这次我们来看一个医疗AI领域的实用项目——基于SHAP解释的放射组学模型专门用于预测全脑放疗后的患者生存获益情况。这个项目结合了影像分析、机器学习模型和可解释性AI技术让临床医生能够更直观地理解模型预测的依据。对于需要进行全脑放疗的脑转移患者来说准确预测治疗效果和生存期至关重要。传统的预测方法依赖临床经验而这个放射组学模型通过提取CT或MRI影像中的定量特征构建预测模型并用SHAP值来解释每个特征对预测结果的贡献度。这样既提高了预测的准确性又让模型决策过程变得透明可解释。本文将重点介绍这个项目的核心能力、适用场景、环境部署方法、模型训练流程、SHAP解释分析以及实际应用中的注意事项。无论你是医学影像研究者、临床医生还是对可解释AI感兴趣的开发者都能从中获得实用的技术方案。1. 核心能力速览能力项说明项目类型医疗影像分析 可解释机器学习主要功能从放射影像提取组学特征预测全脑放疗后生存获益提供SHAP可解释性分析输入数据CT或MRI脑部影像DICOM格式输出结果生存获益预测概率 特征重要性SHAP图模型架构放射组学特征提取 机器学习分类器如随机森林、XGBoost可解释性SHAP值分析可视化特征贡献度硬件要求CPU/GPU均可GPU加速特征提取过程内存需求8GB以上RAM处理大型影像数据集需要更多内存部署方式Python脚本、Jupyter Notebook、Web服务接口适合场景临床研究、疗效预测、医学影像分析项目2. 适用场景与使用边界这个放射组学模型主要适用于脑转移患者全脑放疗后的生存获益预测。在临床实践中医生需要判断哪些患者可能从全脑放疗中获益以及预期的生存期改善程度。传统方法主要依赖临床病理特征而放射组学提供了从影像中提取定量特征的新途径。适用场景包括放疗科医生评估治疗方案效果医学研究人员进行预后因素分析医院影像科建立智能预测系统临床决策支持工具开发使用边界需要特别注意模型预测结果仅供参考不能替代临床医生专业判断需要高质量的DICOM影像数据图像质量直接影响特征提取效果模型训练需要足够多的标注数据小样本容易过拟合SHAP解释基于特征重要性但不能证明因果关系涉及患者隐私数据必须进行脱敏处理符合医疗数据安全规范在实际应用中建议将模型作为辅助工具结合临床经验共同决策。对于关键医疗决策必须有多重验证机制。3. 环境准备与前置条件要运行这个放射组学项目需要准备以下环境操作系统要求Windows 10/11, macOS 10.14, Ubuntu 18.04 等主流系统建议使用Linux系统获得更好的性能表现Python环境Python 3.8-3.10版本3.11可能存在兼容性问题推荐使用conda或venv创建虚拟环境需要安装的关键包pyradiomics, shap, scikit-learn, pandas, numpy, matplotlib医学影像处理依赖SimpleITK或ITK用于DICOM文件读取PyDICOM用于DICOM元数据处理OpenCV或PIL用于图像预处理机器学习框架scikit-learn用于传统机器学习模型可选PyTorch或TensorFlow用于深度学习扩展硬件配置建议CPU: 4核以上推荐8核以上内存: 8GB最小16GB推荐处理大批量影像需要32GB存储: 至少50GB可用空间用于存储影像数据和中间结果GPU: 非必需但可加速大型影像处理CUDA 11.0数据准备收集脑部CT或MRI的DICOM数据准备对应的临床随访数据生存时间、治疗反应等数据需要经过伦理审批和脱敏处理4. 安装部署与启动方式4.1 创建Python虚拟环境# 使用conda创建环境 conda create -n radiomics_shap python3.9 conda activate radiomics_shap # 或者使用venv python -m venv radiomics_shap source radiomics_shap/bin/activate # Linux/macOS radiomics_shap\Scripts\activate # Windows4.2 安装核心依赖包# 安装放射组学核心包 pip install pyradiomics pip install SimpleITK pip install pydicom # 安装机器学习和可解释性包 pip install scikit-learn pip install xgboost pip install shap pip install pandas numpy matplotlib seaborn # 安装Jupyter用于交互式分析可选 pip install jupyter notebook4.3 项目结构准备创建以下目录结构来组织代码和数据radiomics_project/ ├── data/ │ ├── dicom/ # 原始DICOM影像 │ ├── masks/ # 感兴趣区域掩膜 │ └── clinical.csv # 临床数据 ├── features/ # 提取的特征文件 ├── models/ # 训练好的模型 ├── results/ # 分析结果和图表 └── scripts/ ├── extract_features.py # 特征提取脚本 ├── train_model.py # 模型训练脚本 ├── shap_analysis.py # SHAP分析脚本 └── utils.py # 工具函数4.4 基础验证脚本创建一个简单的验证脚本来测试环境是否正常# test_environment.py import sys import pkg_resources required_packages { pyradiomics: 5.0.0, shap: 0.41.0, scikit-learn: 1.2.0, pandas: 1.5.0 } print(检查环境依赖...) for package, version in required_packages.items(): try: installed_version pkg_resources.get_distribution(package).version print(f✓ {package}: {installed_version}) except pkg_resources.DistributionNotFound: print(f✗ {package} 未安装) print(\n测试基本功能...) try: import pyradiomics import shap from sklearn.ensemble import RandomForestClassifier print(✓ 核心包导入成功) except ImportError as e: print(f✗ 导入失败: {e})运行验证脚本python test_environment.py5. 