CWT-based PPG Preprocessing (PPG → CWT-PPG 2D Feature) （中文版本在下方）

This repository provides an open-source implementation of the PPG Continuous Wavelet Transform (CWT, Complex Morlet) preprocessing pipeline aimed at cuffless blood pressure estimation tasks.

Core Objective: Slice raw PPG waveforms into fixed windows and generate 256×256×2 (Real + Imag) CWT tensors, consistent with the reference paper, to serve as input for Deep Learning models.

Note: This repository focuses strictly on CWT data construction and reproducible preprocessing. Model training is not included in this scope but can be extended based on the output of this pipeline.

Resources (资源导航)

📄 Paper (论文): A U-net and Transformer Paralleled Network for Robust Cuffless Blood Pressure Estimation（IEEE SMC 2025 Oral）

📚 Dataset (数据集): Google Drive Link

👨‍🎓 Author (作者主页): Google Scholar Profile

---

1. Project Structure

CWT-based BP estimation/
│
├── data/                     # Raw Data (Read-only)
│   └── dataset_MIMICII.mat (as data is too big, plz download through google drive link: https://drive.google.com/file/d/18A8FjLt1Rdqz5PqMomooJ_-frzGzl2yB/view?usp=drive_link)
│
├── demos/                    # Visualization/Check scripts (Not for training)
│   └── signal_demo.m
│
├── preprocessing/            # Core Preprocessing Functions (Paper-Consistent)
│   ├── cwt_morlet_complex_conv.m
│   ├── reflect_pad_1d.m
│   ├── segment_ppg_windows.m
│   ├── ppg_to_cwt_tensor_batch.m
│   └── cwt_to_tensor_real_imag.m
│
├── scripts/                  # Executable Pipelines (Single/Batch)
│   ├── run_CWT_PPG_PaperConsistent.m
│   └── run_CWT_PPG_PaperConsistent_Batch.m
│
├── outputs/                  # Runtime outputs (Temporary)
│
└── CWT_Dataset/              # Generated Training Input Data (Batch Script Output)

## 2. Data Format (dataset_MIMICII.mat)

`data/dataset_MIMICII.mat` contains:

- `p`: a `1×N` cell array (e.g., `N=1000`)
- each `p{i}` is a `3×T double` matrix (three waveform channels)

This repo assumes the channel order (used by the demo scripts):

- `X(1,:)` → **PPG**
- `X(2,:)` → **ABP**
- `X(3,:)` → **ECG**

> The CWT pipeline only uses **PPG** (channel 1). ABP/ECG are for reference or later extensions.

---

## 3. Requirements

- MATLAB (recommended R2020a or newer; uses basic functions like `imresize`)
- Optional: Signal Processing Toolbox if you use `findpeaks`

