Balance ability test dataset for stroke patients based on flexible array pressure sensors

Abstract

In recent years, flexible array pressure sensors have been extensively employed in balance assessment research owing to their high sensitivity, flexible design, and multi-point measurement capability. This paper introduces a publicly available dataset for balance assessment, based on these sensors, developed to facilitate the detection and analysis of balance impairments in stroke patients. Data collection was conducted at Jiangbin Hospital, located in the Guangxi Zhuang Autonomous Region. The dataset comprises data from 53 volunteers (24 males, 29 females), including 36 healthy controls and 17 stroke patients. Participants' ages ranged from 16 to 86 years, heights from 1.45 to 1.83 m, and weights from 41 to 85 kg. Each participant performed five specific movements in a fixed sequence: sit-to-stand, in-place stepping, stepping on a soft mat with eyes closed, static balance with eyes closed, and static balance with eyes open. An array pressure acquisition system was used for data collection and storage. The public release of this dataset is intended to address the current scarcity of large-scale, high-quality data for balance assessment in stroke patients and to promote the advancement of balance assessment technologies.

Specifications Table

Subject	Computer Sciences
Specific subject area	Balance ability assessment
Type of data	Raw: Raw pressure data (.csv format) Image: Pressure heatmap (.png format)
Data collection	The dataset was collected using an array pressure acquisition system with a flexible seat cushion pressure sensor and a flexible plantar pressure sensor. Their specifications are as follows: The flexible seat cushion pressure sensor measures 430 mm by 420 mm, with an effective pressure-sensing area of 410 mm by 400 mm. It has 256 pressure-sensing points (16 × 16), a standard data transmission rate of 10 Hz, and a baud rate of 115200. The flexible plantar pressure sensor measures 520 mm by 493 mm, with an effective pressure-sensing area of 319.3 mm by 448.7 mm, and contains 2048 pressure-sensing points (32 × 64). It operates at a standard data transmission rate of 20 Hz and a baud rate of 460800. The array pressure acquisition system transmits data to the host computer via serial communication. The experimental design included a stroke patient group and a healthy control group (ages 18–86, weights 41–85 kg). Each participant sat upright on the flexible seat cushion pressure sensor during testing, with their feet flat on the flexible plantar pressure sensor. They performed five movement sequences: sit-to-stand, in-place stepping, eyes-closed soft mat stepping, static balance with eyes closed, and static balance with eyes open, with 10 to 15 seconds of rest between transitions.
Data source location	Institution: Jiangbin Hospital of Guangxi Zhuang Autonomous Region Region: Guangxi Zhuang Autonomous Region Country: China
Data accessibility	Repository name: Mendeley Data Data identification number: doi:10.17632/kcym3ptsyw.1 Direct URL to data: https://data.mendeley.com/datasets/kcym3ptsyw/1
Related research article	None

1. Value of the Data

• •The dataset comprises raw pressure data and heatmap images, enabling quantitative analysis and visualization of pressure patterns. This serves as a foundation for investigating balance ability in stroke patients, enhancing rehabilitation assessments, and developing strategies for sports injury prevention.
• •The dataset comprises five movements: sit-to-stand, in-place stepping, static balance with eyes open, static balance with eyes closed, and stepping on a soft mat with eyes closed. These movements reflect typical activities in the daily lives of stroke patients. By extracting features such as the centre of pressure (COP) and gait cycle using algorithms, researchers can conduct comprehensive assessments and analyses of balance abilities in stroke patients.
• •This dataset enables the training and evaluation of deep learning models for balance assessment. Researchers can either generate heatmaps from the raw pressure data or directly use the provided heatmaps for model training.
• •The open sharing of this dataset facilitates interdisciplinary research at the intersection of biomedical engineering and artificial intelligence. Analogous to medical imaging datasets used for disease detection, it provides a high-quality public resource for research on balance impairment in stroke. It also enhances the comparability and reproducibility of research findings, thereby supporting the development of balance assessment tools, health monitoring systems, and sensing technologies.
• •This dataset provides unique measured data on plantar and sitting pressure in stroke patients during targeted functional tasks. Publicly available datasets in this specialized field remain scarce. Even with 17 stroke patients, this dataset fills the gap of small-sample original data for balance and pressure assessment, provides pilot data, methodological references, and a basis for sample-size estimation for future large-sample cohort studies, and thus offers irreplaceable reusable value for basic and translational research. Furthermore, in terms of the volume of collected data, although only 17 stroke patients were enrolled in this study, the overall data scale remains sufficient after data segmentation and slicing, which can support the training of detection and analysis algorithms and related research in the relevant fields.

