A deep learning method for children's self-care problems classification using represent learning and focal loss

Abstract

Background

An accurate diagnosis of children's self-care problems significantly matters in the growth and development of children. However, various and extensive disorders make the self-care problems classification extremely complex and require much effort and time to solve.

Objective

To deal with the above challenge, a deep learning model is proposed to classify the children's self-care problems intelligently and precisely.

Method

The proposed deep learning model contains two sub-deep neural networks. The first sub-network employs a technology of representing learning named triplet loss. It aims to compress the dimensions of the feature of the children with self-care problems to extract the useful information and exclude the noise, in order to improve classification performance. The second sub-network utilizes a technology for handling the class imbalance problem called focal loss to further improve the classification accuracy.

Result

The experimental results show that the proposed deep learning model outperforms. The averages of accuracy, precision, recall, and F1 score can achieve 99.78%, 0.99, 0.99, and 0.99, respectively.

Conclusion

To the best of our knowledge, the proposed method achieves state-of-the-art results. That can significantly support the rehabilitation and growth of children with self-care issues. Furthermore, this study also provides a demonstration and experience of the application of the deep learning model in the healthcare field.

Keywords

deep neural network represent learning triplet loss class imbalance problem focal loss

1 Introduction

Self-care refers to the behaviors and activities a person undertakes to maintain and improve their physical, mental, and emotional wellbeing.^1,2 Self-care is able to help individuals increase their self- awareness, self-esteem, confidence, and independence. It not only helps to cope with the challenges and stresses of life, but also presents and alleviates a range of lifestyle-related disorders such as anxiety, depression, and stress.³ Self-care is not a one-time act, but requires consistent and determined effort, and becomes a part of everyday life.^4,5

For children and youth with certain functioning or disabilities, self-care is important for their physical and mental developments.^6,7 An accurate assessment of children's self-care problems plays an essential role in the subsequent occupational therapy and professional medical services.⁸ Therefore, a conceptual framework named International Classification of Functioning, Disability, and Health for Children and Youth (ICF-CY) has been developed and released by the World Health Organization (WHO), in order to help evaluate, assess, and classify the corresponding functioning, disability, and health problems of children.⁹ In accordance with the ICF-CY, WHO aims to help the healthy growth of the child physically and mentally.^10–13

However, due to the complexity of disability itself^14,15 and the extensive varieties of disorders,^16,17 it is a hard, labor-intensive, and time-consuming work to accurately assess the children's disabilities and disorders.¹⁸ Besides, the diagnosis of the corresponding self-care problem is also a challenge. A vast wealth of corresponding knowledge, experience, and collaboration among physicians, therapists, special educators, neuropsychologists, and other professionals in relevant fields would be required to conduct a precise diagnosis.

Therefore, in this paper, a deep learning model is proposed to accurately and effectively diagnose and classify children's self-care problems, in order to determine a corresponding appropriate occupational therapy and professional medical services to guarantee the children grow healthily.

1.1 Related work

The issue of self-care was initially studied as early as 1980. Twenty-six women were recruited and kept health diary, which consisted of a structured sheet to record health upsets, for four weeks, in order to conduct an exploratory study of self-care.¹⁹ It indicated that self-care was necessary on more than 80 percent of days when medical issues were present. Moreover, the women reported that numerous “nonmedical” actions or events provided therapeutic benefits. The study not only confirmed that the self-care significantly impacted on both physical and mental well-being, but also highlighted that the self-care necessitated a holistic assessment of intricate interplay among social, psychological, and medical elements.

Then, different from the qualitative study way, such as “diary”¹⁹ and “interview”,²⁰ the statistics method is employed to quantitatively analyze the self-care problem and conduct further research.^21,22 Given the children's abilities of manual dexterity and gross motor functioning, the development of self-care and mobility skills in children with cerebral palsy was examined by Ohrvall AM et al..²³ The data of 195 children diagnosed with cerebral palsy were collected and utilized. The collected data were evaluated using the Pediatric Evaluation of Disability Inventory's self-care and mobility functional skill scales, according to the Manual Ability Classification System (MACS) and the Gross Motor Function Classification System (GMFCS). Through the analysis of stepwise multiple regression, it can be found that the most influential predictor of self-care skills was MACS, and the most influential predictor of mobility skills was GMFCS. Besides, it also observed that children classified as MACS level I or II showed a significant correlation between age and self-care ability.

With the development of modeling technology, data size, and computing power, automatic expert systems based on data science are increasingly prevalent in medical and healthcare research.^24–33 Meanwhile, these methods have also been applied to the study of children's self-care problems. Zarchi MS et al. innovatively introduced SCADI (Self-Care Activities Dataset based on ICF-CY), which was a new standard dataset for children's self-care problems analysis, to contribute to the children's self-care research.³⁴ To comprehensively study children's self-care issues, the SCADI contains the corresponding information as much as possible, and 29 self-care activities are included. In order to assess the applicability and universality of the SCADI, Zarchi MS et al. employed two distinct types of expert systems for classifying self-care problems among children who had physical and motor disabilities. In the first expert system, an Artificial Neural Network (ANN) was used as a classifier and achieved an accuracy of 83.1% in classifying children's self-care problems. In the second expert system, a Decision Tree (DT) algorithm known as C4.5 was employed to accurately extract classification rules for children's self-care problems. Zdrodowska M et al. considered the features contained in the SCADI and their impacts on the accuracy of children's self-care problems classification.³⁵ The correlation method, information gain method, and Chi-Square test-based method were employed to find and select the features that play primary roles in categorizing a child into a specific group with certain self-care problems. Then, the JRip, PART, C4.5, Random Tree (RT), and CART were used to build the classification model. The experiments showed that the accuracy of JRip increased by 4.28% with the 17 selected features, while the performances of other methods all decreased. The motivation for using the above method was to leverage a few key features to achieve satisfactory classification performance for children's self-care problems while keeping computational costs relatively low. Within the same conceptual framework, Choudhury A et al. used the Boruta algorithm to select the useful features, then utilized Random Forest (RF), Support Vector Machine (SVM), Naive Bayes (NB), and Hoeffding Tree (HF) to classify the children's self-care problems based on the selected features.³⁶ The experiment suggested that RF achieved the best performance with 84.75% accuracy. Fatemi et al. introduced a classification model for children's self-care problems that the Genetic Algorithm (GA) was applied for selecting features, and the Probabilistic Neural Network (PNN) was employed as a classifier.³⁷ The experiments demonstrated that the proposed model could classify children's self-care problems with 94.28% accuracy using the selected features across 34 dimensions. Syafrudin M et al. proposed an effective classification model for children's self-care problems by leveraging the Principal Component Analysis (PCA) to extract important features and the Decision Trees (DT) to construct the corresponding classification model.³⁸ The experiments demonstrated that PCA-based feature extraction yielded positive results, enhancing the model's performance by achieving 1.70% increase in accuracy. The proposed classification model achieved 94.29% accuracy. In a recent study, Putatunda S introduced a novel deep learning-based model, which was named Care2Vec, to classify the corresponding self-care problems of physically disabled children.³⁹ In Care2Vec, autoencoders and deep neural networks were employed to construct a two-step modeling process. Firstly, the autoencoders were used to learn the low-dimensional representation of the feature of the self-care problem. Subsequently, the learned low-dimensional representation of the feature was used as the input data that were fed into the deep neural networks to complete the training procedure. The experiments showed that the proposed Care2Vec can achieve 91.43% accuracy in the children's self-care problems classification.

1.2 Review of literature

Based on the above review, it can be seen that self-care contributes significantly to improving physical, mental, and emotional well-being, especially in promoting the healthy growth of children with certain functioning and disabilities. An accurate self-care diagnosis and classification can significantly contribute to the subsequent occupational therapy and professional medical services.

However, diagnosing and classifying self-care problems is a complex challenge due to the diverse range of disorders and disabilities. In practice, achieving a precise diagnosis of self-care problems often necessitates a broad range of knowledge, experience, and collaboration among physicians, therapists, and other related professionals. As a result, numerous studies are being conducted to automatically and accurately diagnose and classify self-care problems. Firstly, qualitative study methods, such as “diary” and “interview”, are used. Then, to obtain a more quantitative description of the self-care problem, the statistics method is employed. Nowadays, modeling methods, computational technology, and deep learning methods with increasing computing power have made great progress and achieved impressive performances in the expert system, image processing, and natural language processing fields. To utilize the outstanding ability of these new emerging methods, researchers have introduced them to the field of self-care.

As far as we know, the best performance of children's self-care problems classification achieves 94.29% accuracy.³⁸ The study treats the self-care problem classification as a two-step modeling process. The first step is to use the PCA to select the most related features. The second step is to utilize the selected feature to categorize the children's self-care problems using the DT method. The advantage of the two-step modeling process is that the feature selection procedure is able to make the classifier only focus on the most related factors and exclude other interfering factors which could lead to a decrease in accuracy and efficiency.⁴⁰ As a result, studies that employ the feature selection procedure achieve rather good performances. Such as Bushehri and Zarchi,³⁷ Syafrudin et al.,³⁸ and Putatunda³⁹ achieve 94.28%, 94.29%, and 91.43% accuracy of children's self-care problems classification, respectively.

