# Hcs 3wt Breast Cancer Diagnosis

> 1. 确认输入参数完整

- Skill: `yakeworld/hcs-3wt-breast-cancer-diagnosis` (Agent Skill, multi-file: 10 files)
- Install (CLI): `npx skillmds@latest add yakeworld/hcs-3wt-breast-cancer-diagnosis`
- Raw SKILL.md: https://api.skillmd.com/api/skills/yakeworld/hcs-3wt-breast-cancer-diagnosis/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- License: MIT
- Author: yakeworld (https://skillmd.com/u/yakeworld)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/yakeworld/hcs-3wt-breast-cancer-diagnosis

---


## Operational Steps
1. 确认输入参数完整
2. 执行核心操作（参考本目录下的 scripts/ 或 references/）
3. 验证输出符合契约
4. 保存结果并报告

## Pitfalls
- 
- 

## Verification
- 
- 
- 
- 
1. 
2. 
3. 
category: research-tools
signature: "hcs-3wt-breast-cancer-diagnosis -> research-tools: HCS-3WT (Hybrid Cascade-Stacking Three-Way Triage) breast cancer diagnostic"
description: HCS-3WT (Hybrid Cascade-Stacking Three-Way Triage) breast cancer diagnostic
author: Synthos
license: MIT
version: 1.1.0