特征提取与数据预处理5.1 DICOM数据读取与标准化放射组学分析的第一步是正确处理DICOM格式的医学影像# dicom_processor.py import pydicom import numpy as np import SimpleITK as sitk from radiomics import featureextractor class DICOMProcessor: def __init__(self, dicom_dir): self.dicom_dir dicom_dir self.extractor featureextractor.RadiomicsFeatureExtractor() def load_dicom_series(self): 加载DICOM序列并转换为SimpleITK图像 reader sitk.ImageSeriesReader() dicom_names reader.GetGDCMSeriesFileNames(self.dicom_dir) reader.SetFileNames(dicom_names) image reader.Execute() return image def create_mask(self, image, methodthreshold): 创建感兴趣区域掩膜 if method threshold: # 基于阈值的简单掩膜生成 mask sitk.BinaryThreshold(image, lowerThreshold-1000, upperThreshold300, insideValue1, outsideValue0) return mask def extract_radiomics_features(self, image, mask): 提取放射组学特征 try: features self.extractor.execute(image, mask) return features except Exception as e: print(f特征提取失败: {e}) return None5.2 放射组学特征配置配置特征提取参数确保提取的特征具有临床意义# feature_config.py import json # 放射组学特征提取配置 feature_settings { imageType: { Original: {}, Wavelet: {} }, featureClass: { firstorder: [], glcm: [Autocorrelation, JointAverage], glrlm: [RunLengthNonuniformity], glszm: [ZonePercentage], gldm: [DependenceVariance], ngtdm: [Coarseness], shape: [Maximum3DDiameter] }, setting: { binWidth: 25, resampledPixelSpacing: None, interpolator: sitkBSpline, label: 1 } } # 保存配置 with open(radiomics_settings.json, w) as f: json.dump(feature_settings, f, indent2)5.3 批量特征提取流程对于大规模数据集需要实现批量处理# batch_feature_extraction.py import os import pandas as pd from dicom_processor import DICOMProcessor def batch_extract_features(data_root, output_fileradiomics_features.csv): 批量提取所有患者的放射组学特征 patients [d for d in os.listdir(data_root) if os.path.isdir(os.path.join(data_root, d))] all_features [] for patient_id in patients: patient_path os.path.join(data_root, patient_id, dicom) mask_path os.path.join(data_root, patient_id, mask.nii.gz) if not os.path.exists(patient_path): continue try: processor DICOMProcessor(patient_path) image processor.load_dicom_series() mask sitk.ReadImage(mask_path) features processor.extract_radiomics_features(image, mask) if features: features[patient_id] patient_id all_features.append(features) except Exception as e: print(f患者 {patient_id} 处理失败: {e}) # 转换为DataFrame并保存 df_features pd.DataFrame(all_features) df_features.to_csv(output_file, indexFalse) return df_features6. 模型训练与验证6.1 数据准备与特征工程在训练模型前需要进行仔细的数据预处理# data_preparation.py import pandas as pd import numpy as np from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler from sklearn.impute import SimpleImputer class SurvivalDataPreprocessor: def __init__(self, features_file, clinical_file): self.features_df pd.read_csv(features_file) self.clinical_df pd.read_csv(clinical_file) def merge_datasets(self): 合并放射组学特征和临床数据 merged_df pd.merge(self.features_df, self.clinical_df, onpatient_id, howinner) return merged_df def preprocess_features(self, df): 