Despite achieving rather good performances, the information among feature dimensions and the correlation between different instances are neglected. Fatemi BS et al. and Syafrudin M et al. use GA³⁷ and PCA³⁸ to select the related features, respectively. However, they all neglect the information among the feature dimensions. Putatunda S realizes the information among the feature dimensions and uses the autoencoders to compress the dimensions of the feature instead of only selecting certain dimensions of the feature.³⁹ But the correlation between different instances represented by features is still not considered. All these factors can result in the degeneration of the classification performance of children's self-care problems. Furthermore, the class imbalance problem in children's self-care problems classification, which may lead to a decrease in classification accuracy, has not been considered.

1.3 Contributions

In order to deal with the above problems (i.e., considering the information among the feature dimensions and the correlation between the different instances, as well as the class imbalance problem), we propose an artificial neural network model, which is based on deep learning, represent learning, and focal loss technology, to furthermore improve the accuracy and efficiency of the classification performance of the children's self-care problems. The contributions are as follows:

1)
The classification of children's self-care problems is treated as a two-step modeling process and the model is consisted of two sub-deep neural networks.
2)
The first sub-deep neural network utilizes the triplet loss method to compress the dimensions of the feature that is to exclude the noise, preserve information among different dimensions, and maintain the correlation between different instances, simultaneously.
3)
The second sub-network employs a technology called focal loss to deal with the class imbalance problem to further improve the classification performance of children's self-care problems.

The remaining paper is structured as follows: In Section 2, the proposed method is presented. In Section 3, numerical tests are performed. Section 4 gives discussions about the proposed method. In Section 5, limitations of the proposed method are presented. Section 6 summarizes the paper.
2 A two-step modeling process of self-care classification

In this paper, the classification of children's self-care problems is viewed as a two-step modeling procedure, which contains two sub-deep neural networks. The first sub-network employs triplet loss to enhance feature separability by enforcing intra-class compactness and inter-class distinctiveness, while the second sub-network adopts focal loss to mitigate class imbalance by focusing on hard and minority-class samples. The integration of these two complementary loss functions leads to improved classification performance. The overall structure diagram is presented in Figure 1.

Figure 1.

The overall structure of the proposed children's self-care problems classification method.

2.1 Basic method: deep neural network

The complete proposed method for classifying children's self-care problems is based on the deep neural network. It is one of the most popular artificial methods these days and achieves impressive performances in image processing and natural language processing. For example, the amazing stable diffusion model, the segment anything model, and the ChatGPT are all based on the deep neural network. Because of the outstanding modeling ability of the deep neural network, we have applied it to our children's self-care problems classification model. The basics of the deep neural network are presented in Figure 2. The deep neural network contains three parts which are the input layer, hidden layers, and output layer. The approach involves using the probability theory to model the relationship of the inputs denoted as $x = [x_{1}, x_{2}, \dots, x_{n}]$ between the outputs referred to as $y = [y_{1}, y_{2}, \dots, y_{n}]$ . The relationship can be represented by the probability distributions represented as $P (y | x)$ . Besides, the hidden layers are responsible for extracting the abstract feature which is named as representing⁴¹ or embedding.⁴²

Figure 2.

The structure of the deep neural network.

The deep neural network is a kind of supervised machine learning method and can be trained by the backpropagation method that the weight parameter w is updated by the stochastic gradient descent method according to the loss function L as

\begin{aligned} w (t + 1) = w (t) - η \frac{\partial L}{\partial w} \end{aligned}

(1)

Where $η$ represents the learning rate. L denotes the loss function which can be determined according to the type of learning problem (such as classification and regression). As an illustration, for the supervised learning of the multi-class classification problem, the usual choice of loss function is the cross-entropy function. And the activation function is typically set to be ReLU.⁴³ More details of the deep neural network can be seen in the reference.⁴⁴

2.2 The first sub-deep neural network

The first sub-deep neural network is responsible for compressing the feature dimension to obtain its corresponding representation also named embedding. The purpose is to help the classifier pay more attention to the important useful information contained in the feature to further improve the classification accuracy as well as the computational efficiency.

Besides, in order to utilize the correlation between different instances, the triplet loss technology⁴⁵ is used. The main idea is to learn the map from the original feature to the embedding in a compact Euclidean space where distances measure the similarity. The flowchart is presented in Figure 3.

Figure 3.

The flowchart of the first sub-deep neural network. The triplet loss is employed to generate the embeddings considering the correlation between different instances.

The objective of triplet loss is to ensure that instances with the same label have their embeddings placed closely together in the embedding space, whereas instances with different labels have their embeddings positioned far apart. To achieve this objective, the loss function can be defined using triplets of embeddings, consisting of an anchor, a positive instance from the same class, and a negative instance from a different class. The triplet loss function can be represented as

\begin{aligned} L_{t r i p l e t} = m a x (d (a, p) - d (a, n) + m a r g i n, 0) \end{aligned}

(2)

In the context of the triplet loss function, the distance between the anchor and the positive instance in the embedding space is represented by $d (a, p)$ , while the distance between the anchor and the negative instance is denoted as $d (a, n)$ . The parameter $m a r g i n$ , which can be adjusted based on specific requirements, ensures the negative instance should be farther away than the positive instance to the anchor. It can be seen that, when we minimize the triplet loss, the goal is to minimize $d (a, p)$ to approach 0 and ensure that $d (a, n)$ is greater than $d (a, p) + m a r g i n$ . This process encourages the positive instances to be closer to the anchor and the negative instances to be sufficiently distant from the anchor in the embedding space.

2.3 The second sub-deep neural network

The second sub-deep neural network is used for classifying the children's self-care problems utilizing the embedding of the feature. Considering the class imbalance problem in the children's self-care problems classification that would result in an accuracy decrease, the focal loss method⁴⁶ is used. The class imbalance problem arises from the disparity in the number of instances across different classes in the dataset. And it can lead to a decrease in the performance of the neural network. To handle the above problem, we combine the focal loss technique with the softmax cross-entropy function.

In the training process, the multiclass classification loss function based on softmax cross-entropy is as

\begin{aligned} L_{C E} (y, \hat{p}) = - l o g ({\hat{p}}_{y}) \end{aligned}

(3)

Where $y \in {0, \dots, K - 1}$ is an integer label and K denotes the number of classes. $\hat{p} = ({\hat{p}}_{0}, \dots, {\hat{p}}_{K - 1}) \in [0, 1]^{K}$ is a vector denoting an estimated probability distribution over the K classes. ${\hat{p}}_{y}$ is the predicted probability of the true class y by the neural network. As an illustration, considering a situation that one class (for instance, class A) has more instances than another class (for instance, class B), it can be seen that the instances of class A would make more contribution than the instances of class B to the overall loss function. As a result, the neural network is trained mostly by the instances of class A. That may cause the degradation of classification performance.

The focal loss is used to handle the above problem. The main idea of focal loss is to reduce the loss contributed by a certain class (such as class A) with a large number of instances, while emphasizing the impact of the loss from another class (like class B) with a small number of instances, simultaneously. The specific approach is adding a factor $(1 - {\hat{p}}_{y})^{γ}$ to the corresponding loss. The focal loss is as

\begin{aligned} L_{F L} (y, \hat{p}) = - (1 - {\hat{p}}_{y})^{γ} \log ({\hat{p}}_{y}) \end{aligned}

(4)

where

γ

represents the focusing parameter that specifies how much the class with many instances (like class A) and the class with few instances (like class B) contributes to the overall loss, respectively. For the class with many instances, the value of

(1 - {\hat{p}}_{y})^{γ}

tends to 0 (because

{\hat{p}}_{y}

tends to 1). On the contrary, for the class with few instances, the value of

(1 - {\hat{p}}_{y})^{γ}

is not much affected (because

{\hat{p}}_{y}

tends to 0). Therefore, the higher the

γ

, the contribution of the class with many instances to the overall loss is more down-weighted, while the contribution of the class with few instances is more high-weighted. Figure 4 shows the details of the focal loss.

Figure 4.

The explanation of the second sub-deep neural network. The focal loss is employed to handle the class imbalance problem by adding a weighted factor $(1 - {\hat{p}}_{y})^{γ}$ to adjust the contributions of the different classes to the overall loss function.

Additionally, the algorithm of the proposed method for the children's self-care problems classification is presented in Algorithm 1.

3 Numerical tests

The proposed method for the classification of children's self-care problems is tested on the SCADI dataset which is generated from the ICF-CY framework published by the WHO. Compared with up-to-date references, the experimental tests show that the proposed method outperforms.