- search_files
    signature: 'data_path: str -> diagnosis_report: dict'
    related_skills:
    - academic-paper-completion
    - adhd-eye-tracking-review
    - arxiv
    - biorxiv
    - blogwatcher
## IO_CONTRACT
- **input**: `request: str, context: dict` — 用户请求描述、上下文信息
- **output**: `result: dict — 技能执行结果（结构因技能而异）`
> 对应原则：P2（机械原子暴露输入输出规范）
# HCS-3WT Breast Cancer Diagnostic AI System
Procedural knowledge for building, evaluating, and optimizing the HCS-3WT (Hybrid Cascade-Stacking Three-Way Triage) system for clinical breast cancer diagnosis.
## System Overview
HCS-3WT is a three-stage cascade classifier that structurally decouples diagnostic pathways:
- **Expert B (Catcher)**: High-recall SVC — Clear Negative if P(malignant) < 0.03
- **Expert A (Refiner)**: Voting Classifier (RF + CatBoost + ExtraTrees) — Clear Positive if P(malignant) > 0.95
- **Expert C (Arbiter)**: Meta-learning stacking classifier — handles Gray Zone cases
Key innovation: The system achieves 70.9% automation rate with 99.3% automation accuracy on WDBC (leakage-free 10x5 CV), with consistent risk-aligned gray zone enrichment across multiple datasets.
## Key Implementation Details
### Feature Engineering
Three novel features capture cytological atypia:
1. `size_shape_interaction`: cell_size_uniformity x cell_shape_uniformity
2. `nuclear_abnormality_score`: bare_nuclei + bland_chromatin + normal_nucleoli
3. `triple_product_score`: clump_thickness x marginal_adhesion x bare_nuclei
### Preprocessing Pipeline
- PowerTransformer (Yeo-Johnson) for distribution normalization
- SelectKBest (ANOVA F-stat) with domain-tailored feature subsets (4 for Expert B, 8 for Expert A)
- Borderline-SMOTE applied only to Expert B and Expert C training sets (NOT Expert A)
### Critical Pitfalls
**SMOTE Single-Class Crash**: Borderline-SMOTE requires at least 2 classes in training data. When running cross-validation on imbalanced or single-class datasets (e.g., Wisconsin Prognostic had 194 samples, 0 malignant), SMOTE will crash.
**Fix**: Before calling SMOTE, check `len(np.unique(y_train)) >= 2`. If only 1 class, skip SMOTE.
**SVC Single-Class Crash**: SVC also crashes when training data has only 1 class.
**Key check pattern**:
```python
n_train_classes = len(np.unique(y_train_cv))
n_test_classes = len(np.unique(y_test_cv))
if n_train_classes < 2 or n_test_classes < 2:
    continue
```
**HCS-3WT All Samples in Gray Zone (0% Automation Rate)**: If all samples end up in the Gray Zone, Expert B is giving all samples probability > low_threshold, and/or Expert A is giving all samples probability < high_threshold. Typically: thresholds are inappropriate for the data distribution. Fix: check probability histograms and adjust thresholds.
**Fixed Thresholds Fail on Low-Separability Datasets**: The default thresholds (theta_low=0.03, theta_high=0.95) calibrated on breast cancer data produce near-zero automation rates on datasets with overlapping feature distributions. PIMA diabetes confirmed this: only 8.58% automation rate with fixed thresholds vs 70-79% on breast cancer datasets. Root cause: SVC/VotingClassifier probability estimates on overlapping clinical data rarely reach 0.95 or 0.03. The architecture correctly defers (0 auto FN) but auto rate collapses.
Fix options:
1. Dataset-specific threshold optimization via cost function grid search
2. Adaptive thresholds via calibration curves (Platt scaling)
3. Relax thresholds (e.g. 0.20/0.80) for low-separability data
4. Document as honesty signal: low auto rate = feature space has poor class separation
Diagnostic: if auto rate drops below 10% on a new dataset, check probability histograms. If >90% of probabilities fall in [0.03, 0.95], the feature space lacks separation for the default thresholds.
**Cross-Validate Meta-Features to Prevent Leakage**: Use cross_val_predict for P_A and P_B features fed to Expert C. Training meta-features on the same data used for testing creates data leakage.
**Long-Running Experiments**: WDBC 10x5 CV (50 folds) takes ~190s. PIMA 10x5 CV takes ~189s. Always use background execution for long experiments.
**SVC probability=True Deprecated in sklearn 1.9+**: Use `CalibratedClassifierCV(SVC(), ensemble=False)` instead. Handle both old and new APIs:
```python
