An exploratory, observational, cross-sectional pilot study was conducted between November 2021 and March 2022 at KK Women’s and Children’s Hospital, Singapore. The study was prospectively registered at ClinicalTrials.gov (NCT05079776).

Eligible participants were children aged between 0 and 18 months whose parents (1) had a smartphone or tablet with access to the internet, (2) were able to complete the study questionnaires, and (3) took and uploaded images onto an online form. Children who were unable to undergo length measurement by the standardized technique recommended by the WHO [23] (eg, children with structural abnormalities of the lower limbs or orthopedic conditions such as club foot and hip dysplasia) were excluded from the study.

The study duration was a maximum of 2 days. On day 1 (clinic setting), investigators measured the participant’s body length using the standardized WHO length measurement technique [23]. They then used a smartphone to take 6 top-view photos of the participant in a supine position. Each photo included a standard-size reference card. On day 1 or 2 (home setting), parents took and uploaded 6 smartphone photos of the participant in a supine position with the reference card. Investigators and parents were given a list of image quality requirements and guidelines for capturing good-quality images. Parents and investigators completed their respective user experience questionnaires after the image upload process.

KK Women’s and Children’s Hospital’s independent ethics committee approved the study before its initiation (approval no.: 2021/2540). The study was conducted in accordance with good clinical practice, the Declaration of Helsinki, and the local laws and regulations of Singapore. Written informed consent was obtained from the parent(s) of each participant before any study-related activities were undertaken. The participants’ parent(s) received vouchers (equivalent to 30 Singapore dollars, or US $22.32) as a token of appreciation for participating in the study. Participant data were pseudonymized for analysis.

The LAI algorithm uses state-of-the-art imaging and machine learning techniques to estimate a participant’s length from a single image, such as a smartphone photo. The current algorithm was designed to predict the length of children up to 18 months. The input to the algorithm is a digital image of the participant in a supine position and a reference object (standard size card, 85.6 mm by 54.0 mm). The first step involves extracting image features for both the participant and the reference card (Figure 1). Landmark extraction models are used to detect landmarks on the participant’s face and body (shoulders, hips, knees, ankles, heels, etc) within the image. These estimate the length of individual body segments in pixels. The card detection and card segmentation models are used to locate the reference card in the image and compare its pixel dimensions against its known physical dimensions to generate a pixel per metric value. The feature extraction step thus generates a set of quantitative features used in the length prediction step (Figure 1). A model incorporating these features predicts the total body length in millimeters. The algorithm returns a predicted length value as output only if the key feature extraction steps (body and card features) are successful.

The investigator and parent datasets consisted of all images taken by the investigators and the parents, respectively.

To maximize the number of images usable for length prediction, photo-taking guidelines were given to investigators and parents. This included no clothing on the head or feet (eg, cap, socks, etc); no loose or baggy clothing to ensure that the body contour was visible; high contrast between the background, participant, and reference card; participant and reference card placed on the same flat and stable surface; participant positioned not more than 10 cm from the reference card; taking the image at an angle of 90 degrees from the surface on which the participant and card were placed; and participant’s legs not bent with the entire body visible to the camera. Images that fail to meet these requirements, such as those shown in Figure 2A-F, may fail to generate length predictions.

Following further testing and optimization, automated flags (warnings) were incorporated into the LAI (Figure 3). This allows the tool to detect uploaded images that do not meet the specified requirements and warn users that length prediction may be unsuccessful. In total, 4 different types of warnings were implemented.

