A Multi-center Retrospective Observational Study Using AI to Integrate Proteomics, Pathology, and Clinical Big Data to Build a Multi-modal Model for Predicting the Effectiveness of HER-2 Targeted ADC Therapy (T-DXd) in HER-2 Positive and Low-expression Advanced Breast Cancer, With External Validation.
试验速览
- 阶段
- 不适用
- 状态
- 尚未招募
- 入组人数
- 900
- 试验地点
- 1
- 主要终点
- Area Under the Receiver Operating Characteristic Curve (AUC) of the Predictive Model
研究概览
简要总结
This study aims to develop an AI-based predictive tool to help clinicians more accurately determine whether breast cancer patients can benefit from HER-2-targeted antibody-drug conjugate (T-DXd) therapy before treatment. While HER-2-targeted ADC drugs have significantly improved outcomes for patients with HER-2 positive and low-expression advanced breast cancer, there are notable individual differences in efficacy. Currently, there is a lack of precise clinical methods to predict response, which means some patients might receive ineffective treatment and face unnecessary drug side effects and financial burden.
This study is a retrospective multicenter observational study, planning to collect pathological images (including HE staining and HER-2, ER, PR, Ki-67 immunohistochemical staining), proteomics data, and clinical efficacy information from HER-2 positive and low-expression advanced breast cancer patients who have received T-DXd treatment.
The research will be carried out in five phases:
- Build a clinical database for ADC drug therapy, integrating basic patient information, treatment plans, efficacy data, and pathology specimen information from multiple centers.
- Use LC-MS/MS proteomics technology to screen for key protein markers related to T-DXd efficacy and use bioinformatics analysis to identify predictive protein indicators.
- Extract IHC staining features from pathological images and evaluate their correlation with efficacy alongside clinical data.
- Integrate proteomics, pathology, and clinical big data, using AI technologies such as foundational pathology models (like TITAN), biomedical large language models (like BioBERT), and protein large language models (like ESM2-15B). Apply a multiple instance learning strategy to build a multimodal efficacy prediction model, and evaluate the model's performance on the training set using 5-fold cross-validation.
- Establish an internal validation cohort (200 cases) and a multicenter external validation cohort (300 cases). Considering that the external validation group may lack proteomics data, the multimodal model will be fine-tuned and distilled into a simplified predictive model based on standard IHC features (HER-2, ER, PR, Ki-67, plus key protein markers identified from proteomics) and clinical text information, then its performance will be verified in the external cohort.
Ultimately, this research will create an AI tool to support clinical decision-making, promoting personalized treatment for HER-2 positive and low-expression breast cancer and the clinical adoption of AI in healthcare.
详细描述
Ambispective study combining retrospective model development and prospective validation.
研究设计
- 研究类型
- Observational
- 观察模型
- Cohort
- 时间视角
- Other
入排标准
- 性别
- Female
- 接受健康志愿者
- 否
入选标准
- •Retrospective Cohort (Modeling and Validation):
- •Female, 18 years or older;
- •Advanced breast cancer confirmed by pathology (AJCC 8th edition, stage IV);
- •HER2 status known;
- •Received at least 2 cycles of Pyrotinib monotherapy;
- •Complete baseline IHC slides (HER2, ER, PR, Ki-67) and HE-stained slides;
- •Efficacy assessed according to RECIST 1.1, with follow-up data (PFS or ORR).
- •Prospective Cohort (External Validation):
- •Meet criteria 1-3 above;
- •Planning to receive Pyrotinib monotherapy as second-line or later treatment; if HER2-positive, previously received neoadjuvant/adjuvant H(P) therapy, and had metastatic recurrence within 12 months after completing treatment, with post-recurrence anti-HER2 therapy considered second-line treatment.
- •Signed informed consent, agreeing to provide clinical info like imaging and pathology data before and after treatment.
排除标准
- •All cohorts:
- •Baseline IHC or HE slides of poor quality (e.g., faded, folded, or tissue loss >10%);
- •Previous treatment with other HER2-ADC drugs;
- •History of other malignancies (except non-melanoma skin cancer or cases with no recurrence for over 5 years);
- •Participation in other interventional clinical trials at the same time (past trials already completed are fine);
- •Lost to follow-up or missing key clinical data during treatment (e.g., efficacy evaluation, dose adjustment records);
- •Special treatment backgrounds that the AI model cannot analyze (e.g., combined local radiotherapy, severe infections, or other confounding factors).
- •Additional exclusions for the prospective cohort:
- •Pregnant or breastfeeding women;
- •Contraindications to Üher (e.g., history of ILD, left ventricular ejection fraction <50%, etc.);
- •Unable to comply with regular follow-up (e.g., living in a remote area, mental disorders).
结局指标
主要结局
Area Under the Receiver Operating Characteristic Curve (AUC) of the Predictive Model
时间窗: Baseline (at initial diagnosis)
次要结局
未报告次要终点
研究者
Hu Hai
Chief Physician
Zhejiang Cancer Hospital