1ICD-9 (MIMIC-III) ──→ ICD-10 ──→ PhecodeX ──→ PhecodeXm (higher hierarchy)
2                                      ↑
3SNOMED-CT (Synthea) ──→ ICD-10 ──────┘

 1synthEHRella/
 2├── environment.yaml              # Conda 環境定義 (Python 3.10 + CUDA 12.1)
 3├── setup.py                      # pip install 入口
 4├── example_commands/             # 範例 shell script
 5│   ├── run_preprocessing.sh
 6│   ├── run_generation.sh
 7│   ├── run_postprocessing.sh
 8│   ├── run_evaluation_mimic3.sh
 9│   └── run_evaluation_mimic4.sh
10└── synthEHRella/
11    ├── run_preprocessing.py      # Step 1 CLI
12    ├── run_generation.py         # Step 2 CLI (方法分派器)
13    ├── run_postprocessing.py     # Step 3 CLI
14    ├── run_evaluation.py         # Step 4 CLI
15    ├── data/
16    │   └── methods/              # 各方法的原始碼
17    │       ├── cor-gan/          # CorGAN + MedGAN + VAE
18    │       ├── EHRDiff/          # Score-based Diffusion
19    │       ├── PromptEHR/        # LLM-based 生成
20    │       ├── plasmode/         # R 統計方法
21    │       └── synthea/          # (需自行下載 .jar)
22    ├── evaluation/
23    │   ├── fidelity.py           # 保真度指標
24    │   ├── utility.py            # 效用指標 (TRTR/TSTR/TSRTR)
25    │   └── privacy.py            # 隱私攻擊 (MIA/AIA)
26    └── utils/
27        ├── data_transform.py     # 編碼轉換引擎
28        ├── ICD9toICD10Mapping.json
29        ├── icd10_to_phecodex_mapping.json
30        ├── phecodex_to_phecodexm_mapping.json
31        └── snomed2icd10.json

 1# 1. 取得原始碼
 2git clone https://github.com/chenxran/synthEHRella.git
 3cd synthEHRella
 4
 5# 2. 建立 Conda 環境 (含 PyTorch + CUDA + scikit-learn + transformers 等)
 6conda env create -f environment.yaml
 7conda activate synthEHRella
 8
 9# 3. 安裝套件
10pip install .
11
12# 4. (選用) 下載 Synthea JAR
13# 從 https://github.com/synthetichealth/synthea/releases 下載 synthea-with-dependencies.jar
14# 放至 synthEHRella/synthEHRella/data/methods/synthea/

 1# ========================================
 2# Step 1: 前處理 MIMIC-III
 3# ========================================
 4MIMIC_III_ADMISSION="data/physionet.org/files/mimiciii/1.4/ADMISSIONS.csv"
 5MIMIC_III_DIAGNOSIS="data/physionet.org/files/mimiciii/1.4/DIAGNOSES_ICD.csv"
 6MIMIC_III_PATIENTS="data/physionet.org/files/mimiciii/1.4/PATIENTS.csv"
 7OUTPUT="testing_output/"
 8
 9python -m synthEHRella.run_preprocessing \
10    $MIMIC_III_ADMISSION \
11    $MIMIC_III_DIAGNOSIS \
12    $MIMIC_III_PATIENTS \
13    $OUTPUT binary mimic3
14
15# 輸出檔案：
16#   testing_output/processed_mimic3.matrix   ← binary matrix (訓練用)
17#   testing_output/processed_mimic3.pid      ← patient ID list
18#   testing_output/processed_mimic3.types    ← ICD-9 code → column index mapping
19#   testing_output/mimic3-real-phecodex.npy  ← PhecodeX 編碼 (視覺化用)
20#   testing_output/mimic3-real-phecodexm.npy ← PhecodeXm 編碼 (量化評估用)
21
22# ========================================
23# Step 2: 用 CorGAN 生成 50,000 筆合成資料
24# ========================================
25NUM_GEN_SAMPLES=50000
26mkdir -p testing_output/corgan
27
28python -m synthEHRella.run_generation \
29    corgan \
30    --real_training_data_path testing_output/processed_mimic3.matrix \
31    --ckpt_dir testing_output/corgan \
32    --num_gen_samples $NUM_GEN_SAMPLES \
33    --params "--n_epochs 1000"
34
35# 輸出：testing_output/corgan/synthetic-50000.npy
36
37# ========================================
38# Step 3: 後處理 — ICD-9 → PhecodeXm
39# ========================================
40python -m synthEHRella.run_postprocessing corgan \
41    --data_path testing_output/corgan/synthetic-50000.npy \
42    --output_path testing_output/postprocessed/corgan-synthetic-phecodexm.npy
43
44# ========================================
45# Step 4: 三維評估
46# ========================================
47python -m synthEHRella.run_evaluation corgan \
48    --synthetic_data_path testing_output/postprocessed/corgan-synthetic-phecodexm.npy \
49    --real_eval_data_path testing_output/mimic3-real-phecodexm.npy \
50    --output_dir testing_output/evaluation/corgan-mimic3-evaluation.json
51
52# 輸出 JSON 包含所有 Fidelity + Utility + Privacy 指標