2. Background

Since the beginning of the 21st century, balance assessment methods, encompassing clinical observation, scales, and laboratory equipment, have become a key focus of research. While scale-based tests provide quantitative assessments, their reliability and validity are often influenced by evaluator experience [1,2]. Conversely, laboratory-grade balance testers, though considered the primary standard, are costly, thereby limiting their accessibility for patients to specialized research facilities [3,4].

Recent studies have highlighted the limitations of current balance assessment technologies. While motion capture-based gait analysis offers high accuracy, its utility is constrained by high costs, the requirement for dedicated laboratory facilities, and limited camera detection ranges [[5], [6], [7]]. Similarly, wearable sensors, though widely used, often necessitate assistance for donning, which restricts their practicality for independent use [[8], [9], [10]]. In contrast, flexible pressure sensor mats overcome many of these barriers. They are self-operated, making them particularly suitable for individuals with limb dysfunction. Furthermore, they eliminate the need for complex venue setup and equipment calibration, thereby facilitating home-based assessments [[11], [12], [13], [14], [15]].

Addressing the current scarcity of array pressure datasets for balance assessment in stroke patients, this study introduces a dataset collected using flexible array sensing equipment. The system is designed for seamless deployment in home environments. By employing advanced data analysis and artificial intelligence algorithms for in-depth exploration of the dataset, this work enables an efficient and precise evaluation of patients' balance abilities. This dataset possesses prominent practical and research value. Although it includes only 17 stroke patient samples, the complete motion acquisition process and sufficient valid data volume support the training and validation of detection models. Meanwhile, given the scarcity of high-quality public experimental data in this field, this dataset effectively fills the research data gap and features remarkable scarcity and reusable value.

3. Data Description

In experiments to evaluate balance ability, the array pressure acquisition system records pressure data from different parts of the subject's body during various movements using a flexible plantar pressure sensor and a flexible seat cushion pressure sensor. The raw pressure data collected is stored in .csv files, and the data collection process is shown in Fig. 1.

Fig. 1.

Data collection and processing flow.

After data collection, these raw data are processed using advanced signal processing algorithms to extract key features related to position transitions. Finally, heatmap images are generated and stored in .png format.

This dataset consists of two main parts: raw pressure data and processed heatmap data. The raw pressure data are stored in a .csv file format, containing pressure data in hexadecimal format. They are organized into corresponding subdirectories based on the five experimental movements (sit-to-stand, in-place stepping, static balance with eyes open, static balance with eyes closed, and standing on a soft mat with eyes closed). Each subdirectory holds multiple .csv files that represent specific experimental records. Users can apply various signal processing methods to these raw data according to their research needs to extract particular features and analytical results.

The processed heatmap data can be used directly for training image recognition models, helping to improve the performance and accuracy of balance ability recognition models and making vision-based research more convenient.

3.1. Data file description

As shown in Fig. 2, each participant's data is stored separately in a folder. For example, the folder “3_RawData” contains the data for Participant 3. Inside the “3_RawData” folder, there are two subfolders: “3_PressureData” and “3_HeatmapData,” which hold the raw pressure data and the generated heatmap data, respectively.

Fig. 2.

Dataset file storage format.

Relevant participant information is provided in the “Participant_Information.csv” file, including age, height, weight, gender, and health status (whether they are stroke patients). Meanwhile, based on the Berg Balance Scale, participants' balance ability is divided into four levels: Normal Balance, Mild Balance Impairment, Moderate Balance Impairment, and Severe Balance Impairment. Since patients with Severe Balance Impairment cannot independently complete any movements, no patients in this category are included in this dataset; additionally, due to physical function limitations, some participants with particular balance abilities can only perform part of the five experimental movements.

Stroke patients present substantial clinical challenges including impaired mobility, complex comorbidities, and low recruitment compliance. A sample size of 17 is reasonable given the constraints of research duration, ethical approval, and clinical recruitment capacity. This sample is sufficient for the preliminary exploration, feature observation, and methodological validation of plantar and sitting pressure distribution during functional movements in stroke patients, and aligns with the established criteria for small-sample exploratory data research in the medical domain.

3.2. Data frame description

Each plantar pressure data frame includes a frame header, frame type, length, packet type, data field, and checksum. As shown in Table 1, the frame header data type is uint16_t, with a specific value of 0xA55A. The frame type data type is uint8_t, with a particular value of 0 × 01. The length of the data type is uint16_t, with a particular value equal to the size of the packet minus 2. The packet type data type is uint8_t, with a particular value of 0 × 01. The data field data type is uint8_t, containing 32 × 64 data points. The checksum data type is uint16_t, with a specific value equal to the sum of all other fields excluding itself.