The LAI model was trained on the set of images collected by investigators (investigator dataset) and the investigator-generated length measurements. As described above, a set of features was generated, along with warnings. Only images that did not produce warnings were used to train the model. A 5-fold cross-validation was performed. Cross-validation is a statistical resampling method commonly used in applied machine learning to evaluate model performance with limited datasets [24,25]. The dataset was divided into 5 folds (subsets of data used for model training and testing) using 3 criteria: each participant’s images should appear in only 1-fold; similar numbers of participants and images per fold; and similar distribution of measured body lengths across folds. In addition, details are provided in the Cross-Validation Procedure section in Multimedia Appendix 1. A bagged model was used for prediction. Hyperparameter optimization (HPO) incorporated a random feature selection step and was implemented using a state-of-the-art HPO framework [26]. The training was performed using a 2-step process. In the first step, a model was trained on all images from the investigators in the current training fold. A subset of these images within the 90th percentile of the training error was then selected, and the model was retrained on these images to ensure that outliers did not affect the training. Finally, all test images (from the investigators and parents) without warnings within the current fold were used to predict and calculate validation errors. The average validation error from all folds was used to drive the HPO framework to find the best model.

Performance metrics were calculated on a per-image and per-participant basis. For a given image i, the body length (p_i) predicted by the trained model was compared with the corresponding participant’s WHO-standardized length measurement (m_i) to derive the following performance metrics: error (E_i, cm: difference between measured and predicted length E_i = [m_i – p_i]); absolute error (AE_i, cm: absolute value of the error AE_i = E_i); and absolute percentage error (APE_i, %: AE as a percentage of the length measurement, ). Bias (average of E, cm; ), mean AE (MAE; average of AE, cm), mean APE (average of APE, %), percentages at each AE cutoff (≤1 cm, ≤2 cm, ≤5 cm, and ≤10 cm), and percentages at each APE cutoff (≤2%, ≤5%, ≤10%, and ≤20%) were also calculated. Missing values due to errors, where the model did not return a predicted length value, were counted and reported separately. For participants with successful length predictions based on at least 9 images, the predictions for each image were averaged to generate a single predicted length for that participant.

The performance of the LAI model was evaluated using 5-fold cross-validation performed on the combined investigator + parent datasets and investigator-generated length measurements. To assess the LAI algorithm’s performance relative to measurements in general clinic settings [8,22], the appropriate performance metrics were compared with published values for the technical error of measurement (TEM), an index commonly used in anthropometry to assess the accuracy and reliability of measurements [19-21].

Due to the exploratory nature of this study, there was no formal sample size calculation. It was estimated that a complete dataset from 200 participants (standardized length measurements, images taken by investigators, images taken by parents, and completed questionnaires from investigators and parents) would allow model performance to be adequately assessed. Assuming a 20% dropout rate, 250 participants were planned for enrolment.

Descriptive statistics were used to summarize participant characteristics and user experience questionnaire responses. Continuous variables were summarized using mean, median, and minimum and maximum values. Discrete variables were summarized using percentages and frequencies by category, including missing values. No statistical testing of formal hypotheses was conducted.

In total, 215 participants were enrolled in the study, of whom 50.7% (n=109) were female. The mean age was 6.1 (SD 5; range: 0.0-17.7) months and the median age was 4.4 months. All participants completed the clinic-based data collection procedures, and 200 participants completed both clinic-based and home-based data collection procedures to provide a complete dataset of length measurements, images, and questionnaires.

A total of 2490 images were taken and uploaded (1290 images by 8 investigators at the clinic and 1200 images by 200 parents at home; Figure 3). Of these, 89.3% (2224/2490) of images were analyzed, and 10.7% (266/2490) were excluded due to image quality or protocol deviations.

The LAI produced a length prediction for 2211 (99.4%) out of 2224 images (Figure 3). In total, 0.6% (13/2224) of images did not produce a prediction due to either unsuccessful pose segmentation (7 images) or unsuccessful card segmentation (6 images). For the set of 1632 high-quality images (those that did not generate any warnings; Figure 3), the bias (mean error) for individual image predictions was minimal (0.03 cm; Table S1 in Multimedia Appendix 1). Most of the length predictions for these individual images (1557/1632, 95.4%) were within 10% of the measured length (Figure 4 and Table S1 in Multimedia Appendix 1). We found that length prediction was improved by averaging over multiple images for a participant. For 88 participants who had predictions for ≥9 images, the majority of these averaged length predictions (71/88, 81%) were within 5% of the measured length (Figure 4 and Table S1 in Multimedia Appendix 1).

