Automated data preparation for in vivo tumor characterization with machine learning

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Data science for oncology

Background:

• Effective data preparation (DP) is crucial for building accurate machine learning (ML) prediction models in cancer research.
• Traditional DP methods often require manual selection, which can be suboptimal for complex clinical datasets.
• The need for automated and optimized DP strategies is paramount for advancing predictive oncology.

Purpose of the Study:

• To propose and validate a novel machine learning-driven data preparation (MLDP) framework for optimizing DP in cancer cohorts.
• To enhance the performance of ML prediction models by automating the selection of appropriate DP algorithms.
• To compare the efficacy of MLDP against manual DP and standard ML approaches in diverse cancer datasets.

Main Methods:

• Developed an MLDP pipeline integrating established DP methods, utilizing evolutionary algorithms and hyperparameter optimization for automated algorithm selection.
• Validated the MLDP approach on single-center glioma and prostate cancer cohorts using Monte Carlo cross-validation (100-fold, 80-20% split).
• Assessed MLDP performance on a dual-center diffuse large B-cell lymphoma (DLBCL) cohort for independent validation, employing five ML classifiers.

Main Results:

• MLDP significantly improved predictive model performance, with 16 out of 20 models showing increased Area Under the Curve (AUC).
• Random Forest (RF) and Support Vector Machine (SVM) models achieved the highest AUC gains (+0.16 and +0.13, respectively) for glioma survival prediction.
• Optimal DP pipelines varied by cohort complexity, with single-center cohorts featuring extensive steps (e.g., outlier detection, feature selection, SMOTE), while dual-center cohorts required fewer steps.

Conclusions:

• MLDP is a highly effective strategy for optimizing data preparation prior to ML model building in cancer cohorts.
• The ML-driven approach yields superior prediction models compared to manual or standard DP methods in both single and multi-center settings.
• This study underscores the importance of intelligent, automated data preparation for advancing predictive accuracy in clinical oncology.