 1#!/bin/bash
 2# batch_benchmark.sh — 批次基準測試
 3
 4NUM_GEN_SAMPLES=50000
 5REAL_DATA="testing_output/processed_mimic3.matrix"
 6REAL_EVAL="testing_output/mimic3-real-phecodexm.npy"
 7METHODS=("corgan" "medgan" "vae" "promptehr" "resample" "pbr")
 8
 9# 需要額外環境的方法另外處理：
10# - ehrdiff: 需要自己的訓練資料格式
11# - synthea: 需要 Java + synthea-with-dependencies.jar
12# - plasmode: 需要 R 4.2+
13
14for METHOD in "${METHODS[@]}"; do
15    echo "=== Processing $METHOD ==="
16    mkdir -p testing_output/$METHOD
17
18    # Step 2: 生成
19    python -m synthEHRella.run_generation \
20        $METHOD \
21        --real_training_data_path $REAL_DATA \
22        --ckpt_dir testing_output/$METHOD \
23        --num_gen_samples $NUM_GEN_SAMPLES
24
25    # Step 3: 後處理
26    # 不同方法的輸出路徑不同，需要對應調整
27    case $METHOD in
28        corgan|vae)
29            SYNTH_PATH="testing_output/$METHOD/synthetic-${NUM_GEN_SAMPLES}.npy" ;;
30        medgan)
31            SYNTH_PATH="testing_output/$METHOD/synthetic.npy" ;;
32        promptehr)
33            SYNTH_PATH="testing_output/$METHOD/promptehr-synthetic.npy" ;;
34        resample|pbr)
35            SYNTH_PATH="testing_output/$METHOD/${METHOD}-synthetic.npy" ;;
36    esac
37
38    python -m synthEHRella.run_postprocessing $METHOD \
39        --data_path $SYNTH_PATH \
40        --output_path testing_output/postprocessed/${METHOD}-synthetic-phecodexm.npy
41
42    # Step 4: 評估
43    python -m synthEHRella.run_evaluation $METHOD \
44        --synthetic_data_path testing_output/postprocessed/${METHOD}-synthetic-phecodexm.npy \
45        --real_eval_data_path $REAL_EVAL \
46        --output_dir testing_output/evaluation/${METHOD}-mimic3-evaluation.json
47done
48
49echo "=== All evaluations complete ==="
50echo "Results in testing_output/evaluation/"

