Sage Journals: Discover world-class research

Abstract

Accurately predicting lower body dimensions in older women is essential for improving garment fit and wearing comfort. However, challenges in data acquisition and the morphological variability among older adults complicate this task. To address these issues, this study adopts a classification followed by modeling strategy, using a small set of key anthropometric inputs to estimate critical lower-body measurements relevant to pants pattern construction. Anthropometric data were collected from 217 women aged 60–80 in Northeast China, capturing 34 lower body parameters. Principal component analysis was performed to reduce dimensionality and extract six primary body shape factors. A two-step clustering method was then applied to determine the optimal number of body types and to analyze their morphological characteristics. Subsequently, stepwise linear regression models were developed for each body type using stature, weight, and abdominal girth as predictor variables. The results demonstrate that the proposed models achieved good prediction accuracy. Overall, this research provides a data-driven foundation for customized pants pattern development and intelligent production of garments tailored for older women.

Keywords

age lower body classification 3D body scanning cluster analysis size prediction anthropometrics

Introduction

With the accelerating global trend of population aging, the number of older women has increased significantly. In 2019, the global population aged 60 years and older reached 1 billion, and this figure is projected to double to 2 billion by 2050, accounting for 22% of the total global population.^1,2 As a vital consumer group in the apparel industry, older women exhibit notable variability in body morphology, commonly characterized by abdominal protrusion, flattened hips, and altered lower-body proportions.^3,4 These changes pose challenges to conventional sizing systems, often resulting in poor fit issues such as tight waistbands, gaping at the hips, or strain in the crotch area, thereby compromising comfort and mobility.^5–7

Accurate anthropometric data serves as the foundation for apparel pattern design and are essential for constructing well-fitting garment templates.^8,9 Body measurement techniques can be categorized into manual contact methods and non-contact three-dimensional (3D) scanning technologies. Traditional tools such as tape measures are heavily dependent on the operator’s skill and experience, leading to subjective and less reproducible outcomes that are difficult to scale for large populations.^8,10 In contrast, 3D body scanning offers high-precision and repeatable data collection. Nevertheless, its high cost, operational complexity, and limited acceptance among older individuals constrain its widespread adoption in older adult garment development.^5,11 Thus, developing lower-body size prediction models based on a few easily accessible measurements is essential to advancing intelligent pants pattern design for older people.

Various prediction-based approaches have been proposed for estimating body dimensions. For example, Tan et al.¹² constructed multiple regression models using height, bust, and waist girth (W_C) to predict 28 body dimensions in women. Li et al.¹³ classified female lower-body shapes based on width–thickness ratios and derived girth estimates from key landmarks to optimize custom pants block generation. Zanwar et al.¹⁴ applied multiple linear regression (MLR) to establish predictive mechanisms using limited parameters, thereby improving measurement efficiency. More recently, machine learning algorithms such as artificial neural networks (ANN) and generalized regression neural networks (GRNN) have been employed in anthropometric modeling with improved fitting accuracy.^15–17 Other studies have used 2D image-based techniques; for instance, Gu et al.¹⁸ extracted frontal and lateral width–depth features from photographs to predict girths.

Although these methods have shown promising results in younger or standard-bodied populations, they often underperform in older women due to complex morphology and high inter-individual variability. Even among individuals of similar height and weight, older women may exhibit substantial differences in waist, abdomen, hip shapes, and posture curvature. Therefore, integrating body type classification into the modeling process—via a classification followed by modeling approach—offers a viable strategy to enhance model accuracy and individual adaptability.^12,19 This study collected lower-body 3D body scan data and manual measurement data from 217 women aged 60–80 in Northeast China and developed a corresponding prediction framework. The goal is to estimate key pants design dimensions from a limited set of easily measurable inputs, providing a data foundation for personalized pants development for older women. As shown in Figure 1, the study comprises the following steps:

Representative lower body variables are extracted using a hybrid approach combining 3D body scanning and manual measurements.

Dimensionality reduction via principal component analysis (PCA) is performed to identify major shape characteristics, followed by two-step clustering (TSC) for body type classification.

Multiple linear regression models are developed using stature (H), weight (W), and abdominal girth (A_C) to predict 12 essential lower-body dimensions for each body type, along with performance evaluation of the models.

Figure 1.

Framework of the proposed prediction method.

Methods

Data collection

This study recruited 217 older women aged 60–80 from Northeast China as the research subjects. Ethical approval was obtained from the university’s institutional review board, and all participants were informed of the study’s purpose and procedures. Written informed consent was obtained from each participant prior to data collection. The study was conducted between November 2020 and January 2021. Three-dimensional body scanning was conducted using the VITUS scanner. The scanning environment was maintained at 25 ± 3°C to ensure consistency with standard nude measurement conditions. The data acquisition procedures followed the ISO 20685-1:2018 standard “3D scanning methodologies for internationally compatible anthropometric databases,” ensuring consistency and comparability of measurement results.²⁰

To address privacy concerns, participants wore ultra-thin, non-restrictive privacy garments during scanning. All scan data were anonymized by removing personally identifiable information, and the resulting files were securely stored in an encrypted, access-controlled system. The research team members responsible for scanning and measurement received formal training in equipment operation and research ethics, including participant privacy protection, standardized anthropometric techniques, and data confidentiality protocols.

To ensure the accuracy of certain key dimensions that are difficult to extract directly from 3D body scan data, supplementary manual measurements were taken using a flexible tape measure. All measurement definitions followed the Chinese National Standard GB/T 16160-2017.²¹ A total of 34 lower-body feature variables were collected, including height, girths, widths, depths, lengths, and angles, as summarized in Table 1. Figure 2 illustrates the anatomical landmarks corresponding to each measurement variable. Each landmark is encoded with letters and numbers: letters represent specific anatomical regions, while numbers denote the horizontal plane position. Specifically, the number “1” corresponds to the right side of the body, “3” to the left side, “2” to the anterior midline, and “4” to the posterior midline.

Table 1.

Details of lower-body feature variables.