Table 1.

Data frame structure of the plantar pressure sensor.

Data Frame Structure	Data Type	Typical Value
Frame Header	Uint16_t	0XA55A
Frame Type	Uint8_t	0 × 01
Length	Uint16_t	Sizeof(pack)-2
Packet Type	Uint8_t	0 × 01
Data Field [1]	Uint8_t	C1R1
Data Field [2]	……	……
Data Field [3]	Uint8_t	C1R64
Data Field […]	……	……
Data Field [2047]	Uint8_t	C32R63
Data Field [2048]	Uint8_t	C32R64
Checksum	Uint16_t	Excluding Self-Sum

Each seat cushion pressure data frame includes a frame header, frame type, length, packet type, data field, and checksum. As shown in Table 2, the frame header data type is uint16_t with a specific value of 0xA55A. The frame type data type is uint8_t with a particular value of 0 × 01. The length data type is uint16_t, equal to sizeof(packet)-2. The packet type data type is uint8_t with a particular value of 0 × 01. The data field contains 16 × 16 point data, with each point represented by a uint8_t. The checksum data type is uint16_t, calculated as the sum of all other fields excluding itself.

Table 2.

Data frame structure of the seat cushion pressure sensor.

Data Frame Structure	Data Type	Typical Value
Frame Header	Uint16_t	0XA55A
Frame Type	Uint8_t	0 × 01
Length	Uint16_t	Sizeof(pack)-2
Packet Type	Uint8_t	0 × 01
Data Field [1]	Uint8_t	[AX0][AY0]
Data Field [2]	……	……
Data Field […]	Uint8_t	[AX0][AY15]
Data Field [254]	……	……
Data Field [255]	Uint8_t	[AX15][AY14]
Data Field [256]	Uint8_t	[AX15][AY15]
Checksum	Uint16_t	Excluding Self-Sum

The raw pressure data stored in CSV format includes timestamps and pressure point data. As shown in Table 3, the timestamp indicates the exact time when each frame of pressure data is collected and saved. Following the timestamp, the complete pressure data for one frame is recorded, with its specific format detailed in Table 1, Table 2.

Table 3.

Storage format of raw pressure data in CSV.

timestamp	frame_data
2025/4/14 11:09:00	A5 5A 01 06 08 01 00 00 00 00 00 00 01 00 00 00 00 00 01 00 00 00 00 01 01 01 00 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 01 00……
2025/4/14 11:09:00	……
2025/4/14 11:09:00	……
2025/4/14 11:09:00	……
2025/4/14 11:09:00	……
2025/4/14 11:09:00	……
2025/4/14 11:09:01	……
……	……
……	……

4. Experimental Design, Materials and Methods

4.1. Data collection process

As shown in Fig. 3, the seat cushion pressure sensor is placed on the chair, and the plantar pressure sensors are positioned on the floor in front of the chair. Before the experiment begins, subjects receive detailed instructions to fully understand the process and the required movements. At the start of the experiment, the subject sits upright on the chair and places their feet flat on the plantar pressure sensors naturally. This standardized initial posture provides a consistent reference point for subsequent posture changes, ensuring the data collected is uniform and comparable.

Fig. 3.

Data collection scenario.

As illustrated in Fig. 4, the subjects perform five specific movements in order: sit-to-stand, in-place stepping, stepping on a soft mat with eyes closed, static balance standing with eyes open, and static balance standing with eyes closed. Between each movement, the volunteers rest for 10 to 15 seconds to allow their bodies to regain a stable state.

Fig. 4.

Five experimental movements: sit-to-stand, in-place stepping, eyes-closed soft mat stepping, static balance with eyes open, and static balance with eyes closed.

4.2. Experimental materials

The equipment employed in this experiment was manufactured by MATRIX INNOVATION, China. The flexible pressure sensor employed in this study is prepared based on percolation materials and belongs to the thin-film pressure sensor category. Its main body adopts a five-layer composite structure design, as illustrated in Fig. 5. From bottom to top, the layers are sequentially: the first outer fabric layer, the second Y-electrode layer (composited of white fabric, conductive fabric, and screen-printed circuits on PET film), the third piezoresistive film layer, the fourth X-electrode layer (with the same structure as the Y-electrode layer, composited of white fabric, conductive fabric, and screen-printed circuits on PET film), and the fifth outer fabric layer.

Fig. 5.

Internal structure of the flexible pressure sensor.

