Exploring the Use of a Length AI Algorithm to Estimate Children's Length from Smartphone Images in a Real-World Setting: Algorithm Development and Usability Study

Abstract

Background: Length measurement in young children younger than 18 months is important for monitoring growth and development. Accurate length measurement requires proper equipment, standardized methods, and trained personnel. In addition, length measurement requires young children's cooperation, making it particularly challenging during infancy and toddlerhood.

Objective: This study aimed to develop a length artificial intelligence (LAI) algorithm to aid users in determining recumbent length conveniently from smartphone images and explore its performance and suitability for personal and clinical use.

Methods: This proof-of-concept study in healthy children (aged 0-18 months) was performed at KK Women's and Children's Hospital, Singapore, from November 2021 to March 2022. Smartphone images were taken by parents and investigators. Standardized length-board measurements were taken by trained investigators. Performance was evaluated by comparing the tool's image-based length estimations with length-board measurements (bias [mean error, mean difference between measured and predicted length]; absolute error [magnitude of error]). Prediction performance was evaluated on an individual-image basis and participant-averaged basis. User experience was collected through questionnaires.

Results: A total of 215 participants (median age 4.4, IQR 1.9-9.7 months) were included. The tool produced a length prediction for 99.4% (2211/2224) of photos analyzed. The mean absolute error was 2.47 cm for individual image predictions and 1.77 cm for participant-averaged predictions. Investigators and parents reported no difficulties in capturing the required photos for most participants (182/215, 84.7% participants and 144/200, 72% participants, respectively).

Conclusions: The LAI algorithm is an accessible and novel way of estimating children's length from smartphone images without the need for specialized equipment or trained personnel. The LAI algorithm's current performance and ease of use suggest its potential for use by parents or caregivers with an accuracy approaching what is typically achieved in general clinics or community health settings. The results show that the algorithm is acceptable for use in a personal setting, serving as a proof of concept for use in clinical settings.

Trial registration: ClinicalTrials.gov NCT05079776; https://clinicaltrials.gov/ct2/show/NCT05079776.

Keywords: AI; algorithm; artificial intelligence; children; computer vision; height; imaging; infant; length; length estimation; mHealth; measure; mobile health; mobile phone; neonatal; newborn; pediatric; smartphone; smartphone images.

Figure 1

LAI algorithm overview. From an input image of a child in a supine position and a standard reference object, anthropomorphic landmarks of the body and face are extracted, along with the detection and segmentation of the reference object (a standard size card, 85.6 mm by 54.0 mm). These are used by the LAI algorithm to predict the length of the child. LAI: length artificial intelligence.

Figure 2

Challenges with image-based length prediction. Images that pose challenges for length prediction by the LAI algorithm include those where (A) the camera is not positioned perpendicularly (90-degree angle) above the participant during image capture, (B) the participant and reference card are not placed on a flat horizontal surface, (C) there is blurring or glare, (D) there are baggy clothes on participant affecting the visibility of body contour, (E) there is low contrast of the participant with background, and (F) the face or body is not fully visible. LAI: length artificial intelligence.

Figure 3

Schematic diagram of image flow. A total of 2490 images were collected in this study and 2224 images were analyzed. A total of 266 images were not analyzed due to protocol deviations (resubmission of images, submission of images outside the stipulated visit window, or images that did not meet the requirements). Of 2224 images analyzed, 2211 images produced a length prediction. The algorithm did not produce a prediction for 13 images due to unsuccessful pose segmentation (m=7) or unsuccessful card segmentation (m=6). High-quality images refer to images that did not generate any warnings. m (%): number and percentage of images in the specified category.

Figure 4

Scatter plot depicting length predictions made by the model versus gold-standard length measurements made by the investigators. For length predictions on individual images, the majority fell within 10% of the participant’s measured length. For averaged length predictions (per participant, for participants who had predictions from ≥9 images), the majority fell within 5% of the measured length. Blue circles represent predictions from all individual images. Red squares represent averaged predictions for children who had predictions from at least 10 images. The thick black line indicates the ideal prediction (ie, length prediction equal to the measured length). Dashed lines represent 5% and 10% deviations from the ideal prediction.

Figure 5

The overall distribution of errors (residuals) for individual image predictions (blue dots) and participant-averaged predictions made by the model. These were presented alongside published interobserver TEMs of “gold standard” length measurements from WHO (0.48 cm and 0.70 cm) and general clinics or community health settings (1.41 cm and 1.25-1.59 cm). The MAE of individual image predictions was 2.47 cm. When averaged, the predictions had an MAE of 1.77 cm, which approaches the TEM range reported in general clinics. MAE: mean absolute error; TEM: technical error of measurement; WHO: World Health Organization.

Figure 6

A plot illustrating the percentage of images available for length prediction and the corresponding MAE under varying scenarios where combinations of different warnings were ignored. Ignoring more warning types allowed more images to be used but yielded less accurate length predictions. Warnings: W0, card and body overlapping; W1, incorrect card aspect ratio; W2, improper card segmentation; W3, width, and height pixel size are too different. MAE: mean absolute error.

Figure 7

Ease of collecting images as rated by investigators and by parents. N represents the number of participants for which the investigator or parent provided a rating.