HSBP-DB: A Synchronized Phonocardiography and Blood Pressure Database for Multi-Class Hypertension Classification
Description
HSBP-DB is a novel, annotated, and synchronised database of phonocardiography (PCG) and arterial blood pressure (BP) recordings, developed to address the absence of a purpose-built, openly accessible benchmark for multi-class hypertension staging from cardiac acoustic signals. The dataset comprises 400 raw, uncompressed .wav PCG recordings acquired from 50 adult participants (32 male, 18 female; age 22.70 ± 2.00 years) under two physiological conditions: resting-baseline (5 sessions per participant) and post-exercise (3 sessions per participant). PCG signals were captured using the 3M Littmann CORE Digital Stethoscope (Model 8480) positioned at Erb's point, sampled at 4,000 Hz. Concurrent ground-truth blood pressure was measured using the Contec ABPM50 oscillometric ambulatory monitor (ESH-IP 2010 validated), applying a sequential reference-first acquisition protocol to ensure temporal alignment between the BP label and the corresponding PCG recording. Each recording is labelled at the recording level according to the 2017 AHA/ACC four-class blood pressure classification schema (Normal, Elevated, Hypertension Stage 1, Hypertension Stage 2). All metadata — participant identifier, session index, physiological state (resting/post-exercise), systolic pressure, diastolic pressure, and mean arterial pressure — are self-contained within a structured, anonymised filename convention, eliminating dependency on external metadata files. Files are organised into four class-stratified directories (Normal, Elevated, HTN-1, HTN-2) corresponding to the AHA/ACC classification of each recording. HSBP-DB is intended to support reproducible research in cuffless blood pressure estimation, PCG-based haemodynamic biomarker extraction, ordinal/multi-class classification modelling, and multimodal cardiovascular signal processing. Ethical approval was granted by the Universiti Kuala Lumpur Research Ethics Committee (UREC Ref: UNIKL REC/2025/FRGS/APPV/01). All data were anonymised at the point of acquisition; no personally identifiable information is contained in or recoverable from the deposited files.
Files
Steps to reproduce
PCG and BP data were collected at the Gymnasium, Universiti Kuala Lumpur- Malaysian Institute of Industrial Technology (UniKL MITEC), Johor Bahru, Malaysia, under UREC approval (Ref: UNIKL REC/2025/FRGS/APPV/01). Each of the 50 participants completed eight sessions: five resting-baseline (RT: R1–R5) and three post-exercise (EX: R1–R3). Resting sessions followed ≥5 minutes of quiet seated rest; post-exercise sessions followed a supervised moderate-intensity step-test, with BP and PCG acquired within a standardised time window after exercise cessation. A sequential reference-first protocol was enforced for every acquisition. Ground-truth systolic (SBP) and diastolic (DBP) blood pressures were measured first using the Contec ABPM50 oscillometric ambulatory monitor (ESH-IP 2010 validated; ±3 mmHg), applied to the left upper arm with the participant seated upright and the arm supported at heart level. PCG signals were then acquired using the 3M Littmann CORE Digital Stethoscope (Model 8480) positioned at Erb's point (left sternal border, fourth intercostal space), sampled at 4,000 Hz, and exported as raw .wav via the Eko application. Prior to deposit, recordings underwent a three-step standardisation pipeline: (1) monaural conversion; (2) DC offset removal; (3) amplitude normalisation to [−1, +1]. No spectral filtering was applied. Each recording was assigned a structured anonymisation identifier embedding the physiological state (RT/EX), participant code, session index, SBP, DBP, and MAP. A custom automated pipeline parsed SBP and DBP from each filename, assigned AHA/ACC 2017 class labels (Normal, Elevated, HTN-1, HTN-2), and allocated each .wav to its corresponding class-stratified directory.
Institutions
- University of Kuala LumpurKuala Lumpur, Kuala Lumpur