The core working mechanism of the sensor is pressure-resistance signal conversion. When pressure is applied to the surface of the sensor film, the pressure signal can be converted into a change in the resistance value of the sensor, thereby enabling the detection and acquisition of pressure information. It possesses distinct piezoresistive properties with a negative correlation between pressure and resistance: the higher the pressure, the lower the resistance value. The calibration formulas are shown in Eqs. (1) and (2), which lays a theoretical foundation for the accurate quantification of pressure data.

Seat Cushion Pressure Calibration Fitting Formula:

$$\{\begin{array}{c}y=3\bullet x\left(0\le x\le 18\right)\hfill \\ y=8·5x-99\left(18<x\le 48\right)\hfill \\ y=4.2\bullet x+107.4\left(48<x\le 75\right)\hfill \\ y=9.2\bullet x-267.6\left(75<x\le 100\right)\hfill \end{array}$$ (1)

Plantar Pressure Calibration Fitting Formula:

$$\{\begin{array}{c}y=3.27\bullet x\left(0\le x\le 20\right)\hfill \\ y=7.16\bullet x-77.8\left(20<x\le 45\right)\hfill \\ y=12.036\bullet x-297.22\left(45<x\le 68\right)\hfill \\ y=13.541\bullet x-399.56\left(68<x\le 88\right)\hfill \\ y=8\bullet x+88.048\left(88<x\le 100\right)\hfill \end{array}$$ (2)

Where $x$ denotes the sampled value, and $y$ denotes the corresponding value in F (N). The corresponding calibration curve is shown in Fig. 6.

Fig. 6.

Calibration curve.

As shown in Fig. 7, the flexible plantar pressure sensor has external dimensions of 520 mm by 493 mm. Its effective pressure-sensing area measures 319.3 mm by 448.7 mm, with 2048 pressure-sensing points (32 by 64). The standard data transmission rate is 20 Hz, and the baud rate is 460800. The flexible seat cushion pressure sensor has outside dimensions of 430 mm by 420 mm; a pressure-sensing area of 410 mm by 400 mm; 256 pressure-sensing points (16 by 16); a standard data transmission rate of 10 Hz; and a baud rate of 115200.

Fig. 7.

Plantar pressure sensor and seat cushion pressure sensor.

4.3. Data preprocessing methods

After collecting the raw pressure data, researchers can process it to generate the heatmaps they need. For the plantar pressure sensor, after removing the frame header, frame type, length, packet type, and checksum from a raw pressure data frame, the remaining data shown in Fig. 8 is the pressure data in the data field. This pressure data from one data field is then converted into a 32 × 64 data matrix to be analysed.

Fig. 8.

Raw pressure data frame.

For the flexible seat cushion pressure sensor, similarly, after removing the frame header, frame type, length, packet type, and checksum from a frame of raw pressure data, the remaining data is the pressure data in the data field. This pressure data from one data field is adjusted into a 16 × 16 data matrix, which can then be analysed.

The relative pressure values collected by the sensor are assembled into a corresponding 32 × 64 matrix, and the corresponding heatmap is generated, as illustrated in Fig. 9. The values ranging from 0 to 64 on the horizontal axis and from 0 to 32 on the vertical axis correspond to the 32 × 64 sampling lattice of the plantar pressure sensor, respectively. The term Pressure Intensity denotes the raw relative pressure values acquired through sampling.

Fig. 9.

Plantar pressure heatmap.

Researchers can utilize this dataset to extract raw pressure-time data from the plantar pressure dataset and perform data cleaning. Initially, missing data is filled using methods such as neighbourhood averaging or interpolation. Since the raw pressure signal may contain high-frequency noise, like slight friction between the foot and the pad, it needs to be smoothed with low-pass filters such as moving average filtering or Butterworth filtering to preserve the true dynamic trend of the Centre of Pressure (COP).

The COP refers to the resultant force point of pressure at all contact points on the plantar surface, and the principle of its coordinate calculation is based on the “moment balance of forces”: as shown in Eqs. (1) and (2), the pressure value$\left(p_{i,j}\right)$ of each sensor is used as a weight to calculate the weighted average of its coordinates ($x_i,y_i$), thereby obtaining the COP coordinates ($CO{P}_x$, $CO{P}_y$) at that moment. Each sampling time t corresponds to a unique set of COP coordinates ($CO{P}_x\left(t\right)$, $CO{P}_y\left(t\right)$).

