m46012002/smart-mattress-bcg

Smart Mattress BCG Heart Rate Dataset

Posture-Robust Lightweight Smart Mattress Edge Heart Rate Monitoring on Low-Cost Pressure Sensor Arrays

Wu Jie-Neng -- National Yang Ming Chiao Tung University, Master's Thesis, 2026 Code & models: https://github.com/m46012002/smart-mattress-edge-hr

---

Dataset Summary

This dataset contains ballistocardiography (BCG) signals collected from a 32-channel FSR (force-sensitive resistor) pressure sensor array embedded in a latex mattress. Simultaneously, a fingertip MAX30105 PPG sensor provides heart rate ground truth.

33 healthy subjects performed 5 body postures (supine, prone, left lateral, right lateral, and on/off-bed transitions) while lying on the mattress. The dataset was used to train and evaluate a lightweight CNN-based channel reliability model for wearable-free, posture-robust heart rate monitoring at the edge.

Key statistics:

• 33 subjects (age 18-35, mixed sex)
• 5 postures per subject
• 13.5 Hz sample rate (32ch) / 27.03 Hz (16ch variant, subject 224)
• ~225 raw recording files, ~270 MB preprocessed

---

Dataset Structure

ESP32_recored/
|-- 100/                   # test/calibration subject (excluded from analysis)
|-- 101/ ... 110/          # Group 1: 10 subjects, 2-min per posture
|   |-- Front/             # Supine (face up)
|   |-- Back/              # Prone (face down)
|   |-- LeftSide/
|   |-- RightSide/
|   |-- UFront/            # Post-exercise variant (stair climbing before recording)
|   |-- UBack/
|   |-- ULeftSide/
|   `-- URightSide/
|-- 200/                   # test/calibration subject (excluded from analysis)
|-- 201/ ... 223/          # Group 2: 23 subjects, 5-min per posture
|   |-- Front/
|   |-- Back/
|   |-- LeftSide/
|   |-- RightSide/
|   `-- Leave/             # on-bed -> off-bed -> on-bed transition
`-- 224_16ch/              # Group 3: 1 subject, 16ch firmware (27.03 Hz)
    |-- Front/
    |-- Back/
    |-- LeftSide/
    |-- RightSide/
    `-- Leave/

Each posture folder contains one file: ppg_data_YYYYMMDD_HHMMSS.txt

---

Data Fields

Each file is comma-separated with no header row, 35 columns per line:

Column	Name	Type	Description
0	timestamp	string	PC wall-clock time "YYYY-MM-DD HH:MM:SS.mmm"
1-32	ch0-ch31	int16	FSR pressure ADC (higher = less pressure, lower = more pressure)
33	ppg	int32	MAX30105 green-LED reading (ground truth reference only)
34	esp32_us	int64	esp_timer_get_time() microsecond timestamp on ESP32

Example row:

2025-05-20 13:14:43.343,21879,21750,20843,...,5092,3267,...,21888,99840,991080717

ADC polarity: higher ADC value means less pressure (no contact 21800-22000); lower means more pressure (under body weight ~1900-11000). BCG signal = tiny dynamic fluctuation (100-1000 ADC counts) riding on a large static DC bias.

PPG note: the PPG column is the raw finger-clip signal used only for ground-truth HR labelling. It is not used as a model input at inference time.

---

Collection Protocol

Group 1 -- Subjects 101-110

• Sample rate: 13.5 Hz (32 channels)
• Duration per posture: 2 minutes (approx. 1,640 rows)
• Postures: Front, Back, RightSide, LeftSide (+ U-variants after stair climbing)
• U-variants (UFront, UBack, ULeftSide, URightSide): subject climbed stairs immediately before recording to create an elevated + recovering heart rate scenario. Merged with base posture in downstream processing.