try:
    from sklearn.calibration import CalibratedClassifierCV
    base = SVC(kernel='rbf', class_weight='balanced', random_state=RS)
    catcher = CalibratedClassifierCV(base, ensemble=False)
except ImportError:
    catcher = SVC(kernel='rbf', probability=True, ...)
```
## Cross-Domain Validation (Confirmed Experimental Data, 2026-06-25)
The system has been rigorously evaluated with leakage-free 10x5 stratified CV, all with fixed thresholds theta_low=0.03, theta_high=0.95:
- **WBC Original** (UCI): 699 samples, 9 features. 79.07%+-4.19% automation rate, 99.35%+-0.71% auto accuracy. Gray zone (20.93%) with 1.22x malignant enrichment. HCS-3WT accuracy 0.9657. FN reduction vs best single (SVC): -28.0%.
- **WDBC** (Wisconsin Diagnostic Breast Cancer, sklearn): 569 samples, 30 features. 70.93%+-6.44% automation rate, 99.26%+-0.95% auto accuracy. Gray zone (29.07%) with 1.21x enrichment. HCS-3WT accuracy 0.9475. FN reduction vs best single (SVC): -47.7%.
- **Pima Indians Diabetes** (auxiliary): 768 samples, 8 features. **8.58%+-2.03% auto rate** with fixed thresholds. Shows threshold sensitivity (see pitfall above). Auto accuracy 99.38% with 0 auto FN. Gray zone 1.77x enrichment.
- **Wisconsin Prognostic**: 194 samples, 32 features. Single class (all benign) — must be skipped.
## Files
- `run_hcs3wt.py` — Main WDBC experiment script (10x5 CV, 50 folds)
- `run_hcs3wt_wbc_original.py` — WBC Original experiment script
- `run_hcs3wt_pima.py` — PIMA cross-domain experiment script
- `experiment_results.json` — WDBC results
- `experiment_results_wbc_original.json` — WBC Original results
- `experiment_results_pima.json` — PIMA results
## References
- `references/error-transcript.md` — Common errors and fixes encountered during development
## 验证清单 · VERIFICATION
- [ ] 单类别折检测：交叉验证中 `len(np.unique(y_train)) < 2` 或 `len(np.unique(y_test)) < 2` 的折已 `continue` 跳过，SMOTE 与 SVC 均未崩溃（HCS-001）
- [ ] 低自动化率诊断：自动化率 < 10% 时已检查概率直方图，>90% 概率落在 [0.03, 0.95] 则判定特征空间分离度不足并记录为诚实信号（HCS-002）
- [ ] 元特征防泄漏：Expert C 的 P_A、P_B 特征通过 `cross_val_predict` 生成，未在测试集上训练（HCS-003）
- [ ] 长实验后台化：10x5 交叉验证（50 折，约 190s）使用后台执行，会话未超时（HCS-004）
- [ ] SVC 概率 API 兼容：sklearn 1.9+ 使用 `CalibratedClassifierCV(SVC(), ensemble=False)` 替代已弃用的 `probability=True`，含 ImportError 回退（HCS-005）
- [ ] SMOTE 作用域：Borderline-SMOTE 仅应用于 Expert B 与 Expert C 训练集，Expert A 训练集未做 SMOTE（HCS-006）
- [ ] 低可分数据集阈值：PIMA 等低可分数据集采用数据集特定阈值优化或放宽阈值（如 0.20/0.80），未硬套默认 0.03/0.95（HCS-007）
## 约束规则 · RULES
1. **输入约束**: 参数类型、范围、格式必须校验
2. **输出约束**: 返回值结构、编码、命名必须一致
3. **异常约束**: 错误信息必须包含上下文和恢复建议
4. **安全约束**: 不执行未验证的任意代码，不暴露内部状态
## Golden 集合 · GOLDEN SET
- **Golden Input**: `data_path="wdbc"`（569 样本、30 特征）→ 执行 10x5 分层交叉验证（50 折，后台执行），theta_low=0.03 / theta_high=0.95
- **Golden Output**: 诊断报告含自动化率 ≈ 70.9%±6.4%、自动化准确率 ≈ 99.3%±1.0%、Gray Zone 29.1% 伴 1.21x 恶性富集、HCS-3WT accuracy ≈ 0.9475、vs 最佳单模型 SVC 的 FN 降低 ≈ 47.7%；单类别折已被 `continue` 跳过（未触发 SMOTE/SVC 崩溃）
- **Golden Error**: PIMA 类低可分数据集（自动化率 < 10%，>90% 概率落在 [0.03, 0.95]）→ 不硬套默认阈值，输出含上下文（"特征空间分离度不足"）与恢复建议（"调整阈值至 0.20/0.80 或采用数据集特定阈值优化"）
> Golden 集合是测试的单一真理来源。所有改进必须通过 golden 测试。
> 违反规则的操作视为不安全，必须拒绝或隔离。
> 每项验证必须可执行、可记录、可复现。验证失败时记录原因和修复。
# Hcs 3Wt Breast Cancer Diagnosis---
> (P032 去重: 以下为合并前第二份中的 1 行独有内容, 保留以防丢失)
# Hcs 3Wt Breast Cancer Diagnosis
## Genes (策略基因)
> 紧凑策略表示。条件→策略。需要深度时参考完整文档。
- **[HCS-001]** 训练数据仅包含单一类别 → 跳过 SMOTE 和 SVC 训练步骤，直接跳过该折以避免崩溃
- **[HCS-002]** 自动化率低于 10% 且概率分布集中在 [0.03, 0.95] → 判定特征空间分离度不足，需调整阈值或记录为诚实信号
- **[HCS-003]** 构建 Expert C 的元特征 (P_A, P_B) → 必须使用 cross_val_predict 生成特征以防止数据泄漏
- **[HCS-004]** 执行 10x5 交叉验证等长耗时实验 → 始终使用后台执行以避免会话超时
- **[HCS-005]** 使用 sklearn 1.9+ 版本且需 SVC 概率输出 → 使用 CalibratedClassifierCV 替代已弃用的 probability=True 参数
- **[HCS-006]** 应用 Borderline-SMOTE 预处理 → 仅应用于 Expert B 和 Expert C 的训练集，严禁应用于 Expert A
- **[HCS-007]** 面对低可分性数据集（如 PIMA） → 采用数据集特定的阈值优化或放宽阈值（如 0.20/0.80）而非固定默认阈值