---

## 4. Quick Start

### 4.1 Three-channel waveform demo

Purpose: verify data loading, channel order, and waveform shapes.

```matlab
cd demos
signal_demo

The script:

• loads ../data/dataset_MIMICII.mat
• uses idx=1 by default
• uses a fixed fs=125 Hz time axis
• plots PPG / ABP / ECG as three subplots

---

4.2 Single-sample CWT generation (paper-consistent)

Purpose: reproduce/inspect the CWT shape on one sample and debug parameters.

cd scripts
run_CWT_PPG_PaperConsistent

---

4.3 Batch CWT dataset generation (recommended)

Purpose: generate CWT tensors for the full dataset.

cd scripts
run_CWT_PPG_PaperConsistent_Batch

Default output folder:

CWT_Dataset/sample_0001/
CWT_Dataset/sample_0002/
...

Suggested per-sample files:

• X_cwt.mat: X_cwt with shape 256×256×2×num_windows (single)
• meta.mat: configuration such as fs, win_sec, stride_sec, frequency band, etc.
• (optional) ppg_peaks.mat: peak detection results if enabled

---

5. Method Summary (Paper-consistent Semantics)

Key design choices:

1. Sliding-window segmentation

• window length: win_sec = 5 s
• stride: stride_sec = 1 s (overlap exists)

2. Reflect padding

• pad_sec = 1 s (example)
• reduces boundary artifacts

3. Complex Morlet CWT

• generates a complex matrix W
• split into Real(W) and Imag(W) as two channels

4. Fixed output size

• out_hw = [256, 256]
• resize real/imag maps to a standard tensor size

Overlap note: If stride_sec < win_sec, adjacent windows overlap by win_sec - stride_sec.

---

6. Preprocessing Call Chain

Typical call chain in the batch pipeline:

run_CWT_PPG_PaperConsistent_Batch
    ↓
ppg_to_cwt_tensor_batch
    ↓
segment_ppg_windows
    ↓
reflect_pad_1d
    ↓
cwt_morlet_complex_conv
    ↓
cwt_to_tensor_real_imag

Function responsibilities:

• segment_ppg_windows.m: windowing by win_sec and stride_sec
• reflect_pad_1d.m: reflective padding for each window
• cwt_morlet_complex_conv.m: Complex Morlet CWT core
• cwt_to_tensor_real_imag.m: real/imag split + resize (and optional normalization)
• ppg_to_cwt_tensor_batch.m: orchestrator that outputs the final tensor

---

7. Can the output be used directly for deep learning?

Yes.

X_cwt.mat is already a fixed-size time–frequency tensor (two channels for real/imag). You can load it in MATLAB or Python:

• MATLAB: load('X_cwt.mat')
• Python: scipy.io.loadmat or h5py (if saved with -v7.3)

Typical tensor layouts:

• (num_windows, 2, 256, 256) or (2, 256, 256) (single window)

Labels (SBP/DBP) should be constructed separately (e.g., from ABP), depending on your training setup.

---

8. FAQ

Q1: “Cannot find dataset_MIMICII.mat”

Make sure the folder layout is:

• demos/ and data/ are siblings
• the file exists at data/dataset_MIMICII.mat

The demo typically uses:

dataFile = fullfile('..','data','dataset_MIMICII.mat');

Q2: Why does my CWT visualization not match the paper’s color tone exactly?

Common reasons:

• different colormap / rendering settings
• different normalization or log-compression
• different frequency-axis scaling (linear vs log)

This repo prioritizes numerical and tensor-construction consistency. Visualization styling can be aligned later by fixing the display parameters.

---

9. Citation

If you use this preprocessing implementation, please cite the associated paper:

• A U-net and Transformer Paralleled Network for Robust Cuffless Blood Pressure Estimation (IEEE SMC 2025 Oral, no index available yet, please star this project to stay updated)

---

CWT-based BP Estimation（PPG → CWT 预处理开源）