No.	Items	Definitions	Sources
1	Stature (H)	Vertical distance from the highest point of the head in the median line to ground	3D body scanner
2	Waist height (W_H)	Vertical distance from the waist level to the ground	3D body scanner
3	Abdominal height (A_H)	Vertical distance from horizontal plane of iliac crest point to ground	3D body scanner
4	Hip height (H_H)	Vertical distance from horizontal plane of most prominent point of buttocks to ground	3D body scanner
5	Crotch height (C_H)	Vertical distance from the horizontal plane of perineum point to ground	3D body scanner
6	Knee height (K_H)	Vertical distance from the midpoint of the patella to the ground	3D body scanner
7	Ankle height (An_H)	Vertical distance from lateral malleolus point to ground	3D body scanner
8	Waist girth (W_C)	Horizontal girth of waistline between the upper end of hip bone and lower edge of rib	3D body scanner
9	Abdominal girth (A_C)	Horizontal girth of the abdomen through iliac crest point	3D body scanner
10	Hip girth (H_C)	Horizontal girth measured at the fullest part of buttocks	3D body scanner
11	Thigh girth (T_C)	Maximum horizontal girth of thigh measured below gluteal groove	3D body scanner
12	Knee girth (K_C)	Measure the girth of the knee when the subject stands upright. When measuring, align the upper edge of the tape measure with the tibia point (knee)	3D body scanner
13	Ankle girth (An_C)	The horizontal girth of the most prominent point of the outer ankle bone when the legs are relaxed	3D body scanner
14	Crotch length (C_L)	The midpoint of the waistline of the front body through the perineum to the waistline of the back	3D body scanner
15	Waist width (W_W)	The horizontal straight line distance in the width direction between the most prominent parts on both sides of the waist section	Sliding anthropometer
16	Abdominal width (A_W)	The horizontal straight line distance in the width direction between the most prominent parts on both sides of the abdominal section	Sliding anthropometer
17	Hip width (H_W)	The horizontal straight line distance between the left and right greater trochanter points of the buttocks	Sliding anthropometer
18	Waist depth (W_T)	The horizontal straight line distance in the thickness direction between the most prominent parts of the front and back of the waist on the waistline	Sliding anthropometer
19	Abdominal depth (A_T)	The horizontal straight line distance in the thickness direction between the most prominent parts of the front and back of the abdomen at the height of the iliac crest point	Sliding anthropometer
20	Hip depth (H_T)	The horizontal straight line distance in the thickness direction between the most protruding parts of the front and back of the buttocks	Sliding anthropometer
21	Buttock inclined angle (W_A)	Angle between the vertical line of W_T midpoint and H_T midpoint and vertical line	Geomagic^® software
22	Lateral ventral angle(LV_A)	Angle between the ventral curve and upper vertical line	Geomagic^® software
23	Lateral buttock angle (LB_A)	Angle between the hip curve and upper vertical line	Geomagic^® software
24	Superior buttock convex angle (S_A)	Angle between the rear hip curve and upper vertical line	Geomagic^® software
25	Inferior buttock convex angle (I_A)	Angle between the rear hip curve and lower vertical line	Geomagic^® software
26	Ventral convex angle (V_A)	Angle between front abdomen curve and lower vertical line	Geomagic^® software
27	Body axis angle (B_A)	Angle between W_T midpoint and protruding point of the lateral ankle bone and upper vertical line	Geomagic^® software
28	H_C--W_C	Difference between H_C and waist girth	Calculate
29	H_C--A_C	Difference between H_C and A_C	Calculate
30	W_C--A_C	Difference between W_C and A_C	Calculate
31	H_T/2--A_T/2	Distance between the projection of W_T midpoint on hip thickness line and H_T midpoint	Calculate
32	W_R	W_W/W_T	Calculate
33	A_R	A_W/A_T	Calculate
34	H_R	H_W/H_T	Calculate

Note. 3D body scanner refers to a non-contact scanning device used to capture human body geometry and extract relevant measurement data; sliding anthropometer refers to a caliper-like instrument used to directly measure lower-body widths and depths from 3D body scan data; Geomagic^® software refers to a commercial 3D processing tool used to measure specific body angles.

Figure 2.

Key measurement points and coding for lower-body analysis.

Data preprocessing

To ensure the accuracy and reliability of the sample data, a series of preprocessing and filtering procedures were performed using SPSS 26.0. The preprocessing included three main procedures: outlier detection and removal, correction of measurement errors, and assessment of data normality. Together, they ensured that the dataset used for modeling was both statistically valid and technically robust.

Boxplots were first employed to detect potential outliers and missing values. Data points falling outside 1.5 times the interquartile range (IQR) were flagged as outliers. These values were cross-validated against the original 3D body scan files. If confirmed as errors, they were either corrected or excluded from the dataset. Missing values were handled based on the source and nature of the variables. When corresponding manual measurements were available, they were used to replace missing 3D body scan data. Variables with unresolvable or irrecoverable missing data were excluded from further analysis.

The normality of all 34 lower-body measurement variables was evaluated using Quantile–Quantile (Q–Q) plots to verify the applicability of parametric statistical methods in subsequent analysis. In the majority of cases, the points closely aligned with the diagonal line in the plots indicated an acceptable approximation to the normal distribution. Figure 3 presents a Q–Q plot for abdominal girth as an illustrative example. After data preprocessing, a total of 209 valid samples were retained for modeling. The Q–Q plot assessment confirmed that all retained variables met the assumptions of normality, supporting their suitability for PCA and regression analysis.

Figure 3.

Q–Q plot of abdominal circumference for the normality test.

Extraction of principal components

Principal component analysis is a multivariate statistical method that reduces data dimensionality while preserving the integrity of the original information. It transforms multiple correlated variables into a smaller set of uncorrelated components.²² In this study, the number of principal components retained was determined based on two criteria: (1) each component must have an eigenvalue greater than 1.0; and (2) components were selected in descending order of rotated factor loadings until the cumulative explained variance reached a sufficient threshold to represent the lower body shape information adequately.