3.1 Datasets

The SCADI dataset includes data with 205 features of 70 children who have various physical and motor disabilities. There is a roughly equal sex distribution, with 29 boys (41.4%) and 31 girls (58.6%). The total median (interquartile range) age of the 70 children is 12 (9–15) years old. For boys, the median (interquartile range) age is 10 (9–14) years old. For girls, the median (interquartile range) age of the girls is 13 (10–15) years old. 29 activities related to the physical and motor disabilities of the children's self-care problems are considered in the SCADI dataset. These activities with the corresponding descriptions and feature codes are shown in Table 1. Table 2 presents all the feature codes used in the SCADI dataset. The feature codes are defined according to the ICF framework.⁹ Furthermore, the children's self- care problems are categorized into 7 distinct classes. Class 1: Caring for body parts problem; Class 2: Toileting problem; Class 3: Dressing problem; Class 4: Washing oneself and caring for body parts and dressing problem; Class 5: Washing oneself, caring for body parts, toileting, and dressing problem; Class 6: Eating, drinking, washing oneself, caring for body parts, toileting, dressing, looking after one's health, and looking after one's safety problem; Class 7: No self-care problem. These classes are also presented in Table 3. More details of the SCADI dataset can be found in the reference.³⁴

Table 1.
Self-care activities considered in the SCADI.

No. Self-care category Activity no. Description Codes

1 Washing oneself 1 Washing body parts d 5100

2 Washing whole body d 5101

3 Drying oneself d 5102

2 Caring for body parts 4 Caring for skin d 5200

5 Caring for teeth d 5201

6 Caring for hair d 5202

7 Caring for fingernails d 5203

8 Caring for toenails d 5204

9 Caring for nose d 5205

3 Toileting 10 Indicating need for urination d 53000

11 Carrying out urination appropriately d 53001

12 Indicating need for defecation d 53010

13 Carrying out defecation appropriately d 53011

14 Menstrual care d 5302

4 Dressing 15 Putting on clothes d 5400

16 Taking off clothes d 5401

17 Putting on footwear d 5402

18 Taking off footwear d 5403

19 Choosing appropriate clothing d 5404

5 Eating 20 Indicating need for eating d 5500

21 Carrying out eating appropriately d 5501

6 Drinking 22 Indicating need for drinking d 5600

23 Carrying out drinking appropriately d 5602

7 Looking after one's health 24 Ensuring one's physical comfort d 5700

25 Managing diet and fitness d 5701

26 Managing medications and following health advice d 57020

27 Seeking advice or assistance from caregivers or professionals d 57021

28 Avoiding risks of abuse of drugs or alcohol d 57022

8 Looking after one's safety 29 Looking after one's safety d 571

No.	Self-care category	Activity no.	Description	Codes
1	Washing oneself	1	Washing body parts	d 5100
		2	Washing whole body	d 5101
		3	Drying oneself	d 5102
2	Caring for body parts	4	Caring for skin	d 5200
		5	Caring for teeth	d 5201
		6	Caring for hair	d 5202
		7	Caring for fingernails	d 5203
		8	Caring for toenails	d 5204
		9	Caring for nose	d 5205
3	Toileting	10	Indicating need for urination	d 53000
		11	Carrying out urination appropriately	d 53001
		12	Indicating need for defecation	d 53010
		13	Carrying out defecation appropriately	d 53011
		14	Menstrual care	d 5302
4	Dressing	15	Putting on clothes	d 5400
		16	Taking off clothes	d 5401
		17	Putting on footwear	d 5402
		18	Taking off footwear	d 5403
		19	Choosing appropriate clothing	d 5404
5	Eating	20	Indicating need for eating	d 5500
		21	Carrying out eating appropriately	d 5501
6	Drinking	22	Indicating need for drinking	d 5600
		23	Carrying out drinking appropriately	d 5602
7	Looking after one's health	24	Ensuring one's physical comfort	d 5700
		25	Managing diet and fitness	d 5701
		26	Managing medications and following health advice	d 57020
		27	Seeking advice or assistance from caregivers or professionals	d 57021
		28	Avoiding risks of abuse of drugs or alcohol	d 57022
8	Looking after one's safety	29	Looking after one's safety	d 571

Table 2.

The feature codes used in the SCADI.

Feature codes^a
1. Age	42. d 5202-4	83. d 53010-3	124. d 5403-2	165. d 5700-1
2. Gender	43. d 5202-8	84. d 53010-4	125. d 5403-3	166. d 5700-2
3. d 5100-0	44. d 5202-9	85. d 53010-8	126. d 5403-4	167. d 5700-3
4. d 5100-1	45. d 5203-0	86. d 53010-9	127. d 5403-8	168. d 5700-4
5. d 5100-2	46. d 5203-1	87. d 53011-0	128. d 5403-9	169. d 5700-8
6. d 5100-3	47. d 5203-2	88. d 53011-1	129. d 5404-0	170. d 5700-9
7. d 5100-4	48. d 5203-3	89. d 53011-2	130. d 5404-1	171. d 5701-0
8. d 5100-8	49. d 5203-4	90. d 53011-3	131. d 5404-2	172. d 5701-1
9. d 5100-9	50. d 5203-8	91. d 53011-4	132. d 5404-3	173. d 5701-2
10. d 5101-0	51. d 5203-9	92. d 53011-8	133. d 5404-4	174. d 5701-3
11. d 5101-1	52. d 5204-0	93. d 53011-9	134. d 5404-8	175. d 5701-4
12. d 5101-2	53. d 5204-1	94. d 5302-0	135. d 5404-9	176. d 5701-8
13. d 5101-3	54. d 5204-2	95. d 5302-1	136. d 5500-0	177. d 5701-9
14. d 5101-4	55. d 5204-3	96. d 5302-2	137. d 5500-1	178. d 57020-0
15. d 5101-8	56. d 5204-4	97. d 5302-3	138. d 5500-2	179. d 57020-1
16. d 5101-9	57. d 5204-8	98. d 5302-4	139. d 5500-3	180. d 57020-2
17. d 5102-0	58. d 5204-9	99. d 5302-8	140. d 5500-4	181. d 57020-3
18. d 5102-1	59. d 5205-0	100. d 5302-9	141. d 5500-8	182. d 57020-4
19. d 5102-2	60. d 5205-1	101. d 5400-0	142. d 5500-9	183. d 57020-8
20. d 5102-3	61. d 5205-2	102. d 5400-1	143. d 5501-0	184. d 57020-9
21. d 5102-4	62. d 5205-3	103. d 5400-2	144. d 5501-1	185. d 57021-0
22. d 5102-8	63. d 5205-4	104. d 5400-3	145. d 5501-2	186. d 57021-1
23. d 5102-9	64. d 5205-8	105. d 5400-4	146. d 5501-3	187. d 57021-2
24. d 5200-0	65. d 5205-9	106. d 5400-8	147. d 5501-4	188. d 57021-3
25. d 5200-1	66. d 53000-0	107. d 5400-9	148. d 5501-8	189. d 57021-4
26. d 5200-2	67. d 53000-1	108. d 5401-0	149. d 5501-9	190. d 57021-5
27. d 5200-3	68. d 53000-2	109. d 5401-1	150. d 5600-0	191. d 57021-6
28. d 5200-4	69. d 53000-3	110. d 5401-2	151. d 5600-1	192. d 57022-0
29. d 5200-8	70. d 53000-4	111. d 5401-3	152. d 5600-2	193. d 57022-1
30. d 5200-9	71. d 53000-8	112. d 5401-4	153. d 5600-3	194. d 57022-2
31. d 5201-0	72. d 53000-9	113. d 5401-8	154. d 5600-4	195. d 57022-3
32. d 5201-1	73. d 53001-0	114. d 5401-9	155. d 5600-8	196. d 57022-4
33. d 5201-2	74. d 53001-1	115. d 5402-0	156. d 5600-9	197. d 57022-8
34. d 5201-3	75. d 53001-2	116. d 5402-1	157. d 5602-0	198. d 57022-9
35. d 5201-4	76. d 53001-3	117. d 5402-2	158. d 5602-1	199. d 571-0
36. d 5201-8	77. d 53001-4	118. d 5402-3	159. d 5602-2	200. d 571-1
37. d 5201-9	78. d 53001-8	119. d 5402-4	160. d 5602-3	201. d 571-2
38. d 5202-0	79. d 53001-9	120. d 5402-8	161. d 5602-4	202. d 571-3
39. d 5202-1	80. d 53010-0	121. d 5402-9	162. d 5602-8	203. d 571-4
40. d 5202-2	81. d 53010-1	122. d 5403-0	163. d 5602-9	204. d 571-8
41. d 5202-3	82. d 53010-2	123. d 5403-1	164. d 5700-0	205. d 571-9

The feature codes are defined according to the ICF framework and the details can be found in reference.⁹

Table 3.

Classes of the children's self-care problems.