The overall distributions of errors for individual image predictions and participant-averaged predictions are illustrated in Figure 5 [8,19,20,22]. Published interobserver TEM ranges for length measurements from the WHO Multicenter Growth Reference Study (0.48 cm [19], 0.70 cm [20]) and general clinics or community health settings (1.41 cm [22], 1.25-1.59 cm [8]) are indicated on the figure for comparison. The MAE for individual image predictions was 2.47 cm, and the MAE for participant-averaged predictions was 1.77 cm, which approaches the TEM range reported in general clinics or community health settings [8,22].

A quarter of the images with successful length predictions (579/2211, 26.2%) generated at least 1 warning (Figure 3). For this study dataset, the most common warning was improper card segmentation. The numbers of images affected by each type of warning are shown in Figure 3. Figure 6 illustrates the percentages of images available for length prediction under scenarios where different warning types are ignored, and the corresponding MAE values. The length prediction workflow can be adjusted to use more or less stringent settings, which affects the number of images retained for prediction and prediction error. Retaining fewer but higher quality images (fewer warnings) for prediction resulted in smaller MAE values (2.47 cm on images without any warnings); conversely, ignoring more warning types allowed more images to be used, but led to an increased MAE (3.39 cm using all images regardless of warnings). Similarly, the MAE for participant-averaged predictions decreased from 2.48 cm (n=155) to 1.77 cm (n=88) when only high-quality images without warnings were used for prediction (Table S1 in Multimedia Appendix 1).

In most cases, investigators and parents reported that they did not find it difficult to capture the required images (Figure 7). Investigators and parents rated the photo-taking process as very easy, easy, or normal for most participants (182/215, 84.7% participants, and 144/200, 72% participants, respectively).

Length measurement is challenging in clinical practice, especially among children younger than 2 years of age, where inaccuracy and interobserver variability in length measurements have been documented in primary care and community health settings [5,8,27]. Without proper equipment and training, it is impractical to expect parents and caregivers to take accurate and reliable length measurements at home regularly. Approaches for computer-assisted body length estimation from digital images have been proposed but involve elaborate requirements for data capture or procedures that may cause disturbances to the child [17,18] and limit their real-world feasibility and applicability. In contrast, the LAI algorithm was designed to have minimal requirements in terms of equipment and user training: only a smartphone and a readily available reference object (standard-size credit card) are required for data collection.

In our study, both parents and investigators found it easy to take the required photos for automated length estimation. Results from this proof-of-concept study suggest that the LAI algorithm, when trained on high-quality length measurements and image data, can predict children’s body length with minimal systematic error and accuracy approaching that achieved in general clinics or community health settings. This indicates the potential for future implementations of the LAI algorithm to make growth monitoring more accessible to nonexpert users. The MAE for individual image predictions was 2.47 cm, and our results suggested that length prediction could be improved by averaging across multiple images of a child: participant-averaged predictions had a smaller MAE of 1.77 cm, which approaches the published interobserver TEM range (1.25-1.59 cm) reported in general clinics and community health settings [8,22]. Thus, the LAI algorithm would enable parents to estimate their child’s length using images captured using their smartphones, with accuracy comparable to length measurements in general clinics or community health settings. Published reports indicate that the level of accuracy considered acceptable ranges from 0.5 cm to 2.0 cm, depending on the clinical indication for which length measurements are required [5,6,9,19-22,27]. For clinical growth monitoring, the WHO Multicenter Growth Reference Study protocols recommend interobserver differences of <0.7 cm for length [20], and the Standardized Monitoring and Assessment of Relief and Transitions (SMART) manual indicates an acceptable limit of <0.5 cm for height [28]. Growth measurement studies commonly report differences of ≤0.5 cm for expert anthropometrists [19-21,29]. This is consistent with the views of the study investigators, who indicated that they considered an accuracy of within 1 cm acceptable in an expert clinical setting.