Cluster analysis

This study employed the TSC method based on the Bayesian information criterion (BIC) to enable automated classification of lower body shapes in older women. Two-step clustering is a hybrid clustering algorithm capable of handling both continuous and categorical variables, and it determines the optimal number of clusters automatically using statistical criteria (e.g. AIC or BIC) rather than relying on arbitrary selection.^23–25 The TSC algorithm proceeds in two stages: the first phase uses distance measures to pre-cluster the data. In contrast, the second phase applies a probabilistic model—similar to Latent Class Analysis—to determine the optimal subgroup structure.^26,27 In this study, the TSC algorithm was implemented using IBM SPSS Statistics 26. Key body shape indicators derived from the PCA results, including horizontal girths, vertical lengths, and angular measurements, were used as input variables for clustering. The BIC values were computed for different clustering solutions, and the solution with the lowest BIC was selected as the optimal number of clusters.

Regression modeling

It is necessary to account for variations in body shape characteristics to improve the accuracy of lower body dimension predictions. Therefore, based on the previously identified three body shape clusters, this study developed separate linear regression models for each body type. Linear regression is a well-established and extensively validated modeling approach widely adopted in apparel design and pattern drafting applications.^28–30

The general mathematical form of the linear regression model is expressed as follows:

Y = β_{0} + β_{1} X_{1} + β_{2} X_{2} + \dots + β_{p} X_{p} + ε

(1)

In the equation, $Y$ denotes the dependent variable, representing the body measurement to be predicted; $X_{1}$ $X_{2}, \dots,$ $X_{p}$ represent the independent variables; $β_{0}$ is the intercept term; $β_{1}$ $β_{2}, \dots,$ $β_{p}$ refer to the regression coefficients; and $ε$ is the random error term. When the linear model includes only one dependent variable and a single independent variable, it is referred to as a simple linear regression model. If the model includes one dependent variable and two or more independent variables, it is defined as a multiple linear regression model.

The predictive accuracy of the regression models was evaluated using the root mean square error (RMSE), a standard measure for assessing model performance. The RMSE is calculated using the following equation:

E_{RMSE} = \sqrt{\frac{1}{n} \sum_{i = 1}^{n} {(y_{i} - {\hat{y}}_{i})}^{2}}

(2)

In the equation, n represents the number of test samples, $y_{i}$ is the $i$ actual measurement, and ${\hat{y}}_{i}$ is the $i$ predicted measurement. The smaller the error value, the better the predictive performance is.

Results

Statistical analysis of body shape characteristics in older women

The descriptive statistics of 209 valid samples of older women aged 60–80 years from Northeast China were analyzed using SPSS to obtain the minimum, maximum, mean, range, and standard deviation of key lower body dimensions and shape indicators, as summarized in Table 2. The results indicate substantial inter-individual variability across age segments. Notably, W_C (mean = 86.86 centimeters (cm), SD = 7.80), H_C (mean = 97.47 cm, SD = 4.88), and A_C (mean = 96.42 cm, SD = 6.69) exhibited considerable dispersion, highlighting the challenge of applying standard sizing systems to this demographic.

Table 2.

Descriptive statistics of anthropometric measurements (n = 209).

Variable	Mean	Min	Max	Range	SD	Variable	Mean	Min	Max	Range	SD
W	62.30	44.50	80.70	36.20	7.44	A_W	31.79	23.90	39.00	15.10	2.46
H	160.90	146.80	173.20	26.40	5.30	H_W	32.86	28.60	39.40	10.80	1.76
W_H	100.71	91.10	112.00	20.90	4.11	W_R	1.24	1.02	1.54	0.52	0.10
A_H	96.18	86.60	105.70	19.10	3.53	A_R	1.31	1.02	1.62	0.60	0.11
H_H	81.97	72.40	90.70	18.30	3.36	H_R	1.54	1.12	1.85	0.73	0.14
C_H	70.14	61.70	78.20	16.50	3.21	H_C--W_C	10.65	−2.10	20.7	22.80	5.12
K_H	43.68	39.00	48.30	9.30	1.71	H_C--A_C	1.08	−7.80	7.60	15.40	3.10
W_C	86.86	68.00	106.80	38.80	7.80	W_C--A_C	−9.57	−15.50	−3.20	12.30	2.39
A_C	96.42	81.40	114.20	32.80	6.69	LV_A	11.33	2.86	19.26	16.40	3.03
H_C	97.47	86.20	110.40	24.20	4.88	LB_A	8.10	−0.90	14.88	15.78	2.72
T_C	54.55	45.60	63.30	17.70	3.50	V_A	20.40	13.95	26.50	12.55	2.82
K_C	36.35	32.10	42.70	10.60	1.88	S_A	15.07	6.99	21.95	14.96	2.98
W_T	21.88	16.50	29.50	13.00	2.86	I_A	21.09	13.68	32.91	19.23	2.88
A_T	24.34	18.50	30.60	12.10	2.48	B_A	4.39	2.11	7.64	5.53	1.05
H_T	21.48	17.00	29.00	12.00	2.31	W_A	7.29	1.87	13.57	11.70	1.68
W_W	27.02	21.00	35.00	14.00	2.54	C_L	76.74	67.20	85.30	18.10	3.47

Note. Min: minimum; Max: maximum; SD: standard deviation; W: weight (unit: kilograms (kg)), length variables are expressed in cm, angular variables in degrees (°), and ratio variables are dimensionless.

Analysis of body shape variability

To further explore the underlying structure of lower body measurements in older women from Northeast China, an initial analysis was conducted on 34 variables. Based on each variable’s practical significance and relevance to body shape characterization, 27 variables were selected for PCA. As shown in Table 3, six principal components were extracted, explaining 24.392%, 19.118%, 16.759%, 8.928%, 7.11%, and 4.656% of the total variance, respectively. The cumulative variance contribution reached 80.963%, indicating that these six components effectively represent the overall structure of the original dataset.

Table 3.

Explained variance and rotated component loadings from principal component analysis.

