To compare the performance of classical ML and Deep Learning (DL) algorithms for cancer classification.
• To evaluate tumor identification accuracy using the 11_tumor database.
• To assess the impact of algorithm tuning on classification accuracy.

Main Methods:

• Utilized the 11_tumor database containing gene expression (microarray) data from multiple human cancer types.
• Applied and compared various classical ML algorithms (e.g., Logistic Regression) and DL algorithms (e.g., Convolutional Neural Networks).
• Employed k-fold cross-validation to evaluate algorithm performance and accuracy.

Main Results:

• Achieved tumor identification accuracies ranging from 90.6% (Logistic Regression) to 94.43% (Convolutional Neural Networks).
• Demonstrated that Convolutional Neural Networks (CNNs) provided superior classification performance compared to classical ML algorithms.
• Showcased the impact of algorithm tuning on classification accuracy, with varying degrees of improvement observed.

Conclusions:

• Gene expression data combined with ML/DL algorithms offers an efficient and accurate method for tumor classification.
• Deep Learning, particularly CNNs, shows significant promise for accurate tumor type prediction in complex, multi-cancer scenarios.
• This comparative study provides valuable insights for selecting optimal algorithms in cancer bioinformatics research.