Class No.	Description
1	Caring for body parts problem
2	Toileting problem
3	Dressing problem
4	Washing oneself and Caring for body parts and Dressing problem
5	Washing oneself, Caring for body parts, Toileting, and Dressing problem
6	Eating, Drinking, Washing oneself, Caring for body parts, Toileting, Dressing, Looking after one's health and Looking after one's safety problem
7	No Problem

Figure 5 shows the statistics of the data. Figure 5(a) presents the proportions of different ages, and Figure 5(b) shows the instances of different classes. Besides, there are a few instances in class 1, class 3, and class 5. For example, class 1, class 3, and class 5 only have 2, 1, and 3 instances, respectively. Since the triplet loss requires at least two positive instances and one negative instance to serve as the anchor, positive, and negative in the embedding space, we doubled the SCADI dataset by augmenting the gender field. The assumption is in line with the fact that the gender field cannot affect the diagnosis of children's self-care problems in the vast majority of cases. Table 4 displays the class distribution, including the number of instances for each class.

Figure 5.

The statistics of children with self-care problems contained in the SCADI dataset. (a) The proportions of different ages in the SCADI dataset; (b) The instances of different classes in the SCADI dataset.

Table 4.

The distribution of the classes and instances.

Class No.	Instances in the SCADI dataset	Instances in the augmented dataset
1	2	4
2	7	14
3	1	2
4	12	24
5	3	6
6	29	58
7	16	32

3.2 Evaluations

As a multiclass classification problem, the classification performances of children's self-care problems can be evaluated by common metrics such as accuracy, precision, recall, and F1 score, respectively.

\begin{aligned} a c c u r a c y & = \frac{c o r r e c t c l a s s i f i c a t i o n s}{a l l c l a s s i f i c a t i o n s} \end{aligned}

(5)

\begin{aligned} p r e c i s i o n_{k} & = \frac{t r u e p o s i t i v e s}{t r u e p o s i t i v e s + f a l s e p o s i t i v e s} \end{aligned}

(6)

\begin{aligned} r e c a l l_{k} & = \frac{t r u e p o s i t i v e s}{t r u e p o s i t i v e s + f a l s e n e g a t i v e s} \end{aligned}

(7)

\begin{aligned} F 1 s c o r e & = \frac{2}{p r e c i s i o n^{- 1} + r e c a l l^{- 1}} \end{aligned}

(8)

where

a c c u r a c y

represents the classification accuracy.

p r e c i s i o n_{k}

and

r e c a l l_{k}

means the precision and recall performances of k-th class, respectively. Furthermore,

p r e c i s i o n

denotes the average of

p r e c i s i o n_{k}

, and

r e c a l l

means the average of

r e c a l l_{k}

In addition, 10-fold cross-validation⁴⁷ is employed to comprehensively and reliably assess the classification performances. The 10-fold cross-validation is a kind of statistical method. It divides the whole dataset into 10 subsets. The training process is conducted on nine of these subsets, while the remaining one subset is used to validate the classification performances. The above process is repeated 10 times until each one subset has been used once as the validation set. Table 5 presents the split of the training and validation datasets in the 10-fold cross-validation.

Table 5.

The split of the training and validation datasets of 10-fold cross-validation.

10-fold (140^a)	Validation dataset (size)	Training dataset (size)
1	The first one subset (14^b)	Remaining nine subsets (126^c)
2	The second one subset (14)	Remaining nine subsets (126)
3	The third one subset (14)	Remaining nine subsets (126)
…	…	…
10	The tenth one subset (14)	Remaining nine subsets (126)

(140) represents that the total size of the SCADI dataset is 140.

(14) denotes that the size of the validation dataset is 14.

(126) means the size of the training dataset is 126.

3.3 Hyperparameters

Several trials have been carried out to select the reasonable hyperparameters of the first sub-deep neural network (which is employed to extract useful information from the features) and the second sub- deep neural network (which is used to handle the imbalance problem).

Given that the classification of children's self-care problems is a multiclass classification problem, the activation function used in the proposed model is set to be the ReLU. The loss function of the first sub- deep neural network is chosen as the triplet loss function, while the loss function of the second sub-deep neural network is selected as the focal loss function.

For both sub-deep neural networks, the selection criterion is to choose the simplest architecture that achieves the best validation performance. The triplet loss margin is set to 0.1 to balance between intra-class compactness and inter-class separability, while the focal loss parameter gamma is set to 2.0 to effectively down-weight easy samples. The dropout regularization technique⁴⁸ is not used in this approach. The Adam optimizer⁴⁹ is employed to train the deep neural networks. The number of training epochs is set to be 50 for stable convergence. The hyperparameters and architecture of the proposed model are determined and presented in Table 6.

Table 6.
The hyperparameters and architecture of the model.

Base model Loss function Architecture

Layer Neurons

The first sub-neural network Triplet loss Input layer 205

$m a r g i n$ = 0.1 Hidden layer 200

Hidden layer 150

Output layer Embedding size

The second sub-neural network
Focal loss
Input layer Embedding size

$γ = 2.0$ Hidden layer 100

Hidden layer 50

Output layer Class size

Base model	Loss function	Architecture
The first sub-neural network	Triplet loss	Input layer	205
	$m a r g i n$ = 0.1	Hidden layer	200
		Hidden layer	150
		Output layer	Embedding size
The second sub-neural network	Focal loss	Input layer	Embedding size
	$γ = 2.0$	Hidden layer	100
		Hidden layer	50
		Output layer	Class size

3.4 Classification performances

The performances of the proposed method for classifying the children's self-care problems are investigated and compared with other methods in the up-to-date references such as probabilistic neural network (PNN), classification and regression tree (CART), genetic algorithm combined probabilistic neural net- work (GA-PNN),³⁷ naive Bayes (NB), support vector machine (SVM), k-nearest neighbor (KNN), principal component analysis based decision tree (PCA-DT),³⁸ and a deep learning based approach named Care2Vec.³⁹ Besides, two scenarios are considered. In the first scenario, feature dimension compression is not implemented. The feature dimension remains unchanged which is 205. In contrast, the second scenario considers the feature compression procedure. The feature dimension is compressed from 205 to 34 and 56.

The experimental environment comprises Python 3.6 with TensorFlow 2.0, running on a 64-bit version of Microsoft Windows 11. The system is equipped with an AMD Ryzen 75800H processor, Nvidia GeForce RTX 3060 GPU, and 16 GB DDR4 RAM. Additionally, the dataset and code have been up-loaded to GitHub, and the URL is https://github.com/yangyugit/self-care_problems_classification.

The numerical results are presented in Table 7. It can be seen that the proposed method outperforms in both scenarios (with and without the feature compression procedure). In the first scenario, CART has the worst accuracy which is 76.82%. PNN obtains an accuracy of 78.56%. Care2Vec and MLP achieve the accuracy rate of 90.00% and 90.47%, respectively. Furthermore, the proposed model obtains 98.41% accuracy, when only employing the triplet loss function in the first sub-deep neural network. Based on this, we are capable of achieving 100% accuracy by incorporating the focal loss in the second sub-deep neural network. In terms of precision, recall, and F1 score, CART outperforms PNN, with 0.90 precision, 0.81 recall, and 0.85 F1 score compared with PNN's 0.81 precision, 0.75 recall, and 0.77 F1 score. Care2Vec and MLP obtain the same precision, which is 0.94. However, Care2Vec exhibits better recall and F1 score than MLP, with 0.88 recall and 0.91 F1 score compared with MLP's 0.85 recall and 0.89 F1 score. The neural network with only the triplet loss function achieves a precision, recall, and F1 score of 0.98. Lastly, the proposed method incorporating both the triplet loss function and the focal loss function achieves a precision, recall, and F1 score of 1.00.

Table 7.
The performances of the children's self-care problems classification.