Component	Initial eigenvalues			Rotation sums of squared loadings
Component	Total	% of variance	Cumulative %	Total	% of variance	Cumulative %
1	7.477	27.693	27.693	6.586	24.392	24.392
2	6.427	23.804	51.498	5.162	19.118	43.51
3	3.323	12.308	63.805	4.525	16.759	60.269
4	1.886	6.987	70.792	2.41	8.928	69.197
5	1.544	5.717	76.509	1.92	7.11	76.307
6	1.202	4.454	80.963	1.257	4.656	80.963

The rotated factor loading matrix is shown in Table 4. Principal component 1 exhibits high loadings on height-related variables, including total height, W_H, A_H, H_H, and C_H, and is defined as the Height Factor. Principal component 2 demonstrates strong associations with girth-related variables, such as A_C, H_C, and T_C, reflecting the Horizontal Girth Factor. Principal Component 3 comprises variables representing differences among the waist, abdomen, and hip regions, including ratio indices and abdominal protrusion angle. Collectively, these features describe the cross-sectional morphology in the torso region. Thus, this component is the Waist–Abdomen–Hip Cross-Sectional Shape Factor. Principal Component 4 features high loadings on hip protrusion angle and torso tilt, characterizing the sagittal body profile—specifically the forward-backward curvature—and is defined as the Lateral Contour Factor. Principal Component 5 consists mainly of the side hip and side abdominal angles, describing the anterior curvature of the lower torso and is referred to as the Frontal Contour Factor. Principal Component 6 includes the B_A and lower hip protrusion, indicating the degree of lower body inclination, and is named the Postural Inclination Factor. Collectively, these six components align well with the anatomical structure of the human body and correspond closely to the 3D spatial design requirements in pants patternmaking—length, girth, and curvature—thereby establishing a robust basis for subsequent clustering analysis.

Table 4.

Rotated component matrix from PCA of 27 lower-body anthropometric variables.

Variable	Component
Variable	1	2	3	4	5	6
A_H	0.982	0.078	−0.023	−0.049	−0.013	0.036
K_H	0.981	0.075	−0.043	−0.060	−0.012	0.038
C_H	0.962	−0.137	0.172	0.061	−0.001	0.002
W_H	0.958	−0.017	0.154	0.088	−0.004	−0.049
H	0.955	0.108	0.203	0.055	−0.010	−0.020
H_H	0.932	0.063	−0.011	−0.092	0.039	0.005
An_H	0.929	0.101	0.227	0.046	0.029	−0.024
K_C	0.075	0.955	−0.162	0.100	0.030	−0.041
T_C	0.061	0.948	−0.093	0.121	0.031	0.070
H_C	0.154	0.897	−0.207	0.021	0.033	0.130
An_C	0.022	0.875	−0.045	0.000	0.049	−0.123
A_C	0.053	0.758	−0.556	−0.109	0.026	0.143
H_C--W_C	0.111	−0.076	0.885	0.358	0.045	−0.027
H_C--A_C	0.135	−0.199	0.878	0.271	−0.004	−0.100
V_A	−0.112	0.060	−0.761	0.286	0.050	−0.019
W_C--A_C	−0.063	−0.095	−0.755	−0.413	−0.101	−0.073
W_C	0.026	0.627	−0.712	−0.221	−0.008	0.101
W_W/W_T	0.097	−0.282	0.581	−0.009	−0.006	0.046
A_W/A_T	0.176	−0.357	0.489	−0.082	0.159	0.113
H_W/H_T	0.132	−0.396	0.416	−0.122	0.090	0.074
S_A	0.008	0.046	0.120	0.925	0.034	−0.005
W_A	−0.014	0.117	0.135	0.915	0.010	0.014
LV_A	−0.015	0.054	0.015	0.036	0.964	−0.038
LB_A	0.020	0.061	0.089	0.016	0.960	−0.014
B_A	−0.053	0.152	0.024	0.005	0.029	−0.787
I_A	−0.137	0.315	−0.205	−0.296	0.044	0.536
H_T/2--A_T/2	−0.001	0.033	0.326	0.210	−0.055	0.496

Cluster analysis of lower body shape in older women

Based on PCA, measurement feasibility, and variables’ representativeness for body shape description, six key morphological features were extracted from six principal components as classification variables. These features included A_H, H_C, H_C--W_C, S_A, LV_A, and B_A. Based on the factor loadings of the principal components and the BIC, a TSC method was employed to categorize the samples into three distinct body types. The resulting model yielded a silhouette coefficient of 0.3. Although this value indicates a moderate level of cluster overlap, it is within an acceptable range for anthropometric studies, particularly given the continuous nature of body shape variation and substantial intra-group heterogeneity among older women. Further analysis of cluster centroids revealed clear distinctions among the three types in terms of H_C--W_C, H_C, and A_H, providing a solid structural basis for developing accurate body type-specific predictive models.

A summary of the clustering results is presented in Table 5. The clusters are ordered from top to bottom by descending proportion, namely Type 1, Type 2, and Type 3, which account for 45.45%, 27.75%, and 26.79% of the total sample, corresponding to 95, 58, and 56 participants, respectively. The importance of the classification variables, ranked in descending order, is as follows: H_C--W_C (1.00), H_C (0.92), A_H (0.63), S_A (0.58), LV_A (0.2), and B_A (0.06). The first four variables contributed significantly more than the others and served as the primary discriminators of lower body shape.

Table 5.

Sample distribution and centroids of clustering variables from lower-body anthropometric analysis.

Cluster type	Sample size	Proportions (%)	Centroid of a clustering variables
Cluster type	Sample size	Proportions (%)	H_C--W_C	H_C	A_H	S_A	LV_A	B_A
1	95	45.45	12.53	94.06	94.38	15.99	11.18	4.52
2	58	27.75	12.96	101.19	99.04	12.78	12.78	4.09
3	56	26.79	5.08	99.59	96.27	12.45	10.12	4.47

To further illustrate the morphological differences between the clustered body types, Figure 4 presents the intermediate shape of three subdivision types. By comparing the morphological characteristics of the three body types, it can be found that there are apparent differences in the front and side morphology of each body type. Based on the centroids of the clustering variables and intermediate shape, the three lower body shape types among older women can be characterized as follows:

Type 1 (Short, slender, with prominent buttocks): This group features the smallest H_C and A_H, but relatively large values for hip-to-waist difference, upper hip protrusion angle, and lateral abdominal angle. The B_A is the highest among all groups. This body shape is petite yet curvaceous, with the buttocks projecting upward and backward, a pronounced waist indentation, and a forward-leaning posture.