特征预处理缺失值处理、标准化 # 分离特征和目标变量 X df.drop([patient_id, survival_status, survival_days], axis1) y (df[survival_days] 180).astype(int) # 6个月生存获益 # 处理缺失值 imputer SimpleImputer(strategymedian) X_imputed imputer.fit_transform(X) # 标准化特征 scaler StandardScaler() X_scaled scaler.fit_transform(X_imputed) return X_scaled, y, scaler, imputer def prepare_train_test(self, test_size0.2, random_state42): 准备训练测试集 merged_df self.merge_datasets() X, y, scaler, imputer self.preprocess_features(merged_df) X_train, X_test, y_train, y_test train_test_split( X, y, test_sizetest_size, random_staterandom_state, stratifyy ) return X_train, X_test, y_train, y_test, scaler, imputer6.2 机器学习模型训练使用多种算法进行模型训练和比较# model_training.py from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier from sklearn.svm import SVC from sklearn.linear_model import LogisticRegression from sklearn.model_selection import cross_val_score, GridSearchCV from sklearn.metrics import accuracy_score, roc_auc_score, classification_report import xgboost as xgb class SurvivalPredictor: def __init__(self): self.models { random_forest: RandomForestClassifier(n_estimators100, random_state42), xgboost: xgb.XGBClassifier(n_estimators100, random_state42), svm: SVC(probabilityTrue, random_state42), logistic: LogisticRegression(random_state42) } self.best_model None self.best_score 0 def train_models(self, X_train, y_train, X_test, y_test): 训练多个模型并选择最佳性能 results {} for name, model in self.models.items(): print(f训练 {name}...) model.fit(X_train, y_train) y_pred model.predict(X_test) y_prob model.predict_proba(X_test)[:, 1] accuracy accuracy_score(y_test, y_pred) auc roc_auc_score(y_test, y_prob) results[name] { model: model, accuracy: accuracy, auc: auc, predictions: y_pred, probabilities: y_prob } print(f{name} - 准确率: {accuracy:.3f}, AUC: {auc:.3f}) if auc self.best_score: self.best_score auc self.best_model model return results def hyperparameter_tuning(self, X_train, y_train): 对最佳模型进行超参数调优 if isinstance(self.best_model, RandomForestClassifier): param_grid { n_estimators: [50, 100, 200], max_depth: [10, 20, None], min_samples_split: [2, 5, 10] } grid_search GridSearchCV( RandomForestClassifier(random_state42), param_grid, cv5, scoringroc_auc ) grid_search.fit(X_train, y_train) self.best_model grid_search.best_estimator_ return self.best_model7. SHAP可解释性分析7.1 SHAP值计算与可视化SHAP分析是理解模型决策的关键# shap_analysis.py import shap import matplotlib.pyplot as plt import pandas as pd import numpy as np class SHAPAnalyzer: def __init__(self, model, feature_names): self.model model self.feature_names feature_names self.explainer None self.shap_values None def create_explainer(self, X_train, X_test): 创建SHAP解释器并计算SHAP值 if hasattr(self.model, predict_proba): # 对于树模型使用TreeExplainer self.explainer shap.TreeExplainer(self.model) self.shap_values self.explainer.shap_values(X_test) else: # 对于其他模型使用KernelExplainer self.explainer shap.KernelExplainer(self.model.predict_proba, X_train) self.shap_values