Method^a Dimension Performance evaluation metric

Accuracy Precision Recall F1 score

CART 205 76.82% 0.90 0.81 0.85

PNN 78.56% 0.81 0.75 0.77

Care2Vec 90.00% 0.94 0.88 0.91

MLP^b 90.47% 0.94 0.85 0.89

Triplet^c 98.41% 0.98 0.98 0.98

Focal^d 100.00% 1.00 1.00 1.00

PCA-PNN 34 81.42% 0.90 0.85 0.87

GA-PNN ³⁷ 94.28% 0.94 0.98 0.95

GA-SVM 87.64% 0.86 0.89 0.87

Care2Vec 91.42% 0.92 0.90 0.91

Triplet 98.41% 0.98 0.98 0.98

Focal 100.00% 1.00 1.00 1.00

NB 56 78.57% 0.91 0.78 0.85

KNN 87.14% 0.91 0.95 0.91

SVM 88.57% 0.87 1.00 0.93

PCA-DT ³⁸ 94.29% 0.98 0.94 0.95

Care2Vec ³⁹ 91.42% 0.93 0.88 0.90

Triplet 98.41% 0.98 0.98 0.98

Focal 100.00% 1.00 1.00 1.00

Method^a	Dimension	Performance evaluation metric
CART	205	76.82%	0.90	0.81	0.85
PNN		78.56%	0.81	0.75	0.77
Care2Vec		90.00%	0.94	0.88	0.91
MLP^b		90.47%	0.94	0.85	0.89
Triplet^c		98.41%	0.98	0.98	0.98
Focal^d		100.00%	1.00	1.00	1.00
PCA-PNN	34	81.42%	0.90	0.85	0.87
GA-PNN ³⁷		94.28%	0.94	0.98	0.95
GA-SVM		87.64%	0.86	0.89	0.87
Care2Vec		91.42%	0.92	0.90	0.91
Triplet		98.41%	0.98	0.98	0.98
Focal		100.00%	1.00	1.00	1.00
NB	56	78.57%	0.91	0.78	0.85
KNN		87.14%	0.91	0.95	0.91
SVM		88.57%	0.87	1.00	0.93
PCA-DT ³⁸		94.29%	0.98	0.94	0.95
Care2Vec ³⁹		91.42%	0.93	0.88	0.90
Triplet		98.41%	0.98	0.98	0.98
Focal		100.00%	1.00	1.00	1.00

The proposed method is compared with the state-of-the-art approaches from recent studies,^37–39 which introduce GA-PNN, PCA-DT, and Care2Vec, respectively. The comparison is conducted under the condition that each of these methods achieves its best performance. Additionally, these studies also evaluate other methods, such as CART, PNN, PCA-PNN, GA-SVM, NB, KNN, and SVM, which are likewise included in our comparisons.

MLP represents a multi-layer perceptron that has the same structure as the second sub-deep neural network in the proposed method.

Triplet denotes the first sub-deep neural network with the triplet loss function.

Focal represents the second sub-deep neural network with the focal loss that has been used based on the first sub-deep neural network (i.e., the entire proposed method).

In the second scenario (the feature has been compressed to 34 dimensions), PCA-PNN has the lowest accuracy, recall, and F1 score, which are 81.42%, 0.85, and 0.87, respectively. GA-PNN performs better by obtaining 94.28% accuracy, 0.94 precision, 0.98 recall, and 0.95 F1 score. Care2Vec achieves 91.42% accuracy, 0.92 precision, 0.90 recall, and 0.91 F1 score. Compared with the performances of PNN and Care2Vec in the first scenario, these results indicate that the feature compressing procedure can extract useful information, eliminate noise, and be beneficial for improving the classification performance. However, GA-SVM performs moderately with 87.64% accuracy, 0.86 precision, 0.89 recall, and 0.87 F1 score. The reason could be attributed to the limited modeling capability of the SVM model compared with the PNN model. The neural network with only triplet loss technology achieves 98.41% accuracy, 0.98 precision, 0.98 recall, and 0.98 F1 score. Lastly, the proposed method with both the triplet loss function and the focal loss function achieves 100.00% accuracy, 1 precision, 1 recall, and 1 F1 score. When the feature is compressed to 56 dimensions, NB performs worst with 78.57% accuracy, 0.91 precision, 0.78 recall, and 0.85 F1 score. KNN and SVM perform similarly, and both outperform NB. Specifically, KNN obtains 87.14% accuracy, 0.91 precision, 0.95 recall, and 0.91 F1 score, while SVM achieves 88.57% accuracy, 0.87 precision, 1.00 recall, and 0.93 F1 score. PCA-DT achieves the third best performance with 94.29% accuracy, 0.98 precision, 0.94 recall, and 0.95 F1 score. Moreover, the neural network with only the triplet loss technology outperforms PCA-DT with 98.41% accuracy, 0.98 precision, 0.98 recall, and 0.98 F1 score. Lastly, the proposed method incorporating both the triplet loss function and the focal loss function achieves the best performance with 100.00% accuracy, 1 precision, 1 recall, and 1 F1 score.

Besides, the execution time of the proposed neural network model is investigated, considering the real-time diagnostic issues. The dataset loading time is 0.0648 s. The training time of the proposed neural network is 61.9405 s. And the classification time is only 0.0013 s. That satisfies the requirements for real-time on-site diagnosis to assist doctors.

Despite the proposed method achieving excellent performance in the classification of the children's self-care problems, its outstanding performance (i.e., 100.00% accuracy, 1 precision, 1 recall, and 1 F1 score) inevitably raises doubts and concerns. Generally, the primary cause of overly good performance is overfitting. Overfitting refers to when a deep neural network model performs well on the training dataset but poorly on the test dataset. This can be understood as the deep learning model overly focusing on the details of the data and neglecting to extract knowledge from the data. Although the excellent performance mentioned above is obtained under 10-fold cross-validation, we still want to further confirm whether overfitting occurs. Therefore, we independently repeat 10 sets of 10-fold cross-validation. This is equivalent to randomly partitioning the dataset 100 times and creating separate training and test datasets each time. Then, the proposed method is tested in the separated datasets. The testing results are presented in Table 8 and Figure 6.

Figure 6.

The testing results of the proposed method repeating 10 sets of 10-fold cross-validation.

Table 8.

The 10-fold cross-validation performances.

Feature dimension	Accuracy (SD)	Precision (SD)	Recall (SD)	F1 score (SD)
205	98.98% (0.04)	0.99 (0.04)	0.99 (0.04)	0.99 (0.04)
34	99.53% (0.03)	0.99 (0.03)	0.99 (0.03)	0.99 (0.03)
56	98.73% (0.04)	0.99 (0.03)	0.99 (0.05)	0.99 (0.04)

Without feature compression (i.e., feature dimension is 205), we observe that out of 100 experimental trials, accuracy, recall, and F1 score could not achieve 1 in 6 trials, precision could not reach 1 in only 5 trials. The mean and standard deviation of accuracy, precision, recall, and F1 score are 98.98% (SD = 0.04), 0.99 (SD = 0.04), 0.99 (SD = 0.04), and 0.99 (SD = 0.04), respectively. When the feature is compressed to 34 dimensions, we observe that out of 100 experimental trials, accuracy, precision, and F1 score cannot reach 1 in only 3 trials, recall cannot reach 1 in only 2 trials. The mean and standard deviation of accuracy, precision, recall, and F1 score are 99.53% (SD = 0.03), 0.99 (SD = 0.03), 0.99 (SD = 0.03), and 0.99 (SD = 0.03), respectively. When the feature is compressed to 56 dimensions, we find that out of 100 experimental trials, accuracy cannot reach 100% in 8 trials, precision cannot reach 1 in 5 trials, recall cannot reach 1 in 9 trials, and F1 score cannot reach 1 in 9 trials. The mean and standard deviation of accuracy, precision, recall, and F1 score are 98.73% (SD = 0.04), 0.99 (SD = 0.03), 0.99 (SD = 0.05), and 0.99 (SD = 0.04), respectively. These mean the proposed method maintains consistently impressive performance across 10 sets of 10-fold cross-validation runs. This indicates that overfitting does not occur. At the very least, it demonstrates that, for this children's self-care problem dataset, the proposed method exhibits impressive and stable classification performance.

To further validate the performance of the proposed method and examine whether overfitting occurs, we also conduct numerical experiments on two additional external datasets in related domains. The first dataset is the breast cancer dataset⁵⁰ which is about the classification of the radiation therapy to treat the breast cancer. It contains 286 instances, each with 9 features, and includes two classification results (i.e., radiation therapy and non-radiation therapy). The second dataset is the nursery dataset⁵¹ which is developed to the rank applications for nursery schools. It contains 12960 instances, each with 8 features, and includes three classification results (i.e., recommended, not-recommended, and priority). Table 9 presents the performance of the proposed method tested in the breast cancer dataset and the nursery dataset. It can be seen that the proposed method achieves a moderate performance in the breast cancer dataset with 82.79% accuracy, 0.83 precision, 0.83 recall, and 0.83 F1 score. However, compared with the Xgboost classification, SVM classification, RF classification, neural network classification, and logistic regression methods, the proposed method outperforms with the state-of-the-art results.⁵⁰ In the nursery dataset, the proposed method achieves outstanding performance with 98.89% accuracy, 0.99 precision, 0.99 recall, and 0.99 F1 score. These testing results in the additional external datasets verify the excellent performance of the proposed method.

Table 9.

The performances of the proposed method testing in the two additional external datasets.