Type 2 (Tall, robust, with protruding buttocks and backward tilt): This type exhibits the most significant overall body dimensions. It has the highest measurements for H_C, A_H, hip-to-waist difference, hip protrusion angle, and the smallest body axis angle. These features reflect a typical “large frame with prominent buttocks and a backward-leaning posture.”

Type 3 (Medium height, flat profile, with upright buttocks): Individuals in this group show intermediate values for A_H and H_C but the lowest values for hip-to-waist difference, hip protrusion angle, and lateral abdominal angle. The body contour is relatively flat, the buttocks are not prominently projected, and the standing posture is more vertically aligned.

Figure 4.

Average frontal and lateral morphology of body shape types (Types 1–3).

Regression modeling of various body types

Selection of independent variables

The core objective of body size prediction is to estimate key dimensions that are difficult to measure directly by using a small number of easily accessible body measurements. This approach aims to reveal the quantitative relationships between variables. A reasonable selection of independent variables improves the model’s fitting performance and stability and enhances its practicality and operability. In this study, the selection of independent variables for regression modeling was grounded in the results of PCA, which served as the basis for dimensionality reduction and variable screening. The analysis revealed that components related to girth and height accounted for the most significant proportion of variance, indicating their dominant role in explaining lower body shape differences among older women. H and A_C exhibited high loading values within the rotated component matrix within their respective factors. Given their strong explanatory power and ease of measurement in practical settings, these two variables were identified as the primary predictors in the regression models. Furthermore, W was incorporated as an additional independent variable due to its evident impact on the overall body contour. Correlation analysis confirmed that W was significantly associated with multiple lower body dimensions. As W is also easily measurable and has been repeatedly cited in the literature as a key determinant of body shape, it was included to enhance both the interpretability and predictive performance of the models. To summarize, H, A_C, and W were ultimately selected as the independent variables for developing regression models tailored to each lower body shape category. These variables provide a practical and reliable foundation for predicting key dimensions related to pants pattern design in older women.

Regression modeling

Multiple linear regression models were developed for each of the three identified body shape types to enhance the prediction accuracy of critical lower body dimensions. Using H, W, and A_C as predictors, regression models were constructed for 12 lower body dimensions closely related to pants pattern construction. Taking the W_C model of body Type 1 as an example, the regression coefficients and associated statistics are presented in Table 6, and the regression equation is shown in equation (3):

W_{C} = 0.99 * A_{C} + 0.117 * W - 15.806

(3)

Table 6.

Coefficients of linear regression model predicting W_C^a.

Model	Unstandardized coefficients		Standardized coefficients	t	Sig.
Model	B	Std error	Beta	t	Sig.
(Constant)	−15.806	4.11		−3.846	0
A_C	0.99	0.066	0.846	15.027	0
W	0.117	0.056	0.117	2.075	0.041

Dependent variable: W_C.

To provide a more comprehensive presentation of the modeling results, Table 7 summarizes the regression equations for all predicted dimensions across the three body type categories. These equations highlight how H, W, and A_C contribute differently to specific dimensions in each cluster, reflecting the morphological diversity among older women.

Table 7.

Regression models of lower body characteristics of three lower-body types.

Parameters	Type 1	Type 2	Type 3
W_H	0.762*H--21.675	0.761*H--22.007	0.758*H--21.378
A_H	0.652*H--9.082	0.609*H--1.773	0.545*H+9.222
H_H	0.576*H--11.073	0.544*H--5.604	0.617*H--16.191
C_H	0.611*H--27.748	0.598*H--26.500	0.559*H--20.205
K_H	0.312*H--6.675	0.290*H--2.931	0.270*H+0.583
An_H	0.044*H+0.060	0.045*H--0.100	0.044*H+0.026
W_C	0.99A_C+0.117W--15.806	1.116A_C--0.065W--17.993	0.904A_C+0.107W--4.620
H_C	0.541A_C+0.159W+35.433	0.509A_C+0.184W+38.317	0.493A_C+0.214W+35.115
T_C	0.285A_C+0.242W+12.675	0.244A_C+0.226W+17.756	0.177A_C+0.313W+16.641
K_C	0.117A_C+0.141W+16.596	0.191A_C+0.046W+15.506	0.095A_C+0.156W+17.020
An_C	0.02A_C+0.092W+13.877	0.094A_C+0.012W+11.839	0.018A_C+0.087W+14.038
C_L	0.168H+0.313W+30.391	0.192H+0.226W+31.939	0.443W+0.200H+16.258

Note. Each equation corresponds to one of the three body types identified in the clustering analysis. All outputs in cm; inputs H, A_C in cm; W in kg.

Significance testing of the regression equation

To evaluate the statistical significance of the constructed regression models, an analysis of variance (ANOVA) was conducted for each body type category. As an example, the ANOVA results for the W_C regression model of the Type 1 body shape are presented in Table 8. The model produced an F-value of 344.895 with a significance level of p < 0.001, indicating that the regression equation is statistically highly significant and possesses strong explanatory power for the variation in W_C within this body type group.

Table 8.

ANOVA of the W_C^a regression model.

Model	Sum of squares	df	Mean square	F	Sig.
Regression	2227.816	2	1113.908	344.895	0.000^b
Residual	297.133	92	3.23	—	—
Total	2524.949	94	—	—	—

Dependent variable: W_C.

Predictor variables: (constant), W, A_C.

