A Machine Learning Approach for Tracing Tumor Original Sites With Gene Expression Profiles

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Identifying the primary tumor site is critical for effective cancer treatment, especially for carcinomas with unknown origins.
• Patients with unknown primary tumors often receive broad-spectrum chemotherapy, leading to poor prognoses.
• Machine learning (ML) offers promising advancements in clinical practice for complex diagnostic challenges.

Purpose of the Study:

• To classify and predict the origins of tumor samples with uncertain primary sites using ML algorithms.
• To evaluate the efficacy of random forest and Naive Bayesian algorithms in identifying metastatic tumor origins.
• To provide a more precise diagnostic tool for cases of carcinoma of unknown primary.

Main Methods:

• Application of random forest and Naive Bayesian algorithms for tumor sample classification.
• Utilizing precision, recall, and other metrics to assess algorithm performance.
• Employing 10-fold cross-validation to validate classification accuracy.

Main Results:

• Achieved a prediction accuracy of 90.4% for 7,713 tumor samples.
• Demonstrated 80% accuracy in classifying 20 metastatic tumor samples.
• 10-fold cross-validation confirmed a classification accuracy of 91%.

Conclusions:

• Random forest and Naive Bayesian algorithms show high accuracy in predicting tumor origins.
• This ML-based approach can aid in developing precise treatment plans for patients with unknown primary tumors.
• The study highlights the potential of computational methods to improve oncological diagnostics and patient outcomes.