self.explainer.shap_values(X_test) return self.shap_values def summary_plot(self, X_test, max_display20): 生成特征重要性摘要图 if self.shap_values is None: raise ValueError(请先计算SHAP值) plt.figure(figsize(10, 8)) shap.summary_plot(self.shap_values, X_test, feature_namesself.feature_names, max_displaymax_display, showFalse) plt.tight_layout() plt.savefig(shap_summary.png, dpi300, bbox_inchestight) plt.close() def dependence_plot(self, feature_name, X_test, feature_indexNone): 生成特定特征的依赖图 if feature_index is None: feature_index list(self.feature_names).index(feature_name) plt.figure(figsize(10, 6)) shap.dependence_plot(feature_index, self.shap_values, X_test, feature_namesself.feature_names, showFalse) plt.title(fSHAP依赖图 - {feature_name}) plt.tight_layout() plt.savefig(fshap_dependence_{feature_name}.png, dpi300, bbox_inchestight) plt.close() def force_plot_single(self, instance_index, X_test): 生成单个预测的力导向图 plt.figure(figsize(12, 4)) shap.force_plot(self.explainer.expected_value, self.shap_values[instance_index, :], X_test[instance_index, :], feature_namesself.feature_names, matplotlibTrue, showFalse) plt.title(f单个预测SHAP解释 - 实例 {instance_index}) plt.tight_layout() plt.savefig(fshap_force_{instance_index}.png, dpi300, bbox_inchestight) plt.close()7.2 临床特征解释分析将SHAP分析结果转化为临床可理解的解释# clinical_interpretation.py import pandas as pd import numpy as np class ClinicalInterpreter: def __init__(self, shap_analyzer, feature_names): self.sha
这次我们来看一个医疗AI领域的实用项目——基于SHAP解释的放射组学模型专门用于预测全脑放疗后的患者生存获益情况。这个项目结合了影像分析、机器学习模型和可解释性AI技术让临床医生能够更直观地理解模型预测的依据。对于需要进行全脑放疗的脑转移患者来说准确预测治疗效果和生存期至关重要。传统的预测方法依赖临床经验而这个放射组学模型通过提取CT或MRI影像中的定量特征构建预测模型并用SHAP值来解释每个特征对预测结果的贡献度。这样既提高了预测的准确性又让模型决策过程变得透明可解释。本文将重点介绍这个项目的核心能力、适用场景、环境部署方法、模型训练流程、SHAP解释分析以及实际应用中的注意事项。无论你是医学影像研究者、临床医生还是对可解释AI感兴趣的开发者都能从中获得实用的技术方案。1. 核心能力速览能力项说明项目类型医疗影像分析 可解释机器学习主要功能从放射影像提取组学特征预测全脑放疗后生存获益提供SHAP可解释性分析输入数据CT或MRI脑部影像DICOM格式输出结果生存获益预测概率 特征重要性SHAP图模型架构放射组学特征提取 机器学习分类器如随机森林、XGBoost可解释性SHAP值分析可视化特征贡献度硬件要求CPU/GPU均可GPU加速特征提取过程内存需求8GB以上RAM处理大型影像数据集需要更多内存部署方式Python脚本、Jupyter Notebook、Web服务接口适合场景临床研究、疗效预测、医学影像分析项目2. 适用场景与使用边界这个放射组学模型主要适用于脑转移患者全脑放疗后的生存获益预测。在临床实践中医生需要判断哪些患者可能从全脑放疗中获益以及预期的生存期改善程度。传统方法主要依赖临床病理特征而放射组学提供了从影像中提取定量特征的新途径。适用场景包括放疗科医生评估治疗方案效果医学研究人员进行预后因素分析医院影像科建立智能预测系统临床决策支持工具开发使用边界需要特别注意模型预测结果仅供参考不能替代临床医生专业判断需要高质量的DICOM影像数据图像质量直接影响特征提取效果模型训练需要足够多的标注数据小样本容易过拟合SHAP解释基于特征重要性但不能证明因果关系涉及患者隐私数据必须进行脱敏处理符合医疗数据安全规范在实际应用中建议将模型作为辅助工具结合临床经验共同决策。对于关键医疗决策必须有多重验证机制。3. 环境准备与前置条件要运行这个放射组学项目需要准备以下环境操作系统要求Windows 10/11, macOS 10.14, Ubuntu 18.04 等主流系统建议使用Linux系统获得更好的性能表现Python环境Python 3.8-3.10版本3.11可能存在兼容性问题推荐使用conda或venv创建虚拟环境需要安装的关键包pyradiomics, shap, scikit-learn, pandas, numpy, matplotlib医学影像处理依赖SimpleITK或ITK用于DICOM文件读取PyDICOM用于DICOM元数据处理OpenCV或PIL用于图像预处理机器学习框架scikit-learn用于传统机器学习模型可选PyTorch或TensorFlow用于深度学习扩展硬件配置建议CPU: 4核以上推荐8核以上内存: 8GB最小16GB推荐处理大批量影像需要32GB存储: 至少50GB可用空间用于存储影像数据和中间结果GPU: 非必需但可加速大型影像处理CUDA 11.0数据准备收集脑部CT或MRI的DICOM数据准备对应的临床随访数据生存时间、治疗反应等数据需要经过伦理审批和脱敏处理4. 