Evaluation of prediction errors

To evaluate the fitting performance of the regression models, 20% of the samples from each body type were randomly reserved for validation. Tables 9 –11 present a detailed comparison between the predicted values (P) generated by the regression models and their corresponding original measurements (O), with sample IDs indicated numerically. Specifically, 19 samples from Body Type 1 (Table 9), 12 samples from Body Type 2 (Table 10), and 11 samples from Body Type 3 (Table 11) were included in the validation sets. These tables illustrate the closeness of fit between P and O for all predicted parameters, providing direct evidence of model accuracy. The overall prediction errors were further summarized using the RMSE, which is reported in Table 12 and discussed below.

Table 9.

Error comparison between predicted and original data for body type 1 (unit: cm).

Group	W_H	A_H	H_H	C_H	K_H	An_H	W_C	H_C	T_C	K_C	An_C	C_L
P1	99.25	94.39	80.34	69.22	42.84	7.04	76.31	90.71	50.03	34.10	20.40	73.33
O1	98.60	93.60	79.50	68.30	42.40	7.00	76.00	89.70	50.50	34.60	20.80	74.20
P2	104.21	98.63	84.08	73.19	44.87	7.33	75.23	90.46	50.45	34.41	20.69	75.52
O2	104.70	98.60	84.00	73.60	44.80	7.30	74.30	90.10	49.70	32.90	18.80	75.30
P3	99.94	94.98	80.86	69.77	43.12	7.08	75.10	89.95	49.56	33.89	20.34	73.35
O3	101.40	95.00	80.80	71.00	43.10	7.00	76.30	88.80	49.00	33.10	20.40	74.40
…	—	—	—	—	—	—	—	—	—	—	—	—
P18	102.38	97.06	82.70	71.72	44.12	7.22	82.71	94.90	53.39	35.78	21.18	76.30
O18	101.60	97.60	83.20	71.50	44.30	7.20	84.80	93.70	51.40	34.00	19.60	74.70
P19	98.72	93.93	79.94	68.79	42.62	7.01	75.88	90.29	49.48	33.79	20.21	72.59
O19	98.40	92.90	78.90	69.00	42.00	7.00	75.90	91.20	49.80	33.50	20.10	73.10

Table 10.

Error comparison between predicted and original data for body type 2 (unit: cm).

Group	W_H	A_H	H_H	C_H	K_H	An_H	W_C	H_C	T_C	K_C	An_C	C_L
P1	104.32	99.32	84.70	72.77	45.21	7.37	83.98	98.14	55.02	36.52	21.52	77.89
O1	103.10	98.90	84.30	73.30	44.90	7.30	83.90	98.50	56.00	37.50	22.00	74.60
P2	104.32	99.32	84.70	72.77	45.21	7.37	86.17	100.35	56.87	37.22	21.80	79.18
O2	104.90	98.90	84.30	73.00	44.90	7.30	86.80	99.10	56.40	36.90	21.80	79.70
P3	103.63	98.77	84.21	72.23	44.95	7.33	82.42	97.04	54.24	36.15	21.35	77.31
O3	103.60	98.80	84.30	71.80	44.90	7.30	83.20	95.00	54.70	37.10	22.00	78.60
…	—	—	—	—	—	—	—	—	—	—	—	—
P11	104.32	99.32	84.70	72.77	45.21	7.37	82.70	98.09	55.34	36.44	21.45	78.46
O11	105.60	99.00	84.40	73.30	45.00	7.30	84.00	97.00	55.20	36.90	21.90	79.70
P12	105.54	100.30	85.57	73.72	45.67	7.44	85.56	100.25	56.92	37.16	21.76	79.67
O12	104.70	100.10	85.50	73.20	45.60	7.40	87.50	99.90	57.00	35.60	20.00	78.60

Table 11.

Error comparison between predicted and original data for body type 3 (unit: cm).

Group	W_H	A_H	H_H	C_H	K_H	An_H	W_C	H_C	T_C	K_C	An_C	C_L
P1	102.63	98.38	84.75	71.25	44.76	7.22	93.06	98.97	55.63	37.13	21.75	79.10
O1	104.00	98.70	84.30	71.30	44.90	7.20	94.50	98.00	54.70	37.70	22.00	81.00
P2	101.42	97.51	83.76	70.35	44.32	7.15	93.15	99.02	55.64	37.14	21.76	78.78
O2	102.90	97.20	82.90	70.30	44.20	7.10	92.70	100.80	56.40	36.90	20.70	81.40
P3	101.42	97.51	83.76	70.35	44.32	7.15	92.41	98.46	55.21	36.92	21.66	78.34
O3	98.60	93.30	79.00	67.00	42.10	7.10	92.60	97.40	54.80	35.80	20.40	79.10
…	—	—	—	—	—	—	—	—	—	—	—	—
P11	103.62	99.09	85.55	71.97	45.11	7.28	93.60	99.41	56.00	37.32	21.84	79.76
O11	102.50	98.60	84.10	71.00	44.80	7.30	95.40	98.60	55.70	38.30	22.60	79.20
P12	97.17	94.46	80.31	67.22	42.81	6.91	96.27	100.53	55.90	37.29	21.72	77.13
O12	97.50	95.40	81.40	67.60	43.40	6.90	99.60	99.20	57.20	38.80	23.40	78.30

Table 12.

RMSE for 12 lower body dimensions across 3 body types (unit: cm).

Cluster type	W_H	A_H	H_H	C_H	K_H	An_H	W_C	H_C	T_C	K_C	An_C	C_L
1	1.13	0.79	0.79	0.79	0.41	0.07	1.16	1.16	1.46	1.08	0.95	1.58
2	1.46	0.57	0.56	1.07	0.33	0.07	1.11	1.37	1.18	1.00	0.86	1.59
3	1.15	1.62	1.72	1.16	0.87	0.03	1.45	1.20	1.44	1.01	1.02	1.59

The RMSE values between the predicted and actual measurements for the 12 pants-related dimensions across the 3 body types are presented in Table 12. As shown, Body Type 2 exhibited the most stable prediction performance, with all RMSE values below 1.6 cm. Body Type 1 showed comparable results, while Body Type 3 demonstrated the highest prediction errors for variables such as A_H, H_H, and C_L, with the maximum RMSE reaching 1.72 cm.