 1import numpy as np
 2from synthEHRella.evaluation.fidelity import (
 3    compute_prevalence,
 4    compute_correlation,
 5    coappearance_matrix,
 6    discriminative_score
 7)
 8from synthEHRella.evaluation.utility import trtr, tstr, tsrtr
 9from synthEHRella.evaluation.privacy import (
10    membership_inference_attack,
11    attribute_inference_attack
12)
13
14# 載入資料 — 假設已轉換為 PhecodeXm 編碼的 binary matrix
15real_data = np.load("testing_output/mimic3-real-phecodexm.npy")
16synthetic_data = np.load("testing_output/postprocessed/corgan-synthetic-phecodexm.npy")
17
18# ── Fidelity 保真度 ──
19real_prev = compute_prevalence(real_data)
20synth_prev = compute_prevalence(synthetic_data)
21
22# 最大平均差異
23mmd = np.abs(real_prev - synth_prev).max()
24print(f"MMD (Max Mean Discrepancy): {mmd:.4f}")
25
26# 相關矩陣 Frobenius 距離
27real_corr = compute_correlation(real_data.astype(float))
28synth_corr = compute_correlation(synthetic_data.astype(float))
29corr_dist = np.linalg.norm(real_corr - synth_corr, 'fro')
30print(f"Correlation Frobenius Distance: {corr_dist:.4f}")
31
32# 判別器分數 (越接近 0.5 越好 → 無法區分真假)
33disc = discriminative_score(real_data, synthetic_data)
34print(f"Discriminative AUC: {disc['auc']:.4f}")
35
36# ── Utility 效用 ──
37# TRTR: 用真實資料訓練，真實資料測試 (基線)
38# TSTR: 用合成資料訓練，真實資料測試 (核心指標)
39# TSRTR: 用合成+真實訓練，真實資料測試 (資料增強)
40trtr_res = trtr(real_data, index=0)
41tstr_res = tstr(real_data, synthetic_data, index=0)
42tsrtr_res = tsrtr(real_data, synthetic_data, index=0)
43
44print(f"\nTRTR AUC: {trtr_res['auc']:.4f} (baseline)")
45print(f"TSTR AUC: {tstr_res['auc']:.4f} (synthetic only)")
46print(f"TSRTR AUC: {tsrtr_res['auc']:.4f} (synthetic + real)")
47print(f"TSTR-TRTR gap: {tstr_res['auc'] - trtr_res['auc']:.4f}")
48
49# ── Privacy 隱私 ──
50# Membership Inference Attack (MIA; 成員推論攻擊)
51mia = membership_inference_attack(real_data, synthetic_data)
52print(f"\nMIA - Mean min L2 distance: {mia['mean_min_distances']:.4f}")
53
54# Attribute Inference Attack (AIA; 屬性推論攻擊)
55# 選 10 個最平衡 + 10 個最不平衡的特徵作為未知屬性
56prevalence = compute_prevalence(real_data)
57sorted_idx = np.argsort(np.abs(prevalence - 0.5))
58unknown_idx = sorted_idx[:10].tolist() + sorted_idx[-10:].tolist()
59known_idx = list(set(range(real_data.shape[1])) - set(unknown_idx))
60
61aia_f1 = attribute_inference_attack(real_data, synthetic_data, known_idx, unknown_idx)
62print(f"AIA F1 Score: {aia_f1:.4f} (lower = better privacy)")

 1# 1. 用 synthcity 生成合成資料
 2from synthcity.plugins import Plugins
 3plugin = Plugins().get("tvae")
 4plugin.fit(real_dataframe)
 5synthetic_df = plugin.generate(count=50000)
 6
 7# 2. 轉換為 SynthEHRella 需要的 binary numpy matrix
 8synthetic_matrix = synthetic_df.values.astype(float)
 9
10# 3. 確保編碼與 SynthEHRella 的 PhecodeXm 一致後，直接呼叫評估
11from synthEHRella.run_evaluation import fidelity_evaluation, utility_evaluation, privacy_evaluation
12fidelity = fidelity_evaluation(real_phecodexm, synthetic_phecodexm)