安装部署与启动方式4.1 创建Python虚拟环境# 使用conda创建环境 conda create -n radiomics_shap python3.9 conda activate radiomics_shap # 或者使用venv python -m venv radiomics_shap source radiomics_shap/bin/activate # Linux/macOS radiomics_shap\Scripts\activate # Windows4.2 安装核心依赖包# 安装放射组学核心包 pip install pyradiomics pip install SimpleITK pip install pydicom # 安装机器学习和可解释性包 pip install scikit-learn pip install xgboost pip install shap pip install pandas numpy matplotlib seaborn # 安装Jupyter用于交互式分析可选 pip install jupyter notebook4.3 项目结构准备创建以下目录结构来组织代码和数据radiomics_project/ ├── data/ │ ├── dicom/ # 原始DICOM影像 │ ├── masks/ # 感兴趣区域掩膜 │ └── clinical.csv # 临床数据 ├── features/ # 提取的特征文件 ├── models/ # 训练好的模型 ├── results/ # 分析结果和图表 └── scripts/ ├── extract_features.py # 特征提取脚本 ├── train_model.py # 模型训练脚本 ├── shap_analysis.py # SHAP分析脚本 └── utils.py # 工具函数4.4 基础验证脚本创建一个简单的验证脚本来测试环境是否正常# test_environment.py import sys import pkg_resources required_packages { pyradiomics: 5.0.0, shap: 0.41.0, scikit-learn: 1.2.0, pandas: 1.5.0 } print(检查环境依赖...) for package, version in required_packages.items(): try: installed_version pkg_resources.get_distribution(package).version print(f✓ {package}: {installed_version}) except pkg_resources.DistributionNotFound: print(f✗ {package} 未安装) print(\n测试基本功能...) try: import pyradiomics import shap from sklearn.ensemble import RandomForestClassifier print(✓ 核心包导入成功) except ImportError as e: print(f✗ 导入失败: {e})运行验证脚本python test_environment.py5. 特征提取与数据预处理5.1 DICOM数据读取与标准化放射组学分析的第一步是正确处理DICOM格式的医学影像# dicom_processor.py import pydicom import numpy as np import SimpleITK as sitk from radiomics import featureextractor class DICOMProcessor: def __init__(self, dicom_dir): self.dicom_dir dicom_dir self.extractor featureextractor.RadiomicsFeatureExtractor() def load_dicom_series(self): 加载DICOM序列并转换为SimpleITK图像 reader sitk.ImageSeriesReader() dicom_names reader.GetGDCMSeriesFileNames(self.dicom_dir) reader.SetFileNames(dicom_names) image reader.Execute() return image def create_mask(self, image, methodthreshold): 创建感兴趣区域掩膜 if method threshold: # 基于阈值的简单掩膜生成 mask sitk.BinaryThreshold(image, lowerThreshold-1000, upperThreshold300, insideValue1, outsideValue0) return mask def extract_radiomics_features(self, image, mask): 提取放射组学特征 try: features self.extractor.execute(image, mask) return features except Exception as e: print(f特征提取失败: {e}) return None5.2 放射组学特征配置配置特征提取参数确保提取的特征具有临床意义# feature_config.py import json # 放射组学特征提取配置 feature_settings { imageType: { Original: {}, Wavelet: {} }, featureClass: { firstorder: [], glcm: [Autocorrelation, JointAverage], glrlm: [RunLengthNonuniformity], glszm: [ZonePercentage], gldm: [DependenceVariance], ngtdm: [Coarseness], shape: [Maximum3DDiameter] }, setting: { binWidth: 25, resampledPixelSpacing: None, interpolator: sitkBSpline, label: 1 } } # 保存配置 with open(radiomics_settings.json, w) as f: json.dump(feature_settings, f, indent2)5.3 