As shown in Fig. 10, based on the plantar pressure data collected during static standing, the centre of pressure points was extracted and sequentially connected to form a COP trajectory plot. The horizontal axis represents the physical displacement range of the COP in the lateral direction (X-axis) of the plantar pressure sensor, with an effective measurement range of 448.7 mm in this direction; the actual distribution range of the COP trajectory on the horizontal axis is from 108 mm to 113 mm. The vertical axis represents the physical displacement range of the COP in the anterior-posterior direction (Y-axis) of the plantar pressure sensor, with an effective measurement range of 319.3 mm in this direction; the actual distribution range of the COP trajectory on the vertical axis is from 188.4 mm to 189.4 mm.

$$CO{P}_x=\frac{{\sum }_{i=1}^M{\sum }_{j=1}^N\left(p_{i,j}\times x_{i,j}\right)}{{\sum }_{i=1}^M{\sum }_{j=1}^N{p}_{i,j}}$$ (3)

$$CO{P}_y=\frac{{\sum }_{i=1}^M{\sum }_{j=1}^N\left(p_{i,j}\times y_{i,j}\right)}{{\sum }_{i=1}^M{\sum }_{j=1}^N{p}_{i,j}}$$ (4)

Fig. 10.

COP path curve.

By analysing the morphology, position, and motion features of the COP curve, researchers can make an objective assessment of human motor function and balance ability.

Researchers can combine pressure data from the flexible plantar pressure sensor and the flexible seat cushion pressure sensor in this dataset to generate time-varying curves. As shown in Fig. 11, this graph enables the analysis of synchronized variation curves of the subject’s pressure data in two dimensions (plantar and hip) and the process of pressure value changes in the pads during the subject’s sit-to-stand transition.

Fig. 11.

Synchronized variation curves of plantar pressure and seat cushion pressure.

Researchers can directly identify the gait cycle using the dataset's heatmaps or segment it with raw pressure data based on the “pressure peak variations” of the plantar pressure distribution. One gait cycle is divided into three main phases, each characterized by a specific pressure pattern.

• •Heel Strike Phase: The heel initially touches the ground, with pressure focused on the heel. Pressure values quickly increase to a small initial peak.
• •Stance Phase: The entire plantar surface contacts the ground, with pressure shifting from the heel to the midfoot and forefoot, resulting in a full-plantar pressure distribution. In healthy individuals, moderate pressure is observed in the midfoot and low pressure in the arch area during this phase.
• •Toe-Off Phase: The heel lifts off the ground, pressure concentrates in the forefoot and toes, and pressure values gradually decrease to zero as the toes push off the ground.

As shown in Fig. 12, for the left foot: phase (a) is the Stance Phase; phases (b), (c), and (d) are the Toe-Off Phase; phases (e) and (f) are the Heel Strike Phase; and phases (g), (h), and (i) are the Stance Phase. Additionally, researchers can identify the gait cycle phase for each pressure data frame using the corresponding pressure heatmaps. After extracting features with deep learning and completing training, they can directly segment the gait cycle using raw pressure data.

Fig. 12.

Gait cycle heatmap.

Limitations

These two sensors exhibit notable limitations in dynamic response and environmental tolerance: the seat cushion pressure sensor features a 16 × 16 array with a frame rate of 10 Hz, while the plantar pressure sensor offers a 32 × 64 array with a frame rate of 20 Hz, which may be insufficient for applications involving dynamic or high-frequency pressure changes. Additionally, their operating temperature ranges from –20°C to 60°C, with the resistance increases sharply above 50°C, potentially exceeding the circuit’s measurement range and leading to failure. In terms of mechanical durability, the plantar pressure sensor should not be bent beyond 90°, whereas the seat cushion pressure sensor imposes specific constraints on the distribution of its sensing area, requiring that pressure-sensitive regions not be positioned at the centre of large-deflection deformation to avoid compromising measurement performance or causing irreversible damage.

Ethics Statement

The authors confirm that all participants gave informed consent before data collection. No personal identification information was recorded, and all data were fully anonymized to safeguard the participants' privacy. This study complies with ethical research standards and has received formal approval from the Ethics Committee of Jiangbin Hospital of Guangxi Zhuang Autonomous Region (Ethics-KY2025-002).

Credit Author Statement

Shun Wang: Data collection, Data processing, Writing. Jian Guo: Writing-Review & Editing. Yue Zhang: Methodology, Data collection. Huachun Liu: Supervision. Huabin Su: Project administration. Yunpeng Zhang: Conceptualization, Manuscript proofreading.

Data Availability

• Mendeley DataStroke Patients’ Balance Ability Test Dataset Utilizing Array Pressure Acquisition System (Original data)