Discussion

This study aims to establish a lower-body structural size prediction method for older women based on body shape classification. By simplifying the measurement process and enhancing model adaptability, it provides accurate support for pants pattern design. To this end, we combined PCA and TSC to classify the lower-body shapes of older women into three representative categories. Subsequently, we built separate linear regression models for each group to predict 12 key structural dimensions.

In this research, the use of TSC not only enables automatic determination of the optimal number of clusters but also handles complex situations involving both continuous and categorical variables. Compared with the K-means clustering method employed by Lee et al.³¹ in male obesity samples, this approach is more suitable for capturing the diverse body shapes of the elderly population. Meanwhile, the principal components extracted through PCA align with the “height- and girth-dominated classification logic” emphasized by Liu et al.³² in their study on young female samples. Furthermore, our study identifies additional posture-related factors such as pelvic tilt and body axis orientation within the older sample, thereby extending the structural dimensions considered in body shape classification. The strategy of “classification followed by modeling” not only improves the adaptability of predictive models but also echoes a broader technical pathway increasingly adopted in apparel engineering research. For instance, Wang et al. used clustering to generate lower-body shape prototypes as a basis for structural surface construction, while Jin et al. applied shape classification to enhance the design process of female apparel blocks.^29,33 Although the specific techniques vary, these studies consistently emphasize the role of classification as a prerequisite to modeling, facilitating enhanced interpretability and adaptability across diverse body types. Building upon this foundation, the present study further demonstrates the engineering potential of this strategy by enabling prediction based on only three easily obtainable measurement variables.

The results demonstrated that satisfactory prediction accuracy was achieved across all body shape types, although the model performance varied among different categories. Evaluated using RMSE as the metric, Type 2 (tall, full-hipped) demonstrated the most accurate predictions, with the lowest RMSEs across most of the 12 key dimensions. This may be attributed to this group’s higher internal shape consistency and more distinct morphological features favoring linear regression modeling. Type 1 (short, curvy) also exhibited stable and accurate results, suggesting that linear models can effectively accommodate pronounced body contours and provide appropriate pattern designs for this body type. In contrast, Type 3 (medium height, straighter shape) showed larger prediction errors in variables such as A_H, H_H, and C_L, with the maximum RMSE reaching 1.72 cm. This indicates that linear models have limitations when dealing with smoother contours or more complex body structures, which may result in decreased accuracy in pattern design. Moreover, the notable differences in prediction error among body types suggest that factors such as fat distribution and posture variability may influence the accuracy of size prediction. These findings are consistent with the previous research by Yan et al.³⁴ which also reported reduced prediction accuracy in atypical body shapes. Notably, the constructed regression models rely solely on 3 easily accessible variables—H, W, and A_C—to predict 12 key structural dimensions essential for pants design. Most of the predicted values achieved a RMSE within ±2 cm, demonstrating a sound balance between prediction accuracy and engineering feasibility.

Despite the encouraging results, some limitations remain. First, the study sample was limited to older women from Northeast China, which may not fully represent the body characteristics of the national population or other ethnic groups, thus limiting the model’s generalizability. Second, the current regression models are linear, which may constrain their performance when predicting sizes for individuals with less distinct contours or more complex body structures. Third, although the selected input variables are easy to obtain and facilitate practical application, they fail to capture other potentially influential features of lower-body structure, such as hip width, leg length, and pelvic orientation. This omission may reduce the model’s predictive accuracy for atypical or transitional body types.

Conclusion

This study addresses the challenges of body shape diversity and poor sizing compatibility in traditional garments for older women by developing a predictive framework for pants structural dimensions based on a classification followed by modeling strategy. A total of 217 older women from Northeast China were recruited, combining 3D body scans with manual anthropometric measurements to establish a systematic pipeline from body type classification to size prediction. Six principal components were extracted through PCA, representing critical morphological factors related to height, girth, and body shape features. Based on these components, a TSC algorithm was employed to categorize the sample into three representative body types. Separate stepwise linear regression models were then constructed for each body type, using only 3 easily obtainable input variables—H, W, and A_C—to predict 12 pants construction-related dimensions, including W_H, H_H, A_H, H_C, and C_L.

Model evaluation showed that most predicted variables across the three body types achieved prediction errors within ±2 cm, demonstrating strong practical applicability. In practical apparel design, a circumference difference of 1.5–2.5 cm is generally regarded as an acceptable tolerance range.³⁵ Accordingly, the prediction errors obtained in this study fall within this reasonable industry allowance, which can effectively reduce fit problems such as an overly tight waistband or excessive looseness at the hips, thereby underscoring the practical applicability of the proposed method. The predicted dimensions can be directly applied as inputs for pants pattern construction. The proposed framework simplifies the modeling process and reduces both labor and technical barriers in pattern design. The findings of this study not only provide reliable input dimensions for pants pattern construction but also offer data support for optimizing pants structures and enabling personalized design across different body shape groups.

Future research can proceed in several directions: (1) expanding the sample size and regional coverage to include elderly women from a broader range of areas with more diverse body shapes, thereby enhancing the representativeness and applicability of the model; (2) exploring nonlinear modeling approaches (e.g. SVR, ANN, and ensemble learning), in combination with transfer learning or few-shot learning strategies, to improve the model’s performance in handling complex or transitional body shapes; and (3) integrating 3D anthropometric parameters with image recognition techniques to investigate the potential of multimodal inputs for apparel pattern modeling.

Footnotes

ORCID iDs

Jianmei Sun

Tong Yao

Jun Wang

Baixue Zhou

Li Pan

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 Liaoning Provincial Department of Education scientific research project (JYTMS20230401).

Declaration of conflicting interests

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

References

Peng

On the Chinese approach to actively addressing population ageing. Soc Sci China 2022; (9): 46–66.

Bloom

DE.

Macroeconomic implications of population ageing and selected policy responses. In: United Nations Expert Group Meeting on social and economic implications of changing population age structures, Mexico City, 31 August–2 September 2005.

Rosenblad-Wallin

Karlsson

Clothing for the elderly at home and in nursing homes. J Consum Stud Home Econ 1986; 10: 343–356.

