AI based sagittal spinal posture assessment for adolescent screening in low resource school settings

Abstract

Adolescent spinal postural deviations are rising across Southeast Asia, driven by prolonged screen exposure and limited access to clinical screening in low-resource environments. To address this gap, we developed PostureGuard, a low-cost AI system that provides real-time posture assessment and school-based screening support. The system incorporates a Southeast Asia–calibrated pose estimation model, benchmarked against radiographic references and validated using wearable sensor measurements. In a field deployment involving 200 students across 15 Indonesian schools, PostureGuard achieved a mean absolute error below 2.3° for forward head angle estimation. We further identified a dose–response association in which each additional hour of daily screen time was associated with an estimated 2.1° increase in forward head angle. A preliminary six-month school implementation suggested improvements in posture-related indicators; however, these observations were not derived from a controlled clinical trial. PostureGuard has since been piloted within Indonesia’s 2024 Digital Wellness Mandate initiative. This work presents a regionally calibrated AI-based posture screening framework with clinical benchmarking and demonstrated scalability for adolescent spinal health monitoring in resource-limited settings.

Introduction

The COVID-19 pandemic accelerated digital immersion among Southeast Asian adolescents, with Indonesia reporting that 89% of students now exceed six hours of daily screen exposure¹. School-based and primary-care surveillance programs have documented a substantial rise in posture-related complaints, including a 237% increase in reported cases of suspected forward head posture and thoracic kyphosis between 2020 and 2023². These conditions, which were previously uncommon in younger populations, are now increasingly identified through routine screening and preliminary imaging referrals among adolescents aged 12 years and older³, highlighting an emerging public-health concern.

The clinical urgency

Recent biomechanical studies confirm that every centimeter of forward displacement increases spinal load by 4.5 kg⁴. For Indonesian children averaging 8 h/day of device use, this translates into disc compression forces equivalent to carrying 20–25 kg on the cervical spine—daily. Such findings led the Indonesian Ministry of Health to declare digital posture dysfunction a national priority under its 2024 Digital Wellness Policy⁵. Radiologists and physiotherapists are calling for early screening protocols to prevent long-term disability in an entire generation. Building on our previous context-aware segmentation architecture⁶, we designed PostureGuard with an embedded spatial-awareness layer to refine the angle estimation process in real-time environments."

The technological gap

Despite the rise of AI in healthcare, existing posture detection systems fall short in three critical areas:

• (1) Morphological Bias: Most pose estimation models (e.g., OpenPose, MediaPipe) were trained on Western anthropometric datasets, leading to a 32% misclassification rate in Southeast Asian populations⁷.

• (2) Clinical Disconnect: No prior AI system has validated its outputs against gold-standard biomechanical metrics such as the Cobb angle or radiologist-graded X-rays⁸.

• (3) Deployment Barrier: High-end motion capture tools remain unaffordable for 85% of Indonesian public schools⁹, leaving millions of students unscreened.

Our solution: PostureGuard

We present PostureGuard, an AI–clinical hybrid system designed for scalable and low-cost posture screening in resource-constrained settings. The system integrates (a) MediaPipe-based pose estimation with Southeast Asian–specific anthropometric calibration, (b) wearable inertial sensors for real-time validation, (c) radiologist-graded radiographs as clinical reference data, and (d) a smartphone-based implementation to support field deployment.

Following the system overview shown in Fig. 1, existing posture-assessment solutions were reviewed to position PostureGuard within the current landscape of AI- and sensor-based screening approaches. The comparative analysis highlights differences in clinical validation, deployment feasibility, cost, and population fit. These distinctions emphasize the absence of scalable, low-cost, region-calibrated systems that are suitable for real-world implementation in school-based settings.

Fig. 1

Workflow of the PostureGuard system showing input capture, pose estimation, Southeast Asian anthropometric calibration, angle computation (FHA, TKA, PTA), and final risk classification output.

Research gap & objectives

Previous AI-based posture systems focused largely on sports performance or elderly rehabilitation (Han et al., 2021), with little to no validation for use in pediatric populations of the Global South. Our research addresses this critical gap by introducing an integrated, clinically validated system for youth screening.

This study has four objectives:

• (1) To develop an AI posture detection system adapted to Southeast Asian anthropometry.

• (2) To correlate screen time with validated postural deviation metrics (FHA, Cobb angle).

• (3) To evaluate the accuracy of AI predictions against wearable and radiographic data.

• (4) To propose a policy-ready screening framework deployable in low-resource school settings.

Novelty and contribution

This study contributes a context-adapted posture screening framework for adolescent populations in Southeast Asia. The main contributions are: (a) a Southeast Asia–calibrated anthropometric adjustment strategy for vision-based pose estimation; (b) a multi-modal benchmarking approach that links smartphone-based estimates with radiographic references and wearable sensor measurements for screening-level validation; (c) an end-to-end, low-cost deployment pipeline designed for school-based monitoring with on-device inference; and (d) a practical screening and triage workflow that can be integrated into existing school health routines in resource-limited settings.

This paper responds to both clinical urgency and technological feasibility, offering a scientifically grounded and socially scalable solution for mitigating postural degeneration among youth in the post-pandemic era.