批量特征提取流程对于大规模数据集需要实现批量处理# batch_feature_extraction.py import os import pandas as pd from dicom_processor import DICOMProcessor def batch_extract_features(data_root, output_fileradiomics_features.csv): 批量提取所有患者的放射组学特征 patients [d for d in os.listdir(data_root) if os.path.isdir(os.path.join(data_root, d))] all_features [] for patient_id in patients: patient_path os.path.join(data_root, patient_id, dicom) mask_path os.path.join(data_root, patient_id, mask.nii.gz) if not os.path.exists(patient_path): continue try: processor DICOMProcessor(patient_path) image processor.load_dicom_series() mask sitk.ReadImage(mask_path) features processor.extract_radiomics_features(image, mask) if features: features[patient_id] patient_id all_features.append(features) except Exception as e: print(f患者 {patient_id} 处理失败: {e}) # 转换为DataFrame并保存 df_features pd.DataFrame(all_features) df_features.to_csv(output_file, indexFalse) return df_features6. 模型训练与验证6.1 数据准备与特征工程在训练模型前需要进行仔细的数据预处理# data_preparation.py import pandas as pd import numpy as np from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler from sklearn.impute import SimpleImputer class SurvivalDataPreprocessor: def __init__(self, features_file, clinical_file): self.features_df pd.read_csv(features_file) self.clinical_df pd.read_csv(clinical_file) def merge_datasets(self): 合并放射组学特征和临床数据 merged_df pd.merge(self.features_df, self.clinical_df, onpatient_id, howinner) return merged_df def preprocess_features(self, df): 特征预处理缺失值处理、标准化 # 分离特征和目标变量 X df.drop([patient_id, survival_status, survival_days], axis1) y (df[survival_days] 180).astype(int) # 6个月生存获益 # 处理缺失值 imputer SimpleImputer(strategymedian) X_imputed imputer.fit_transform(X) # 标准化特征 scaler StandardScaler() X_scaled scaler.fit_transform(X_imputed) return X_scaled, y, scaler, imputer def prepare_train_test(self, test_size0.2, random_state42): 准备训练测试集 merged_df self.merge_datasets() X, y, scaler, imputer self.preprocess_features(merged_df) X_train, X_test, y_train, y_test train_test_split( X, y, test_sizetest_size, random_staterandom_state, stratifyy ) return X_train, X_test, y_train, y_test, scaler, imputer6.2 机器学习模型训练使用多种算法进行模型训练和比较# model_training.py from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier from sklearn.svm import SVC from sklearn.linear_model import LogisticRegression from sklearn.model_selection import cross_val_score, GridSearchCV from sklearn.metrics import accuracy_score, roc_auc_score, classification_report import xgboost as xgb class SurvivalPredictor: def __init__(self): self.models { random_forest: RandomForestClassifier(n_estimators100, random_state42), xgboost: xgb.XGBClassifier(n_estimators100, random_state42), svm: SVC(probabilityTrue, random_state42), logistic: LogisticRegression(random_state42) } self.best_model None self.best_score 0 def train_models(self, X_train, y_train, X_test, y_test): 训练多个模型并选择最佳性能 results {} for name, model in self.models.items(): print(f训练 {name}...) model.fit(X_train, y_train) y_pred model.predict(X_test) y_prob model.predict_proba(X_test)[:, 1] accuracy accuracy_score(y_test, y_pred) auc roc_auc_score(y_test, y_prob) results[name] { model: model, accuracy: accuracy, auc: auc, predictions: y_pred, probabilities: y_prob } print(f{name} - 准确率: {accuracy:.3f}, AUC: {auc:.3f}) if auc self.best_score: self.best_score auc self.best_model model return results def hyperparameter_tuning(self, X_train, y_train): 对最佳模型进行超参数调优 if isinstance(self.best_model, RandomForestClassifier): param_grid { n_estimators: [50, 100, 200], max_depth: [10, 20, None], min_samples_split: [2, 5, 10] } grid_search GridSearchCV( RandomForestClassifier(random_state42), param_grid, cv5, scoringroc_auc ) grid_search.fit(X_train, y_train) self.best_model grid_search.best_estimator_ return self.best_model7. SHAP可解释性分析7.1 SHAP值计算与可视化SHAP分析是理解模型决策的关键# shap_analysis.py import shap import matplotlib.pyplot as plt import pandas as pd import numpy as np class SHAPAnalyzer: def __init__(self, model, feature_names): self.model model self.feature_names feature_names self.explainer None self.shap_values None def create_explainer(self, X_train, X_test): 创建SHAP解释器并计算SHAP值 if hasattr(self.model, predict_proba): # 对于树模型使用TreeExplainer self.explainer shap.TreeExplainer(self.model) self.shap_values self.explainer.shap_values(X_test) else: # 对于其他模型使用KernelExplainer self.explainer shap.KernelExplainer(self.model.predict_proba, X_train) self.shap_values self.explainer.shap_values(X_test) return self.shap_values def summary_plot(self, X_test, max_display20): 生成特征重要性摘要图 if self.shap_values is None: raise ValueError(请先计算SHAP值) plt.figure(figsize(10, 8)) shap.summary_plot(self.shap_values, X_test, feature_namesself.feature_names, max_displaymax_display, showFalse) plt.tight_layout() plt.savefig(shap_summary.png, dpi300, bbox_inchestight) plt.close() def dependence_plot(self, feature_name, X_test, feature_indexNone): 生成特定特征的依赖图 if feature_index is None: feature_index list(self.feature_names).index(feature_name) plt.figure(figsize(10, 6)) shap.dependence_plot(feature_index, self.shap_values, X_test, feature_namesself.feature_names, showFalse) plt.title(fSHAP依赖图 - {feature_name}) plt.tight_layout() plt.savefig(fshap_dependence_{feature_name}.png, dpi300, bbox_inchestight) plt.close() def force_plot_single(self, instance_index, X_test): 生成单个预测的力导向图 plt.figure(figsize(12, 4)) shap.force_plot(self.explainer.expected_value, self.shap_values[instance_index, :], X_test[instance_index, :], feature_namesself.feature_names, matplotlibTrue, showFalse) plt.title(f单个预测SHAP解释 - 实例 {instance_index}) plt.tight_layout() plt.savefig(fshap_force_{instance_index}.png, dpi300, bbox_inchestight) plt.close()7.2 临床特征解释分析将SHAP分析结果转化为临床可理解的解释# clinical_interpretation.py import pandas as pd import numpy as np class ClinicalInterpreter: def __init__(self, shap_analyzer, feature_names): self.sha
