1# 建議使用 uv 建立隔離環境
 2uv venv temporai-env
 3source temporai-env/bin/activate
 4
 5# 從 PyPI 安裝
 6pip install temporai
 7
 8# 或從 source 安裝（取得最新版本）
 9git clone https://github.com/vanderschaarlab/temporai.git
10cd temporai
11pip install .

1# 安裝含測試依賴
2pip install temporai[testing]
3
4# 安裝完整開發依賴（含 testing + linting + docs）
5pip install temporai[dev]
6
7# 執行測試
8pytest -vsx

1conda create -n temporai python=3.10
2conda activate temporai
3pip install temporai

1from tempor import plugin_loader
2
3# 列出所有可用插件
4plugins = plugin_loader.list()
5print(f"Total plugins: {len(plugins)}")
6for category, methods in plugins.items():
7    print(f"  {category}: {methods}")

 1from tempor import plugin_loader
 2
 3# ----- 步驟 1: 載入資料 -----
 4# PBC (Primary Biliary Cholangitis) 是內建的時序存活分析資料集
 5dataset = plugin_loader.get(
 6    "time_to_event.pbc",
 7    plugin_type="datasource"
 8).load()
 9
10print(f"樣本數: {len(dataset)}")
11print(f"靜態特徵: {dataset.static.dataframe().columns.tolist()}")
12print(f"時序特徵: {dataset.time_series.dataframe().columns.tolist()}")
13
14# ----- 步驟 2: 初始化 Dynamic-DeepHit 模型 -----
15model = plugin_loader.get(
16    "time_to_event.dynamic_deephit",
17    n_iter=200,        # 訓練迭代次數
18    batch_size=100,    # 批次大小
19)
20
21# ----- 步驟 3: 訓練 -----
22model.fit(dataset)
23
24# ----- 步驟 4: 預測不同時間點的風險 -----
25# horizons: 預測 25%, 50%, 75% 時間點的風險
26prediction = model.predict(
27    dataset,
28    horizons=[0.25, 0.50, 0.75]
29)
30
31# 查看第一位病患的風險預測
32print("Patient 0 risk predictions:")
33print(prediction[0])
34
35# ----- 步驟 5: 模型序列化 -----
36from tempor.utils.serialization import save, load
37
38buffer = save(model)
39reloaded_model = load(buffer)  # 可重新使用

 1import numpy as np
 2from tempor import plugin_loader
 3
 4# ----- 步驟 1: 載入含時序治療的資料集 -----
 5dataset = plugin_loader.get(
 6    "treatments.temporal.dummy_treatments",
 7    plugin_type="datasource",
 8    temporal_covariates_missing_prob=0.0,    # 無缺失值（簡化示範）
 9    temporal_treatments_n_features=1,         # 1 個治療變數
10    temporal_treatments_n_categories=2,       # 2 種治療（例：藥物 A vs 藥物 B）
11).load()
12
13# ----- 步驟 2: 初始化 CRN Regressor -----
14model = plugin_loader.get(
15    "treatments.temporal.regression.crn_regressor",
16    epochs=20,
17    encoder_rnn_type="LSTM",   # 使用 LSTM 作為 encoder
18)
19
20# ----- 步驟 3: 訓練 -----
21model.fit(dataset)
22
23# ----- 步驟 4: 定義反事實情境 -----
24# 為每位病患定義預測時間範圍（取後半段時間點）
25horizons = [
26    tc.time_indexes()[0][len(tc.time_indexes()[0]) // 2:]
27    for tc in dataset.time_series
28]
29
30# 定義兩種治療情境：全程用藥 A vs 全程用藥 B
31treatment_scenarios = [
32    [
33        np.asarray([1] * len(h)),    # 情境 1: 全程治療 A
34        np.asarray([0] * len(h)),    # 情境 2: 全程治療 B
35    ]
36    for h in horizons
37]
38
39# ----- 步驟 5: 預測反事實結果 -----
40counterfactuals = model.predict_counterfactuals(
41    dataset,
42    horizons=horizons,
43    treatment_scenarios=treatment_scenarios,
44)
45
46# 查看第一位病患的兩種治療情境預測
47print("Patient 0 - Treatment A outcome:", counterfactuals[0][0])
48print("Patient 0 - Treatment B outcome:", counterfactuals[0][1])
49print("Treatment Effect (A - B):", 
50      counterfactuals[0][0].values - counterfactuals[0][1].values)