Group 2 -- Subjects 201-223

• Sample rate: 13.5 Hz (32 channels)
• Duration per posture: 5 minutes (approx. 4,100 rows)
• Posture order: Front -> RightSide -> Back -> Leave -> LeftSide
• Front/RightSide/Back/LeftSide: subject performs 10-15 squats to elevate HR, then lies still for 5 minutes
• Leave: supine 2 min -> fully leave mattress 30 sec -> return supine 2.5 min (continuous recording)

Group 3 -- Subject 224

• Firmware variant: 16-channel only (ch0-ch15 active; ch16-ch31 are all zero)
• Sample rate: 27.03 Hz (halved scan channels -> doubled sample rate)
• Same 5 postures as Group 2

---

Subject Demographics

Group	Subjects	N	Weight range (kg)	Sex
1	101-110	10	44.5-105.0	7M, 3F
2	201-223	23	42.1-113.0	12M, 11F
3	224	1	--	--

Subjects 100 and 200 are test/calibration runs, excluded from all analyses (EXCLUDE_SUBJECTS = [100, 200]).

---

Hardware

Component	Specification
MCU	ESP32-S3 (Xtensa LX7, 240 MHz, 512 KB SRAM)
Pressure sensors	FSR x32, 16-column x 2-row grid on latex mattress
ADC	ADS1115 16-bit @ 860 SPS x 2 chips, via 4-channel MUX
PPG (reference)	MAX30105 fingertip clip, Green LED, 800 SPS
Firmware	ESP-IDF FreeRTOS; sequential polling ch0->ch31, ~2.28 ms/channel

Channel mapping:

ch = mux_idx x 8 + adc_idx
mux_idx in {0,1,2,3}   # MUX select pins GPIO35/36
adc_idx in {0..7}       # 0-3 -> ADS1115 #1 (addr 0x48), 4-7 -> ADS1115 #2 (addr 0x49)

Inter-channel sampling delay: ch0 is first, ch31 is last; max delay ~72 ms (0.97 samples at 13.5 Hz).

---

Usage

Load a single recording

import pandas as pd

cols = ['timestamp'] + [f'ch{i}' for i in range(32)] + ['ppg', 'esp32_us']
df = pd.read_csv('ESP32_recored/201/Front/ppg_data_20250601_100000.txt',
                 header=None, names=cols)

print(df.shape)           # (approx. 4100, 35)
print(df['ch0'].describe())

Load all subjects for ML training

The load_all_data() function in pipeline/common.py handles sliding-window extraction, PPG FFT ground-truth labelling, and per-channel reliability labelling.

from pipeline.common import load_all_data
windows, labels = load_all_data()
# windows: (N, 32, 128) float32  -- N windows, 32 channels, 128 samples each
# labels:  (N, 32) binary        -- 1 = reliable channel, 0 = unreliable

Ground-truth labelling parameters

Parameter	Value
Window size	128 samples
Overlap	50%
Window function	Hanning
HR frequency range	0.75-3.0 Hz (45-180 BPM)
Channel reliable if	bpm_error < 5 AND FFT SNR > 3
Outlier correction	Forward fill if adjacent HR jump > 20 BPM

---

Download

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id='m46012002/smart-mattress-bcg',
    repo_type='dataset',
    local_dir='data',
)

After download, place the ESP32_recored/ folder at data/ESP32_recored/ relative to the code repo root.

---

Related Resources

• Code & pre-trained models: https://github.com/m46012002/smart-mattress-edge-hr
• Method details: docs/method.md in the code repo
• Hardware & data format: docs/data_description.md in the code repo

---

Citation

@mastersthesis{wu2026smartmattress,
  author  = {Wu, Jie-Neng},
  title   = {Posture-Robust Lightweight Smart Mattress Edge Heart Rate Monitoring
             on Low-Cost Pressure Sensor Arrays},
  school  = {National Yang Ming Chiao Tung University},
  year    = {2026}
}

---

License

MIT License -- Copyright 2026 Wu Jie-Neng, NYCU