1. Elucidate Key Biomarkers Differentiating Classes: Achieve a comprehensive understanding of which biomarkers exhibit the most significant differences between healthy individuals and ovarian cancer patients. This will involve statistical analysis, such as the t-test, to identify biomarkers with the greatest discriminatory power based on their concentrations in blood samples.
2. Develop and Compare Classification Models: Construct a variety of classification models, including k-Nearest Neighbors (kNN), Support Vector Machines (SVM), ensemble models, and neural networks, to classify blood samples into healthy and ovarian cancer categories. Evaluate and compare their performance metrics (e.g., accuracy, sensitivity, specificity) to determine the most effective approach.
3. Identify High-Impact Biomarkers: Determine which biomarkers have the greatest influence on distinguishing between the two classes through feature importance analysis. This will provide insights into the biological relevance of specific markers and guide future diagnostic development.
4. Assess Model Stability: Validate the stability and robustness of the optimal classification model using appropriate methods, such as cross-validation or bootstrapping, to ensure consistent performance across diverse datasets and conditions.
5. Develop Regression Models for Predictive Indexing: Create regression models to link biomarker concentrations to a probability-based predictive index, analogous to the Risk of Ovarian Malignancy Algorithm (ROMA). This will enable quantitative risk assessment for ovarian cancer based on clinical biomarker data.
6. Integrate Dual Detection Approaches: Devise a method to combine the machine learning-enhanced electronic nose approach (based on volatile organic compounds) with the clinical biomarker-based detection method. This unified strategy aims to enhance diagnostic accuracy and early-stage prediction by leveraging complementary data sources.

Jens Eriksson, Linköping University
Donatella Puglisi, Linköping University