Main Session
Sep
28
PQA 04 - Breast Cancer, Patient Reported Outcomes/QoL/Survivorship, Functional Radiation Medicine, Hematologic Malignancies, Palliative Care, and International/Global Oncology
2932 - Machine Learning-Based Multimodal Prognostic Model and Exploration of Prognostic Markers in Triple-Negative Breast Cancer
Presenter(s)
Mengli Zhou, MD - ShenZhen Cancer Hospital, ShenZhen, Guangdong
M. Zhou; Center Hospital Chinese Academy of Medical Sciences, Shenzhen Center, Shenzhen, guangdong, China
Purpose/Objective(s): Triple-negative breast cancer (TNBC) is a highly aggressive subtype with strong heterogeneity and poor prognosis. This study aimed to establish a pathomics and deep learning-based prognostic model using machine learning, identify robust prognostic genes, and analyze immune microenvironment differences to improve risk stratification for TNBC.
Materials/Methods:
Whole-slide images (WSIs) of TNBC samples from the TCGA cohort were preprocessed and segmented into 512×512 patches. Deep features were extracted using a pre-trained ResNet-50 model, and quantitative pathological features including intensity, texture, morphology, and distribution were obtained using CellProfiler in accordance with the 2025 TITAN guidelines. Univariate Cox and LASSO regression analyses were applied to screen prognostic pathological features and construct a risk score model; a prognostic nomogram was established and validated. Differentially expressed genes (DEGs) between high- and low-risk groups were identified using DESeq2. Overlapping genes between DEGs and candidate genes were further analyzed by univariate and multivariate Cox regression to identify independent prognostic genes.
Results:
A total of 17 key pathological features were extracted. After univariate Cox and LASSO regression, 2 prognostic pathological features were retained and integrated into the risk score model as protective factors: Total_Intensity_StdIntensity_Hematoxylin (coef = -0.04047386) and Total_Intensity_StdIntensity_OrigGray (coef = -0.239927434). Risk score was calculated as S(Coefficient(i) × Feature(i) expression). Patients in the high-risk group exhibited significantly poorer overall survival than those in the low-risk group. A total of 1896 DEGs were identified (1191 upregulated, 705 downregulated), and 19 overlapping genes were obtained between DEGs and candidate genes. Finally, 8 independent prognostic genes were identified: protective factors (HR < 1): FADS2, LIFR, AIM2, PHGDH, AKR1C1; risk factors (HR > 1): CPEB1, COX7A1, GSTM3.
Conclusion:
We constructed a stable pathomics-based, machine learning-driven prognostic model for TNBC and identified a panel of 8 robust prognostic genes. The model enables effective risk stratification and provides novel potential biomarkers. These findings may support individualized prognostic evaluation and clinical decision-making for TNBC patients.