Simpler predictive models provide higher accuracy for ovarian cancer detection.

Area of Science:

• Oncology
• Biomarker Discovery
• Machine Learning in Medicine

Background:

• Ovarian cancer poses a significant threat to women's health, necessitating improved screening methods.
• Established protein biomarkers CA125 and HE4 demonstrate high accuracy in ovarian cancer detection, especially when combined.
• Previous research introduced DELFI-Pro, a logistic regression (LR) model integrating cell-free DNA (cfDNA) features with protein concentrations.

Purpose of the Study:

• To re-evaluate the efficacy of the DELFI-Pro screening model by addressing potential confounding factors in its training data.
• To determine if the cfDNA features in DELFI-Pro offer a significant advantage over protein-only biomarkers for ovarian cancer detection.

Main Methods:

• Analysis of a dataset used in prior DELFI-Pro research, focusing on cfDNA-derived features and protein concentrations (CA125, HE4).
• Identification and removal of training data samples with anomalous chromosomal copy number values that could introduce technical variation.
• Comparative performance evaluation of the refined DELFI-Pro model against protein-only logistic regression classifiers using cross-validation.

Main Results:

• Confounding technical variation was identified within the cfDNA features of the original DELFI-Pro training data.
• Exclusion of 42 outlier cancer samples from the training set revealed that DELFI-Pro did not outperform protein-only models.
• The combined CA125 and HE4 protein model achieved an area under the curve (AUC) of 0.99, indicating high accuracy.

Conclusions:

• The cfDNA features in DELFI-Pro do not provide sufficient added value to justify their inclusion over simpler, protein-based models.
• Simpler machine learning models, like the protein-only classifiers, tend to generalize better to new data.
• Current evidence does not adequately support the complex DELFI-Pro model for ovarian cancer screening.