Predicting HER2 overexpression in prostate cancer using machine learning: implications for personalized therapy

TL;DR

This study uses machine learning and MRI data to predict HER2 overexpression in prostate cancer, which could help guide personalized treatment.

Contribution

The novel contribution is combining radiomics features from MRI with clinical data to predict HER2 status in prostate cancer patients.

Findings

The combined model achieved an AUC of 0.841 in predicting HER2 overexpression.

The model showed balanced performance with 78% accuracy, 77% sensitivity, and 79% specificity.

The model outperformed clinical-only and radiomics-only models in discrimination and clinical utility.

Abstract

Human Epidermal Growth Factor Receptor 2 (HER2), a component of the epidermal growth factor receptor family, is thought to be related to advanced prostate cancer (PCa) when overexpressed. Currently, most research on HER2 is limited to molecular pathology, with relatively few studies focused on imaging aspects.

To develop a predictive model by extracting high-throughput radiomics features from magnetic resonance imaging and combining them with clinical characteristics for predicting HER2 overexpression.

A total of 201 patients who underwent radical prostatectomy and HER2 immunohistochemistry were retrospectively enrolled. These patients were randomly divided into a training set (n=160) and a test set (n=41). Multimodal radiomics features extracted from T2-weighted imaging (T2WI) and apparent diffusion coefficient maps (ADC) were selected using Mann-Whitney U test and least absolute shrinkage and selection operator (LASSO) with ten-fold cross-validation. Predictive models were developed and evaluated based on discrimination and clinical utility.

The combined model integrating ISUP Grade, PSA and Radscore achieved an area under the curve (AUC) of 0.841 (95% CI: 0.697-0.955) in the test set, significantly outperforming the clinical model (AUC = 0.580; p = 0.02, DeLong test) and demonstrating a modest improvement over the Radscore model (AUC = 0.838 (0.693-0.951). Evaluation results showed consistent discriminatory power: 0.78 accuracy, 0.77 sensitivity, and 0.79 specificity, indicating well-balanced performance between positive and negative classes. Decision curve analysis and Waterfall plot demonstrated strong clinical applicability.

The combined model effectively predicts HER2 overexpression in prostate cancer, with potential to inform more personalized treatment strategies for HER2-overexpressing PCa patients.

Figures