Dataset	Accuracy	Precision	Recall	F1 score
Breast cancer	82.79%	0.83	0.83	0.83
Nursery	98.89%	0.99	0.99	0.99

Additionally, we perform a Shapley Additive Explanations (SHAP) analysis⁵² to interpret the proposed method for children's self-care problems. Figure 7 and Figure 8 present the results of SHAP analysis. Figure 7(a) presents the mean SHAP values, which are computed as the average of the absolute SHAP values across the seven classes, representing the overall feature importance in classifying children's self-care problems. It can be seen that Feature 1 has the highest SHAP value, indicating its dominant influence on the model's decision-making process. Several other features, such as Feature 186, Feature 102, and Feature 184, also exhibit relatively high SHAP values, suggesting their significant contributions to classification accuracy. Furthermore, the color distribution along each feature axis highlights the impact of high and low feature values on the model's predictions, where red indicates higher feature values and blue represents lower values. This visualization demonstrates how specific features affect the decision boundaries of the model, reinforcing the interpretability of the proposed approach. Figure 7(b)–(d) and Figure 8 show the feature importance for each of the classes of the children's self-care problems. Each subgraph illustrates the impact of different features on the model's classifications for a specific class. It can be observed that Features 1, 186, 102, 116, 74, 99, 185, 55, 199, 97, 135, 104, 65, and 184 contribute more significantly to the classification of children's self-care problems across classes 1 to 7, respectively.

Figure 7.

The average feature importance across all classes and the feature importance for classes 1-3. (a) Average feature importance across all classes. (b) The feature importance for class 1. (c) The feature importance for class 2. (d) The feature importance for class 3.

Figure 8.

The feature importance for classes 4-7 (continued from fig. 7). (a) The feature importance for class 4. (b) The feature importance for class 5. (c) The feature importance for class 6. (d) The feature importance for class 7.

Lastly, the effect of data augmentation on the SCADI dataset in classifying the children's self-care problems is evaluated. From Table 4, it can be seen that, when the gender information, which is less relevant to the classification of the children's self-care problems, is excluded, the SCADI dataset can be doubled in size. Furthermore, Table 10 presents the effect of the data augmentation on the classification performance. It can be seen that clear and consistent performance improvements are achieved across all evaluation metrics (i.e., accuracy, precision, recall, and F1 score) after applying the augmentation method to the SCADI dataset. For example, without the feature compression procedure, the accuracy increases from 77.18% to 98.98%, the precision increases from 0.83 to 0.99, the recall increases from 0.77 to 0.99, and the F1 score rises from 0.80 to 0.99, in the original dataset to the augmented dataset, respectively. When the feature dimension is compressed to 34, the accuracy improves from 79.96% to 99.53%, the precision improves from 0.85 to 0.99, the recall improves from 0.77 to 0.99, and the F1 score improves from 0.81 to 0.99, respectively. When the feature dimension is compressed to 56, the accuracy improves from 79.96% to 98.73%, the precision improves from 0.86 to 0.99, the recall improves from 0.79 to 0.99, and the F1 score improves from 0.82 to 0.99, respectively. Besides, the standard deviations across all metrics are also significantly reduced after the dataset augmentation. That demonstrates the higher model stability and generalization capability. These results highlight that the proposed data augmentation strategy can effectively enrich the SCADI dataset size and enable more accurate and reliable classification of the children's self-care problems, regardless of the feature dimension. It is worth noting that, due to the fact that the class 3 contains only a single instance, it is not feasible to apply the triplet loss function in the original SCADI dataset. Therefore, we adopt an autoencoder network with the same structure as the first sub-neural network used in the proposed model to extract representative embeddings of the instances. These embeddings are then fed into the same second sub-neural network and trained with the focal loss function for a fair comparative experiment.

Table 10.

The effect of data augmentation in classifying the children's self-care problems (tested by 10-fold cross-validation).

Dataset size	Feature dimension	Accuracy (SD)	Precision (SD)	Recall (SD)	F1 score (SD)
Original	205	77.18% (0.11)	0.83 (0.11)	0.77 (0.11)	0.80 (0.11)
	34	79.96% (0.13)	0.85 (0.14)	0.77 (0.11)	0.81 (0.11)
	56	79.96% (0.12)	0.86 (0.10)	0.79 (0.12)	0.82 (0.10)
Augmented	205	98.98% (0.04)	0.99 (0.04)	0.99 (0.04)	0.99 (0.04)
	34	99.53% (0.03)	0.99 (0.03)	0.99 (0.03)	0.99 (0.03)
	56	98.73% (0.04)	0.99 (0.03)	0.99 (0.05)	0.99 (0.04)

3.5 Ablation experiment

To furthermore quantitatively verify the effects, which are contributed by the triplet loss (employed by the first sub-deep neural network) and the focal loss (used in the second sub-deep neural network), in improving the classification performance of the children's self-care problems, several trials that consider different dimensions of feature extraction are conducted. Figure 9 presents the improving effects contributed by triplet loss and focal loss, respectively.

Figure 9.

The improving effect contributed by the triplet loss (used in the first sub-deep neural net- work) and focal loss (used in the second sub-deep neural network). MLP denotes the multi-layer perceptron with the same structure as the second sub-deep neural network. Triplet represents the performance obtained by the triplet loss used in the first sub-deep neural network. Focal denotes the performance achieved by the focal loss used in the second sub-deep neural network with the first sub-deep neural network (i.e., the entire proposed method). (a) The accuracy; (b) The precision; (c) The recall; (d) The F1 score.

It can be seen that both the triple loss and the focal loss can improve the classification performances of the children's self-care problems. Furthermore, based on the triplet loss which is employed to extract useful information from the feature, the focal loss is able to further improve the classification performance. As an illustration, we can observe that the triple loss improves the accuracy from 90.47% to 98.41%, the precision from 0.94 to 0.98, the recall from 0.85 to 0.98, the F1 score from 0.89 to 0.98, with compressing the feature to 34 dimensions, respectively. Additionally, the focal loss can further improve the accuracy from 98.41% to 100.00%, the precision from 0.98 to 1, the recall from 0.98 to 1, and the F1 score from 0.98 to 1, respectively. By contrast, when we compress the feature dimension to 56, we find the triple loss also can improve the accuracy, precision, recall, and F1 score from 90.47% to 98.41%, from 0.94 to 0.98, from 0.85 to 0.98, and from 0.89 to 0.98, respectively. Besides, the focal loss is capable of further improving the accuracy, precision, recall, and F1 score to 100%, 1, 1, and 1, respectively. It can be seen that the triplet loss and the focal loss can improve the classification performances (evaluated by accuracy, precision, recall, and F1 score) of the children's self-care problems about 10% totally, when the feature has been compressed to 34 and 56 dimensions. Furthermore, to comprehensively investigate the improving abilities of triplet loss and focal loss on the classification performances of the children's self-care problems, we conduct more experimental trials from 20 to 205 feature dimensions. It can be found that the triplet loss is able to at most improve 9.53% accuracy (from 90.47% to 100.00% with feature compressed to 80 dimensions), 0.06 precision (from 0.94 to 1 with feature compressed to 80 and 100 dimensions), 0.15 recall (from 0.85 to 1 with feature compressed to 80 dimensions), 0.11 F1 score (from 0.89 to 1 with feature compressed to 80 dimensions). Moreover, based on the triplet loss, the focal loss is capable of further improving at most 3.17% accuracy (from 96.83% to 100.00% with feature compressed to 20, 30, 90, 125, and 175 dimensions), 0.03 precision (from 0.97 to 1 with feature compressed to 90, 125, and 175 dimensions), 0.03 recall (from 0.97 to 1 with feature compressed to 20, 125, and 175 dimensions), 0.03 F1 score (from 0.97 to 1 with feature compressed to 125 and 175 dimensions). At worst, the focal loss may not further improve the classification performances. When we compress the feature to 80 dimensions, it can be seen that the focal loss cannot further improve the accuracy, precision, recall, and F1 score. The improvements in accuracy, precision, recall, and F1 score contributed by the focal loss are all 0, whereas relying solely on the triplet loss has resulted in improvements in all metrics to 1. When we compress the feature to 70 dimensions, we can observe that the focal loss only improves 1.59% accuracy (from 98.41% to 100.00%), 0.02 precision (from 0.98 to 1), but decreases the recall by 0.01 (from 0.98 to 0.97). However, in the whole range of the dimension of feature extraction, the effects of improving the classification performances of the triplet loss and focal loss are prominent. The averages of accuracy, precision, recall, and F1 score achieved by the triplet loss are improved from 90.47%, 0.94, 0.85, and 0.89 to 97.77%, 0.98, 0.97, and 0.97, respectively. Additionally, based on the improving effect of the triplet loss, the focal loss improves the averages of accuracy, precision, recall, and F1 score to 99.78%, 0.99, 0.99, and 0.99, respectively. Therefore, we can observe that the triplet loss technology is able to improve 7.30% accuracy, 0.04 precision, 0.12 recall, and 0.08 F1 score. The focal loss technology is capable of improving 2.01% accuracy, 0.01 precision, 0.02 recall, and 0.02 F1 score, respectively. According to the above analysis, we can conclude that the triplet loss and the focal loss can contribute to improving the classification performances of the children's self-care problems.

In a nutshell, the numerical tests verify the following: 1)

The triplet loss improves accuracy by 7.30%, precision by 0.04, recall by 0.12, and F1 score by 0.08, while the focal loss improves accuracy by 2.01%, precision by 0.01, recall by 0.02, and F1 score by 0.02. Together, these two sub-networks complement each other and lead to the state-of-the-art performance.