Lee

Park

Inclusive design process for user-centered wearable products: a case study on lumbar support for older women with chronic low back pain. Cloth Text Res J 2025; 43(4): 322–338.

Lee

Damhorst

Lee

, et al. Older women’s clothing fit and style concerns and their attitudes toward the use of 3D body scanning. Cloth Text Res J 2012; 30: 102–118.

Brownbridge

Gill

Grogan

, et al. Fashion misfit: women’s dissatisfaction and its implications. J Fash Mark Manag 2018; 22: 438–452.

Chan

MHE

Limbu

KD.

Wearable solutions: addressing dressing challenges among Hong Kong’s elderly through adaptive clothing design. Cloth Text Res J 2024; 42(2): 123–137.

Simmons

Istook

CL.

Body measurement techniques: comparing 3D body scanning and anthropometric methods for apparel applications. J Fash Mark Manag 2003; 7: 306–332.

Bolaji

Qiu

McQuerry

, et al. A comparative anthropometric analysis of US female firefighters versus the general female population. J Safety Res 2025; 93: 206–213.

10.

Xia

Guo

, et al. Comparison of different body measurement techniques: 3D stationary scanner, 3D handheld scanner, and tape measurement. J Text Inst 2019; 110: 1103–1113.

11.

Choi

Ashdown

SP.

3D body scan analysis of dimensional change in lower body measurements for active body positions. Text Res J 2011; 81: 81–93.

12.

Tan

Lin

Wang

Cluster size intelligence prediction system for young women’s clothing using 3D body scan data. Mathematics 2024; 12: 497.

13.

Calculating method of characteristics girth of young female body by 3-D scanning data. J Text Res 2017; 38: 110–114.

14.

Zanwar

Shukla

, et al. Prediction of anthropometric dimensions using multiple linear regression and artificial neural network models. J Inst Eng India Ser C 2023; 104: 307–314.

15.

Meyer

Birregah

Beauseroy

, et al. Missing body measurements prediction in fashion industry: a comparative approach. Fash Text 2023; 10: 37.

16.

Liu

Morris

, et al. An exploratory study of body measurement prediction using machine learning and 3D body scans. Cloth Text Res J 2025; 43(3): 187–203.

17.

Huang

Shen

Shi

, et al. Artificial intelligence-based precise prediction of anthropometric data for female garment pattern-making. J Text Inst 2025; 116: 733–741.

18.

Yan

, et al. Predicting human dimensions in body images for automatic generation of female pants patterns. Text Res J 2019; 89: 3792–3801.

19.

Pan

Yao

, et al. Classification and regression modeling of breast shape of young females in northeast China. Silk 2022; 59: 64–70.

20.

ISO. ISO 20685-1:2018. 3-D scanning methodologies for internationally compatible anthropometric databases – Part 1: evaluation protocol for body dimensions extracted from 3-D body scans. Geneva, Switzerland: International Organization for Standardization, 2018.

21.

Standardization Administration of China. GB/T 16160-2017: human body measurements for garment use—definitions and measuring methods. Standards Press of China, 2017.

22.

Jolliffe

Cadima

Principal component analysis: a review and recent developments. Phil Trans R Soc A 2016; 374: 20150202.

23.

Marbac

Sedki

Variable selection for mixed data clustering: a model-based approach. arXiv preprint, 2017. DOI: 10.48550/arXiv.1703.02293.

24.

Fonseca

JRS

Cardoso

MGMS

. Mixture-model cluster analysis using information theoretical criteria. Intell Data Anal 2007; 11: 155–173.

25.

Benassi

Garofalo

Ambrosini

, et al. Using two-step cluster analysis and latent class cluster analysis to classify the cognitive heterogeneity of cross-diagnostic psychiatric inpatients. Front Psychol 2020; 11: 1085.

26.

CH.

Exploratory data analysis in the context of data mining and resampling. Int J Psychol Res 2010; 3: 9–22.

27.

Hong

Tian

Identification of unknown abnormal conditions of absorption stabilization system in catalytic cracking process based on cyclic two-step clustering analysis and convolutional neural network. Processes 2023; 11: 1443.

28.

Chi

Zeng

Bruniaux

, et al. An intelligent recommendation system for personalised parametric garment patterns by integrating designer’s knowledge and 3D body measurements. Ergonomics 2025; 68(3): 317–337.

29.

Jin

Individualized generation of women’s prototype based on the classification of body shape. Int J Ind Ergon 2024; 103: 103631.

30.

Tian

Zhang

Wang

Effects of wind speed and clothing wearing method on total thermal insulation of down jackets. AATCC J Res 2025; 12: 24723444251314917.

31.

Lee

Song

Kim

Categorization of lower body shapes of abdominal obese men using a script-based 3D body measurement software. Fash Text 2020; 7: 3.

32.

Liu

Wang

Tao

, et al. Fuzzy classification of young women’s lower body based on anthropometric measurement. Int J Ind Ergon 2016; 55: 60–68.

33.

Wang

Pan

, et al. Parametric 3D modeling of young women’s lower bodies based on shape classification. Int J Ind Ergon 2021; 84: 103142.

34.

Yan

Y-J

Zhong

A-H

J-W.

Classification and regression modelling of abdominal convexity morphology in middle-aged and elderly women. J Fiber Bioeng Inform 2023; 16: 45–55.

35.

Petrova

Ashdown

SP.

Three-dimensional body scan data analysis: body size and shape dependence of ease values for pants’ fit. Cloth Text Res J 2008; 26: 227–252.

Size prediction of older women’s lower body based on shape classification

Abstract

Keywords

Introduction

Methods

Data collection

Data preprocessing

Extraction of principal components

Cluster analysis

Regression modeling

Results

Statistical analysis of body shape characteristics in older women

Analysis of body shape variability

Cluster analysis of lower body shape in older women

Regression modeling of various body types

Selection of independent variables

Regression modeling

Significance testing of the regression equation

Evaluation of prediction errors

Discussion

Conclusion

Footnotes

ORCID iDs

Funding

Declaration of conflicting interests

References