Regarding the development of sensing device, we first decided the design requirements of the device. The previous devices had measurement circuits outside the device. This limits their swing range and portability. To overcome these limitations, we developed a new cylindrical sensing device. It could contain the measurement circuit (Figure 1). The device height and diameter were 85 mm and 40 mm, respectively. The sensor consists of five layers: protective, driving, insulating, detection, and electrode layers (Figure 2a). The protective layer was added because the children torn through the drive layer during the preliminary experiments. The protective layer consisted of a 1-mm-thick yellow felt cloth. The driving layer consisted of a conductive silicone sheet (EC-20BH, Shin-Etsu Chemical Co. Ltd.). This layer was connected to a 3.3 V DC voltage source. The DC voltage source used a 3.3 V pin microcontroller (ESP32-DevKitC, Espressif Systems). The insulating layer was a glass fiber sheet (13-7127, KLASS). The detection layer was composed of a conductive sheet (ZC-85, ENGINEER) with a surface resistance of 10 kΩ/sq. When the driving layer contacts with the detection layer, the electrical circuit is closed. As a result, current flows to the electrodes through the detection layer (Figure 2b). The electrode and detection layer were bonded using a conductive epoxy (CW2400, CircuitWorks). The electrode layer was an original flexible printed circuit board with 16 electrodes and a diameter of 2 mm. One of these electrodes was used as the ground condition and the other was used as the measurement electrode. This operation was repeated until all electrodes were used in all conditions. Thus, 256 voltage data points (16 grounding conditions × 16 electrodes) are obtained. The multiplexer (MUX) used was CD74HC4067 (Texas Instruments). The measurement period was 0.25 s per frame.

A tomographic tactile sensor requires a solver to reconstruct pressure distribution from the measured voltage vector. This solver addresses an ill-posed problem because the output dimension (reconstructed image) is larger than the input dimension (measured voltage vector). Therefore, we used the linear reconstruction method of Tikhonov regularization with two-dimensional finite element method (FEM) model, based on our previous studies (Asahi et al., 2024b; Yoshimoto et al., 2020). The hyperparameter of the Tikhonov regularization was set to 5000. These reconstruction processes were performed using MATLAB 2023b (MathWorks Inc.).

In this study, participants were required to meet two criteria: (1) the ability to perform both power grip and precision grip, and (2) an age at which they could sufficiently understand verbal instructions. Regarding (1) the ability to perform power grip and precision grip, as mentioned in the introduction, primitive grasping has been reported to emerge between 2 and 22 weeks, while precision grip develops at 5.97 months. Regarding (2) the ability to sufficiently understand verbal instructions, B. Buckley have reported that three-year-old children are capable of communicating using language. Additionally, four-year-old children can focus on and follow verbal instructions even without explicit cues, such as being called by name (Buckley, 2003). Based on these considerations, this study targeted four-year-old children as participants. The participants of this study were 17 children (4.43 ± 0.30 years old, 8 boys, 9 girls). In order to ensure that the participants had adequate communication skills and no developmental disorders, we administered the KINDER INFANT DEVELOPMENT SCALE questionnaire type C prior to the experiment (Hassanein, 1982). None of the participants had any serious disease or disorder. In the analysis, 11 children (5 boys and 6 girls) were included, excluding those who stopped the measurement halfway through because they could not listen to the experimenter’s instructions or did not want to participate (4 participants), those whose actual dominant hand seemed to differ from that reported by their parents (1 participant), and those who grasped without placing their palm on the object during the power grip (1 participant). Evaluation of the modified Japanese version of the FLANDERS handedness questionnaire (Okubo et al., 2014) indicated 10 right-handed children and one left-handed child.

This study was approved by the Ethics Committees of Shibaura Institute of Technology and Keio University. The experiment was conducted only when informed consent was obtained from the parents of the participating children.

The participants practiced freely grasping the device, without external assistance. Voltage measurements were also performed during the hardware and software testing. The participants were then instructed to perform either a power grip or precision grip, with the order being counterbalanced. Each grip was measured ten times. However, if the hand was released during the measurement process or if the grasp force was not applied (i.e., the object was grasped only by the frictional force of the fingers), the grasp was excluded from the analysis, and an additional measurement was performed. Ten times per grasp were measured, that is, 100 measurement frames (10 measurements × 10 times) were obtained for each grasping category.

The measured voltage vectors and reconstructed images were used as features to classify the power grip and precision grip. For classification using the measured voltage vectors, 256 data points were used (16 electrodes × 16 measurements). For classification using the reconstructed images, the FEM values normalized from 0% to 100% were used. The size of FEM model was 43 × 66 × 1.

In this study, we considered more practical applications, such as educational toys or diagnostic systems for developmental disabilities. For such applications, we conducted a comparative analysis using the k-Nearest Neighbors (KNN) algorithm (Bansal et al., 2022), which is computationally efficient and easy to implement, and the Convolutional Neural Network (CNN), which has been reported to achieve high classification accuracy (Park et al., 2021). In KNN, we used the MATLAB function (fitcknn) with a k value of 1. The network architecture of the CNN consisted of 17 layers, as listed in Table 2. The input layer was configured to input each feature with dimensions of 16 × 16 × 1 for classification using the measured voltage vectors, and 43 × 66 × 1 for classification using the reconstructed images. Subsequently, a three-step convolution was performed. The convolution layers were organized with filter sizes of 3 × 3 × 32, 3 × 3 × 64, and 3 × 3 × 128, in that order. In each convolution layer, the ReLU was applied as the activation function after batch normalization. In steps one and two, a 2 × 2 max pooling layer is utilized, resulting in downsampling with a stride of two. In step three, a fully connected layer was used to classify the data into two classes. The Softmax function was applied to the output layer, resulting in a final classification into two classes: power grip and precision grip. Stochastic gradient descent was employed for training with an initial learning rate of 0.001. In addition, the learning rate was configured to be reduced by 95% after ten epochs. The maximum number of epochs was set to 36, and the data were randomized at the beginning of each epoch.

Cross-validation was performed to validate the classification accuracy of these two types of features and classification methods. Verification was performed for each participant, with one grasp (10 frames) as the test data and the remaining grasps (190 frames) as the training data, and was repeated until all grasps were the test data.

For the evaluation of classification results, the following accuracy was calculated for each participant as Equation 1: A c c u r a c y = T P o w + T P r e   T P o w + T P r e + F P o w + F P r e (1) where, initial character T or F indicates whether the class predicted by the classifier matches the true class or not. The characters Pow and Pre indicate power grip and precision grip, respectively. Thus, T _Pow is the matching case of the power grip as predicted class by the classifier and the power grip as true class. Subsequently, to evaluate the classification performance in different grasps, a confusion matrix was calculated for each grasp. Precision (Prec), recall, and F-measure were then calculated for each grasp as follow Equations 2–4: P r e c   P o w   o r   P r e = T P o w   o r   P r e   T P o w   o r   P r e + F P o w   o r   P r e (2) R e c a l l   P o w   o r   P r e = T P o w   o r   P r e   T P o w   o r   P r e + F P r e   o r   P o w (3) F − m e a s u r e   P o w   o r   P r e = 2 × P r e c × R e c a l l P r e c + R e c a l l (4)

These index values ranged from 0 to 1. In addition, t-Distributed Stochastic Neighbor Embedding (t-SNE) has been used to confirm the distribution of features and clustering trends (Van Der Maaten and Hinton, 2008). The classification features were compressed into a two-dimension map by t-SNE. The distances between points in the t-SNE plot reflect similarity relationships in the original high-dimensional space. Additionally, clearly separated clusters indicate natural groupings based on differences in classification features.