The first sub-deep neural network, equipped with the triplet loss, contributes by learning discriminative feature embeddings through dimensionality compression, key information extraction, and noise reduction, thereby enhancing intra-class compactness and inter-class separability.

The second sub-deep neural network, equipped with the focal loss, addresses the class imbalance problem by down-weighting easy samples and emphasizing hard and minority-class samples, which further boosts the classification performance.

4 Discussion

This study proposes a novel neural network that integrates the representation learning and the focal loss technique to accurately classify the children's self-care problems. The numerical tests validate its state-of-the-art classification performances that achieve average accuracy, precision, recall, and F1 score of 99.78%, 0.99, 0.99, and 0.99, respectively.

Compared with previous studies, the superior performance of the proposed method could be attributed to two factors i.e., the two sub-deep neural networks with the triple loss and focal loss technologies, as well as the augmentation method.

The first contributing factor is the utilization of the triplet loss and focal loss technologies. Fundamentally, triplet loss, which is used in the first sub-neural network, is a representation learning technique. Its objective is not merely to reduce dimensionality but to learn a meaningful and highly discriminative embedding space. The principle lies in learning feature embeddings based on the relative similarity of data instances. For each training sample (the “anchor”), the model considers a “positive” sample (from the same class) and a “negative” sample (from a different class). This process compels the model to discern and learn the most critical underlying features for distinguishing between different classes of children's self-care problems, rather than focusing on superficial or irrelevant features. In complex diagnostic tasks such as classifying children's self-care problems, the initial high-dimensional feature space is often contaminated with irrelevant or redundant features (i.e., “noise”). Such noise can mislead a standard classifier, causing it to learn spurious correlations and thereby degrading its generalization performance. Triplet loss inherently facilitates noise reduction through its learning paradigm. The second sub-neural network, which utilizes the focal loss function, can deal with the class imbalance problem to further improve the classification performance. The ablation experiment verifies that the improvement in classification performance can be attributed to the triplet loss and focal loss. The second contributing factor is the augmentation method of the SCADI dataset.

To the best of our knowledge, this is the first application of the data augmentation method on the SCADI dataset for studying the classification of children's self-care problems. We also conduct the corresponding experiment to verify the effect of data augmentation on the classification performance.

Besides, the proposed method can be easily transferred or applied to other datasets in related fields to deal with the classification problem. We have already conducted the classification experiments on two additional external datasets (i.e., the breast cancer dataset and the nursery dataset) to validate the performance of the proposed method. The experiments provide valuable insights into the model's generalizability, accuracy, and the inherent characteristics of the datasets themselves. First, the high performance on the large nursery dataset (12,960 instances) strongly indicates that our model architecture is robust and generalizes well to new data. Achieving nearly 99% accuracy on a completely different domain task confirms that the model does not simply overfit to the specific characteristics of the SCADI dataset. The core architecture, which combines triplet loss for discriminative feature learning and focal loss for classification, proves to be effective across different problem spaces. Second, the 82.79% accuracy on the breast cancer dataset suggests it is an intrinsically more challenging problem. This can be attributed to the small sample size. This dataset is relatively small, containing only 286 instances. Deep learning models are data-intensive, and with limited samples, it is significantly more challenging to learn complex, non-linear decision boundaries without overfitting. While 82.79% is numerically lower than the results on the other datasets, our method still outperforms other established machine learning models (such as Xgboost, SVM, RF, etc.) and achieves the state-of-the-art result. This insight is critical. It demonstrates that even when faced with a more challenging, smaller dataset, our proposed method is powerful enough to push the performance boundary and excel relative to other baselines.

Additionally, the experimental environment includes Python 3.6 with TensorFlow 2.0, an AMD Ryzen 7 5800H, an Nvidia GeForce RTX 3060 GPU, and 16GB DDR4 RAM. Given that software like Python 3.6 and TensorFlow are free, the primary costs are associated with the CPU, GPU, and memory. Based on current market prices, the estimated cost of the hardware components for the experimental environment ranges from approximately $1360 to $2150 USD, depending on the specific laptop model and configuration. If the user already owns a desktop computer and is only considering a GPU upgrade, the cost would be approximately $350 to $450 USD. Cloud computing services from providers like Amazon and Google can also be considered to further reduce costs.

5 Limitation

Despite the promising performance of the proposed model, there are still several limitations that need to be addressed.

First, the performance of deep learning models is highly dependent on the size and quality of the dataset. In this study, the SCADI dataset is relatively small and potentially suffers from class imbalance and data bias. In particular, the limited demographic diversity, such as underrepresentation of certain age groups, genders, and socio-cultural backgrounds, may introduce potential biases and restrict the model's ability to generalize across broader populations. This raises concerns about the equity, especially in healthcare-related applications where biased decisions could have critical consequences.

Although the corresponding data augmentation method is applied to mitigate these issues, generating diverse and realistic instances for tabular data remains a challenging task compared with image or text domains. In essence, the most effective solution is to collect more comprehensive and demographically balanced data. Expanding the dataset to include a wider range of representative samples can significantly improve the model's generalization ability and reduce bias. In future work, we also plan to explore the use of generative models to enrich tabular datasets in a more robust way.

Second, like many deep learning-based methods, the proposed model functions as a black box. That limits its interpretability. Understanding the mechanism of why the model makes certain classifications is crucial in healthcare-related applications. Incorporating explainable AI techniques to interpret the learned embeddings and classification outcomes would be a valuable direction for future research.

6 Conclusion

An accurate classification of children's self-care problems has important implications for the corresponding subsequent treatment. In this study, we propose a novel deep learning-based model to precisely classify the children's self-care problems. The proposed model includes two sub-deep neural networks. The first sub-deep neural network employs the triplet loss to extract key information from the feature, considering the correlation between the different instances. Then, the embedding is used by the second sub-deep neural network to classify the children's self-care problems. Meanwhile, considering the imbalance problem, the focal loss is used to further improve the classification performance. The numerical tests show that the proposed model achieves the state-of-the-art classification performance of the children's self-care problems. In all ranges of the dimension, the averages of accuracy, precision, recall, and F1 score are 99.78%, 0.99, 0.99, and 0.99, respectively. Besides, according to the ablation experiment, the triplet loss technology is able to improve 7.30% accuracy, 0.04 precision, 0.12 recall, and 0.08 F1 score. The focal loss technology is capable of improving 2.01% accuracy, 0.01 precision, 0.02 recall, and 0.02 F1 score, respectively. In future work, we plan to further investigate the application of the proposed model on larger and more diverse datasets, including different age groups and cultural backgrounds. Additionally, we aim to integrate temporal data to explore the evolution of self-care abilities over time, which could potentially support early intervention strategies. Finally, incorporating expert knowledge and clinical guidelines into the model design could further enhance its interpretability and applicability in pediatric rehabilitation.

Footnotes

Acknowledgements

The authors would like to thank the editor and the anonymous reviewers for their valuable comments and constructive suggestions, which have greatly improved the quality of this article.

Ethical approval

This study did not involve direct experimentation with human or animal subjects. All data used were obtained from public datasets or previously published studies. Therefore, ethical approval was not required.

Author contributions

Yang Yu and Jieqiong Liu conceived and designed the study. Yang Yu implemented the algorithm. Yijia Tang, Xiaoyan Zhang, and Tingyu Zhang performed the validation. The manuscript was drafted by Yang Yu, and all authors contributed to reviewing, editing, and final approval. Jieqiong Liu supervised the project.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the Shanghai Pujiang Program, (grant number 22PJC074) and the General Project of the Education Department of Zhejiang Province, (grant number Y202352869).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

The dataset and code have been up-loaded to GitHub, . They also can be available from the corresponding author upon reasonable request.

Copyright and originality

The authors declare that this manuscript is original, has not been published previously, and is not under consideration for publication elsewhere. All authors have read and approved the final version of the manuscript.

ORCID iDs

Yang Yu

Yijia Tang

Xiaoyan Zhang

Tingyu Zhang

Jieqiong Liu

References

Richard

Shea

. Delineation of self-care and associated concepts. J Nurs Scholarsh. 2011; 43: 255–264.

Riegel

Dunbar

Fitzsimons

, et al.

Self-care research: where are we now? Where are we going?

Int J Nurs Stud. 2021; 116: 103402.

W.H. Organization. The role of self-care in achieving the right to health, 2022.

MacKichan

Paterson

Henley

, et al. Self-care in people with long term health problems: a community based survey. BMC Fam Pract. 2011; 12: 1–10.

Spedale

Luciani

Attanasio

, et al. Association between sleep quality and self-care in adults with heart failure: a systematic review. Eur J Cardiovasc Nurs. 2021; 20: 192–201.