 1from tempor import plugin_loader
 2from tempor.methods.pipeline import pipeline
 3from tempor.benchmarks import benchmark_models
 4from tempor.automl.seeker import PipelineSeeker
 5
 6# ===== Part A: 手動 Pipeline 組合 =====
 7
 8# 載入資料
 9dataset = plugin_loader.get(
10    "time_to_event.pbc",
11    plugin_type="datasource"
12).load()
13
14# 建立 Pipeline: 前處理 (scaling) -> 模型 (存活分析)
15MyPipeline = pipeline([
16    "preprocessing.imputation.temporal.ffill",           # 時序缺失值前向填補
17    "preprocessing.scaling.temporal.ts_minmax_scaler",   # 時序 MinMax 標準化
18    "time_to_event.dynamic_deephit",                     # Dynamic-DeepHit 存活分析
19])
20
21# 初始化 Pipeline 並設定超參數
22pipe = MyPipeline({"dynamic_deephit": {"n_iter": 100}})
23pipe.fit(dataset)
24prediction = pipe.predict(dataset, horizons=[2.0, 4.0, 6.0])
25
26# ===== Part B: Benchmark 模型比較 =====
27
28testcases = [
29    ("Pipeline (FFill+MinMax+DeepHit)", pipe),
30    ("DeepHit alone", plugin_loader.get("time_to_event.dynamic_deephit", n_iter=100)),
31    ("TS-CoxPH", plugin_loader.get("time_to_event.ts_coxph", n_iter=100)),
32]
33
34aggr_score, per_test_score = benchmark_models(
35    task_type="time_to_event",
36    tests=testcases,
37    data=dataset,
38    n_splits=2,         # 2-fold Cross-Validation
39    random_state=42,
40    horizons=[2.0, 4.0, 6.0],
41)
42
43print("=== 模型比較結果 ===")
44print(aggr_score)
45
46# ===== Part C: AutoML 自動搜尋最佳 Pipeline =====
47
48seeker = PipelineSeeker(
49    study_name="pre_ind_survival_study",
50    task_type="prediction.one_off.classification",
51    estimator_names=[
52        "cde_classifier",     # Neural CDE
53        "ode_classifier",     # Neural ODE
54        "nn_classifier",      # RNN/LSTM/Transformer
55    ],
56    metric="aucroc",           # 以 AUROC 為評估指標
57    dataset=dataset,
58    return_top_k=3,            # 回傳前 3 名
59    num_iter=50,               # 搜尋 50 次迭代
60    tuner_type="bayesian",     # Bayesian Optimization
61    # 前處理選項
62    static_imputers=["static_tabular_imputer"],
63    temporal_imputers=["ffill", "bfill"],
64    temporal_scalers=["ts_minmax_scaler"],
65)
66
67# 執行搜尋（可能需要數分鐘到數小時）
68best_pipelines, best_scores = seeker.search()
69
70print("=== AutoML Top 3 Pipelines ===")
71for i, (pipe, score) in enumerate(zip(best_pipelines, best_scores)):
72    print(f"  Rank {i+1}: Score={score:.4f}")

1Bio-SDG Domain 1: 醫療結構化資料合成
2├── A. Patient Simulation    ← Synthea (全病患生命週期模擬)
3├── B. Privacy-Preserving    ← DP 工具群 (差分隱私)
4├── C. Medical Tabular SDG   ← synthcity, COR-GAN (靜態 EHR 合成)
5├── D. Clinical Trial Data   ← VICTRE, tsynth (虛擬臨床試驗)
6├── E. Medical Time Series   ← **TemporAI** (時序 ML 分析)  ◄ 你在這裡
7└── F. Specialized           ← SDM-TIB (特化領域)

 1# 理想的整合流程（概念示範）
 2# 1. 用 synthcity 生成合成時序 EHR
 3from synthcity.plugins import Plugins
 4syn_model = Plugins().get("timegan")
 5syn_model.fit(real_temporal_data)
 6synthetic_data = syn_model.generate(count=1000)
 7
 8# 2. 轉換為 TemporAI Dataset 格式
 9from tempor.data.dataset import TemporalDataset
10dataset = TemporalDataset(
11    static=synthetic_static_df,
12    time_series=synthetic_temporal_df,
13    targets=synthetic_targets_df,
14)
15
16# 3. 用 TemporAI 進行分析
17from tempor import plugin_loader
18model = plugin_loader.get("time_to_event.dynamic_deephit")
19model.fit(dataset)

1# 安裝 MIMIC-IV Data Pipeline 適配器
2pip install temporai-mivdp

1Synthea FHIR JSON → pandas DataFrame (ETL) → TemporAI TemporalDataset

1Synthea ──FHIR──→ ETL ──→ ┐
2                           │
3synthcity (TimeGAN) ──────→├──→ TemporAI ──→ 存活預測 / 治療效果 / 預測
4                           │                    │
5MIMIC-IV ─temporai-mivdp─→┘                    ├──→ Benchmark 報告
6                                                ├──→ 反事實分析
7DP tools (隱私保護) ──→ 合成資料 ──→ TemporAI   └──→ AutoML 最佳模型