本仓库开源并复现论文中的 PPG 连续小波变换（CWT, Complex Morlet）预处理管线，用于无袖带血压估计任务的时频表示构建。核心目标是：将原始 PPG 波形切分为固定窗口，并生成论文一致的 256×256×2（Real + Imag） CWT 张量，作为深度学习模型输入。

说明：本仓库聚焦 CWT 数据构建与可复现的预处理实现。模型训练部分不在本仓库范围内（可在后续扩展）。

---

1. 项目结构

CWT-based BP estimation/
│
├── data/                     # 原始数据（只读）
│   └── dataset_MIMICII.mat   （数据过大，请使用谷歌云盘下载：https://drive.google.com/file/d/18A8FjLt1Rdqz5PqMomooJ_-frzGzl2yB/view?usp=drive_link）
│
├── demos/                    # 展示/检查脚本（不参与训练）
│   └── signal_demo.m
│
├── preprocessing/            # 预处理核心函数（论文一致）
│   ├── cwt_morlet_complex_conv.m
│   ├── reflect_pad_1d.m
│   ├── segment_ppg_windows.m
│   ├── ppg_to_cwt_tensor_batch.m
│   └── cwt_to_tensor_real_imag.m
│
├── scripts/                  # 可执行 pipeline（单次/批量）
│   ├── run_CWT_PPG_PaperConsistent.m
│   └── run_CWT_PPG_PaperConsistent_Batch.m
│
├── outputs/                  # 运行时输出（临时，可删）
│
└── CWT_Dataset/              # 生成的训练输入数据（批量脚本输出）

---

2. 数据格式说明（dataset_MIMICII.mat）

data/dataset_MIMICII.mat 中包含变量：

• p：1×N cell（例如 N=1000）
• 每个 p{i} 为 3×T double（三通道波形）

本仓库默认通道顺序为（已在 demo 中使用）：

• X(1,:) → PPG
• X(2,:) → ABP
• X(3,:) → ECG

pipeline 仅使用 PPG（第 1 行） 构建 CWT；ABP/ECG 用于参考或后续扩展。

---

3. 环境依赖

• MATLAB（建议 R2020a 及以上；需要 imresize 等基础函数）
• 若使用 findpeaks（可选）：Signal Processing Toolbox

---

4. 快速开始

4.1 波形展示 Demo（三通道）

用途：快速确认数据可读、通道顺序正确、波形形态正常。

cd demos
signal_demo

脚本将：

• 从 ../data/dataset_MIMICII.mat 读取 p
• 默认取 idx=1 的样本
• 固定 fs=125 Hz 构建时间轴
• 绘制三通道：PPG / ABP / ECG

---

4.2 单样本 CWT 生成（Paper-consistent）

用途：调参、对照论文图像形态，验证 Real/Imag 两通道输出。

cd scripts
run_CWT_PPG_PaperConsistent

---

4.3 批量生成 CWT Dataset（推荐）

用途：对全数据集生成训练输入张量。

cd scripts
run_CWT_PPG_PaperConsistent_Batch

默认输出目录：

CWT_Dataset/sample_0001/
CWT_Dataset/sample_0002/
...

建议每个 sample 文件夹包含：

• X_cwt.mat：X_cwt（256×256×2×num_windows，single）
• meta.mat：采样率、窗口长度、stride、频带等配置
• （可选）ppg_peaks.mat：峰值信息（若启用）

---

5. 预处理方法与参数（与论文一致的核心语义）

关键设计：

1. 滑动窗口切分

• window length：win_sec = 5 s
• stride：stride_sec = 1 s（存在重叠）

2. 边界反射 padding

• pad_sec = 1 s（示例）
• 减弱边界伪影

3. Complex Morlet CWT

• 输出复矩阵 W
• 拆分为 Real(W) 与 Imag(W) 两通道

4. 统一到图像尺寸

• out_hw = [256, 256]
• resize 到固定大小并拼成张量

重叠说明：只要 stride_sec < win_sec，窗口就会重叠，重叠长度为 win_sec - stride_sec。

---

6. preprocessing/ 目录函数职责（调用链）

批量脚本典型调用链：

run_CWT_PPG_PaperConsistent_Batch
    ↓
ppg_to_cwt_tensor_batch
    ↓
segment_ppg_windows
    ↓
reflect_pad_1d
    ↓
cwt_morlet_complex_conv
    ↓
cwt_to_tensor_real_imag

文件职责概览：

• segment_ppg_windows.m：按 window/stride 切分 PPG
• reflect_pad_1d.m：每个窗口反射 padding
• cwt_morlet_complex_conv.m：Complex Morlet CWT（核心）
• cwt_to_tensor_real_imag.m：拆 Real/Imag + resize（可含归一化）
• ppg_to_cwt_tensor_batch.m：整合流程并输出张量（核心 orchestrator）

---

7. 输出数据能不能直接用于深度学习训练？

可以。

X_cwt.mat 已经是固定尺寸的二维时频表示张量（Real/Imag 两通道）。可在 MATLAB / Python 中直接加载：

• MATLAB：load('X_cwt.mat')
• Python：scipy.io.loadmat 或 h5py（若 -v7.3）

常见张量布局：

• (num_windows, 2, 256, 256) 或 (2, 256, 256)（单窗）

标签（SBP/DBP）需要你根据训练设置另行构建（例如从 ABP 提取）。

---

8. 常见问题（FAQ）

Q1：运行时报“找不到 dataset_MIMICII.mat”

请确认目录结构：demos/ 与 data/ 同级，且 data/dataset_MIMICII.mat 存在。

脚本通常使用：

dataFile = fullfile('..','data','dataset_MIMICII.mat');

Q2：为什么 CWT 的颜色/色调和论文图不完全一致？

常见原因包括：

• colormap / 渲染设置不同
• 归一化/对数压缩/裁剪策略不同
• 频率轴缩放不同（linear vs log）

本仓库优先保证 数值与张量构造一致；可视化色调可在后续通过固定显示参数对齐论文图。

---

9. 引用（Citation）

如果你使用了本仓库的预处理实现，请引用对应论文：

• *A U-net and Transformer Paralleled Network for Robust Cuffless Blood Pressure Estimation*（IEEE SMC 2025）

---

10. 资源导航（Resources）

Resources (资源导航)

📄 Paper (论文): IEEE SMC 2025 (oral)

📚 Dataset (数据集): Google Drive Link

👨‍🎓 Author (作者主页): Google Scholar Profile