Kitson

Feo

Lawless

, et al. Towards a unifying caring life-course theory for better self-care and caring solutions: a discussion paper. J Adv Nurs. 2022; 78: 6–20.

Zhu

Tanaka

Tomisaki

, et al. Do it yourself: the role of early self-care ability in social skills in Japanese preschool settings. Sch Psychol Int. 2022; 43: 71–87.

W.H. Organization. WHO guideline on self-care interventions for health and well-being. Geneva/Switzerland: World Health Organization, 2021.

W.H. Organization. International Classification of Functioning, Disability, and Health: Children & Youth Version: ICF- CY. Geneva/Switzerland: World Health Organization, 2007.

10.

Björck-Åkesson

Wilder

Granlund

, et al. The International Classification of Functioning, Disability and Health and the version for children and youth as a tool in child habilitation/early childhood intervention–feasibility and usefulness as a common language and frame of reference for practice. Disabil Rehabil. 2010; 32: 125–138.

11.

Chien

C-W

Rodger

Copley

, et al. Comparative content review of Children’s participation measures using the international classification of functioning, disability and health–children and youth. Arch Phys Med Rehabil. 2014; 95: 141–152.

12.

Schiariti

Selb

Cieza

, et al. International Classification of Functioning, Disability and Health Core Sets for children and youth with cerebral palsy: a consensus meeting. Dev Med Child Neurol. 2015; 57: 149–158.

13.

de Schipper

Lundequist

Coghill

, et al. Ability and disability in autism spectrum disorder: a systematic literature review employing the international classification of functioning, disability and health-children and youth version. Autism Res. 2015; 8: 782–794.

14.

Edwards

. Disability: Definitions, value and identity. bingdon/Oxon/United Kingdom: Radcliffe Publishing, 2005.

15.

Robert

Schalock

Tassé

. Intellectual Disability: Definition, Diagnosis, Classification, and Systems of Supports. 12th Edition. Washington, DC/USA: American Association on Intellectual and Developmental Disabilities, 2021.

16.

Spitzer

Endicott

. Medical and mental disorder: proposed definition and criteria. Annales Médico- Psychologiques, Revue Psychiatrique 2018; 176: 656–665.

17.

Telles-Correia

Saraiva

Gonçalves

. Mental disorder—the need for an accurate definition. Front Psychiatry. 2018; 9: 64.

18.

Eller

Lev

Yuan

, et al. Describing self-care self-efficacy: definition, measurement, outcomes, and implications. Int J Nurs Knowl. 2018; 29: 38–48.

19.

Freer

. Self-care: a health diary study. Med Care. 1980; 18: 853–861. DOI: 10.1097/00005650-198008000-00006

20.

Roberto

Reynolds

. Older women’s experiences with chronic pain: daily challenges and self-care practices. J Women Aging. 2002; 14: 5–23.

21.

Riegel

Dickson

Goldberg

, et al. Factors associated with the development of expertise in heart failure self-care. Nurs Res. 2007; 56: 235–243.

22.

Mahmoudi

Jalali

Ahmadi

, et al. Identifying critical success factors in Heart Failure Self-Care using fuzzy DEMATEL method. Appl Soft Comput. 2019; 84: 105729.

23.

ÖHRVALL

A-M

ELIASSON

A-C

Löwing

, et al. Self-care and mobility skills in children with cerebral palsy, related to their manual ability and gross motor function classifications. Dev Med Child Neurol. 2010; 52: 1048–1055.

24.

Eliwa

EHI

El Koshiry

Abd El-Hafeez

, et al. Secure and transparent lung and colon cancer classification using blockchain and microsoft azure. Adv Respir Med 2024; 92: 95.

25.

Ameen

Fattoh

Abd El-Hafeez

, et al. Advances in ECG and PCG-based cardiovascular disease clas- sification: a review of deep learning and machine learning methods. J Big Data. 2024; 11: 59.

26.

Mostafa

Mahmoud

Abd El-Hafeez

, et al. Feature reduction for hepatocellular carcinoma prediction using machine learning algorithms. J Big Data. 2024; 11: 88.

27.

Mostafa

Mahmoud

Abd El-Hafeez

, et al. The power of deep learning in simplifying feature selection for hepatocellular carcinoma: a review. BMC Med Inform Decis Mak. 2024; 24: 87.

28.

Abdel Hady

Abd El-Hafeez

. Utilizing machine learning to analyze trunk movement patterns in women with postpartum low back pain. Sci Rep. 2024; 14: 18726.

29.

Abdel Hady

Mabrouk

El-Hafeez

. Employing machine learning for enhanced abdominal fat pre- diction in cavitation post-treatment. Sci Rep. 2024; 14: 11004.

30.

Mabrouk

Hady

DAA

El-Hafeez

. Machine learning insights into scapular stabilization for alleviating shoulder pain in college students. Sci Rep. 2024; 14: 28430.

31.

Abdel Hady

Abd El-Hafeez

. Revolutionizing core muscle analysis in female sexual dysfunction based on machine learning. Sci Rep. 2024; 14: 4795.

32.

Eliwa

EHI

El Koshiry

Abd El-Hafeez

, et al. Tilizing convolutional neural networks to classify monkeypox skin lesions, Sci Rep. 13 (2023), 14495.

33.

Abdel Hady

Abd El-Hafeez

. Predicting female pelvic tilt and lumbar angle using machine learning in case of urinary incontinence and sexual dysfunction. Sci Rep. 2023; 13: 17940.

34.

Zarchi

Bushehri

Dehghanizadeh

. SCADI: a standard dataset for self-care problems classification of children with physical and motor disability. Int J Med Inf. 2018; 114: 81–87.

35.

Zdrodowska

Dardzińska-Głebocka

. Classification and action rules in identification and self-care assessment problems. Technol Health Care. 2022; 30: 257–269.

36.

Choudhury

Greene

Classification of functioning, disability, and health for children and youth: ICF-CY Self Care (SCADI Dataset) using predictive analytics. In: Proceedings of the 2019 IISE Annual Conference, 2019, pp.1-6.

37.

Bushehri

Zarchi

. An expert model for self-care problems classification using probabilistic neural network and feature selection approach. Appl Soft Comput. 2019; 82: 105545.

38.

Syafrudin

Alfian

Fitriyani

, et al. Improving efficiency of self-care clas- sification using PCA and decision tree algorithm. In: 2020 International Conference on Decision Aid Sciences and Application (DASA), IEEE, 2020, pp.224–227.

39.

Putatunda

. Care2Vec: a hybrid autoencoder-based approach for the classification of self-care problems in physically disabled children. Neural Comput Appl 2020; 32: 17669–17680.

40.

Solorio-Fernández

Carrasco-Ochoa

Martínez-Trinidad

. A review of unsupervised feature selection methods. Artif Intell Rev. 2020; 53: 907–948.

41.

Zhong

Wang

L-N

Ling

, et al. An overview on data representation learning: from traditional feature learning to recent deep learning. J Finance Data Sci 2016; 2: 265–278.

42.

Zhu

Dornaika

Ruichek

. Semi-supervised elastic manifold embedding with deep learning architecture. Pattern Recognit. 2020; 107: 107425.

43.

Maas

Hannun

. Rectifier nonlinearities improve neural network acoustic models. In: ICML Work- shop on Deep Learning for Audio, Speech and Language Processing , 2013.

44.

Goodfellow

Bengio

Courville

. Deep Learning. Cambridge/MA/USA: MIT, 2016.

45.

Schroff

Kalenichenko

Philbin

. Facenet: a unified embedding for face recognition and clustering. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp.815–823.

46.

Lin

T-Y

Goyal

Girshick

, et al. Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision , 2017, pp.2980–2988.

47.

Wong

T-T

Yeh

P-Y

. Reliable accuracy estimates from k-fold cross validation. IEEE Trans Knowl Data Eng. 2019; 32: 1586–1594.

48.

Srivastava

Hinton

Krizhevsky

, et al. Dropout: a simple way to prevent neural networks from overfitting. J Mach Learn Res 2014; 15: 1929–1958.

49.

Kingma

Adam: A method for stochastic optimization. In: the 3rd International Conference on Learning Representations, 2015, pp.1-15.

50.

Zwitter

Soklic

. Breast Cancer Dataset. https://archive.ics.uci.edu/dataset/14/breast+cancer.

51.

Rajkovic

. Nursery Dataset. https://archive.ics.uci.edu/dataset/76/nursery.

52.

Lundberg

Lee

S-I

. A unified approach to interpreting model predictions. In: Proceeding of the 31st Conference on Neural Information Processing Systems , 2017, pp.1–10.

Base model	Loss function	Architecture
		Layer	Neurons
The first sub-neural network	Triplet loss	Input layer	205
	$m a r g i n$ = 0.1	Hidden layer	200
		Hidden layer	150
		Output layer	Embedding size
The second sub-neural network	Focal loss	Input layer	Embedding size
	$γ = 2.0$	Hidden layer	100
		Hidden layer	50
		Output layer	Class size