Although smartphones greatly facilitate data collection for LAI algorithm, it is not always possible to consistently capture high-quality images of a child, which affects the accuracy of predictions. Commonly encountered image quality issues include the visibility of body landmarks obscured by clothing or limb position (Figure 2). Within the current LAI algorithm, we implemented a system that detects issues with uploaded images and generates warnings that these could affect length prediction (Table S1 in Multimedia Appendix 1). The settings can be tuned to ignore warnings and allow more images to be used for length prediction but at the cost of generating less accurate predictions. We envision that future versions of the tool will be integrated into a smartphone application that offers real-time feedback to guide parents or caregivers in positioning the baby and the standard-size reference object. Such feedback, including alerts about potential image quality issues and guidance on how to avoid or reduce key sources of error, would help users to take photos that are more likely to result in accurate predictions.

Other technologies, such as depth-of-field sensors available on some consumer devices, could potentially be integrated to further improve performance. Time-of-flight or light detection and ranging sensors can provide depth-of-field information that could be used with the current image-based algorithm to increase length prediction accuracy. Relevant practical considerations include the availability and accuracy of these technologies, especially for smaller participants like infants.

This proof-of-concept study collected images and corresponding length measurements and used these to explore the feasibility of a digital tool that can be used in home environments and in clinics to monitor growth over time. Due to the limited data available, a cross-validation approach was adopted to maximize the information that we could obtain about the tool’s performance. This lack of an external validation sample may introduce limitations such as underestimation of error when the model is applied to new datasets, and systematic errors if the model’s assumptions do not hold on to other datasets. Although care was taken to ensure that the reference measurements were as accurate as possible by using a standardized protocol (WHO method) and trained personnel, we have not formally accounted for the possibility of human error in the length measurements used for training.

Findings from our study have been used to refine the tool further, and projects to evaluate the feasibility and acceptability of the improved version are ongoing. At present, the proof-of-concept results suggest that the LAI algorithm’s current performance may be compatible with personal use, such as general growth tracking at home, as its performance approaches that of manual length measurement in general clinics or community health settings [8,22]. Using the LAI algorithm, parents could record length measurements more frequently and conveniently at home. This idea is consistent with feedback from parents of the children in our study. Most parents reported that they did not measure their child’s length at home; on the other hand, they indicated that an automated measurement tool was desirable and would be used at least monthly or more frequently. Besides automated length measurement, parents wanted a tool capable of estimating other anthropometric parameters and allowing them to track their child’s growth with reference to WHO growth charts [3]. This user feedback provides valuable insights that can guide future development of the tool.

One of the main strengths of the current LAI approach is its simplicity and practicality for nonexpert end users. Data collection requires no specialized equipment or training; physical discomfort and disturbance to the child are minimized. This innovative approach explored the feasibility of image-based automated body length estimation that can be conveniently performed in a wide range of environments by any user. The potential value of such tools to nonexpert users is underscored by the range of studies that explore the possibilities of digital technology-assisted anthropometry [30,31]. The feedback collected from parents and clinicians in this study will inform future versions of the tool to better cater to the unique requirements of different users. It should be noted that the LAI algorithm’s performance was evaluated only using data from healthy children (those without known growth-related conditions). Further studies with different populations will be needed to guide the design and optimization of the LAI algorithm for use in more specialized clinically oriented tasks such as monitoring for abnormal growth.

It should be noted that this was a proof-of-concept study to demonstrate the feasibility and acceptability of the LAI tool for estimating body length in young children (aged <18 months) in the home and similar environments. The performance achieved to date suggests that the current version of the LAI algorithm would not replace the standard clinical method used by health care professionals, but there is potential for future development to enhance its accuracy and applicability. Although the tool does not currently meet the requirements for highly accurate measurement in specialist clinical applications, our findings suggest that improving image quality is one way to increase length prediction accuracy. Integration of other technologies, such as depth-of-field sensors available on some consumer devices, could be explored to further improve performance. The current performance of the LAI algorithm, coupled with its ease of use, suggests it has the potential to be a feasible method of measuring a child’s length for use by parents with accuracy approaching that of general clinic or community health settings.