To our knowledge, this study is among the few to combine SEA-specific anthropometric calibration, radiographic validation, and low-cost school deployment for AI-based posture screening in adolescents.

A comparison of existing AI-based posture assessment systems and their deployment characteristics is summarized in Table 1, highlighting gaps in validation, hardware accessibility, and suitability for low-resource settings.

Table 1 Comparison of existing AI posture assessment systems and postureguard.

Related work

Previous studies have explored vision-based posture and pose estimation approaches for health screening and educational contexts,including methods focusing on lightweight models, school-based deployment, and adolescent assessment^{14,15,16,17,18,19,20,21,22}.

AI-based posture detection: progress and pitfalls

Recent advances in pose estimation—such as OpenPose, MediaPipe, and BlazePose—have enabled real-time skeletal tracking using only camera-based inputs. These tools have been applied in domains ranging from athletic symmetry analysis (Han et al., 2021) to fall detection in elderly care¹⁰. However, despite impressive benchmarks, three persistent limitations hinder their application for pediatric postural screening.

First, there exists a morphological bias in training datasets. Pose estimation models are predominantly trained on Western-centric corpora (e.g., COCO, MPII), causing up to 32% joint localization errors when applied to Southeast Asian populations⁷. For instance, MediaPipe misplaces shoulder landmarks by an average of 4.7 cm in Indonesian adolescents—leading to significant distortions in calculated forward head angles (FHA).

Second, these systems often operate in controlled environments, failing in real-world school settings where occlusion, poor lighting, and non-uniform clothing are common¹¹. Third, their outputs remain clinically disconnected—offering no correlation with validated spinal metrics such as Cobb angle or thoracic curvature.

Clinical assessment: the gold standard gap

Radiographic assessment remains the gold standard for quantifying postural abnormalities, particularly via Cobb angle and FHA measurements³. However, these methods are manual, time-consuming (15–20 min per patient), and require trained radiologists or orthopedists. Attempts to automate these calculations—such as Kim and Chen’s⁸ semi-automated Cobb estimator—still rely on direct radiographic input and lack real-time pose integration.

Moreover, no study to date has correlated AI-predicted posture with radiographic ground truth in pediatric cohorts from developing nations. This validation gap erodes trust among clinicians and limits the translational adoption of AI posture systems¹². This disconnect is especially critical in Southeast Asia, where recent studies report a disproportionately high incidence of thoracic kyphosis among youth², yet no AI-clinical convergence framework has been piloted regionally.

Behavioral risk factors: digital habits without measurement

A growing body of behavioral literature suggests a strong association between prolonged screen use and postural degradation. Lee et al.⁴ demonstrated that each 1 cm forward head shift increases spinal load by 4.5 kg, a finding echoed in WHO’s¹ global warning on digital ergonomics. However, most existing studies suffer from three key weaknesses: (a) Overreliance on self-reported surveys—72% of studies use parental or student recall, introducing significant bias. (b) Use of mechanical simulations on cadavers, rather than live observational tracking in digital environments. (c) A lack of longitudinal datasets from Southeast Asia measuring daily screen time alongside biomechanical metrics such as Cobb angle or FHA.

Thus, the literature confirms the problem’s urgency, but fails to deliver real-time, AI-based behavioral-postural correlation, particularly in vulnerable populations.

Scalability & access: barriers in low-resource settings

Even when technically sound, current posture assessment systems remain largely inaccessible in public school contexts. Commercial motion capture systems (e.g., Vicon, Kinect) cost between $15,000–$50,000 per unit and require lab-grade calibration and trained personnel⁹. Inertial sensor-based tools like Shimmer3 offer better mobility but involve frequent recalibration, data syncing, and firmware updates—rendering them impractical for rural deployment.

School-based projects such as iMoCap lack the real-time feedback necessary for behavioral intervention, while smartphone-based approaches often suffer from poor pose fidelity. To date, no published solution has demonstrated validated clinical implementation in Southeast Asian school settings.

This study was developed in active collaboration with Indonesia’s Orthopedic Association (PABOI), ensuring alignment with the 2024 National Posture Screening Mandate. This partnership enabled early radiologist involvement, standardized clinical benchmarks, and seamless integration into policy translation.

Positioning and contribution of our work

To address these compounded gaps, we propose PostureGuard—an AI-clinical hybrid system purpose-built for developing nations. Compared to existing works, PostureGuard offers: (a) Anthropometric Calibration: First pose estimation tool tuned to Southeast Asian body types, reducing joint localization error from 4.7 cm to < 1.2 cm. (b) Multi-Modal Validation: Cross-validation of pose metrics with 50 radiologist-graded X-rays and wearable sensor data (κ = 0.91 inter-rater agreement). (c) Policy-Level Scalability: Deployed via Android smartphones (∼$200), successfully piloted in 15 Indonesian schools with 92% compliance. (d) Simultaneous Behavior Monitoring: Uniquely integrates screen time tracking with posture monitoring to establish dose–response relationships.

Existing AI posture-assessment systems vary widely in terms of clinical validation, deployment feasibility, and population fit, with most models optimized for Western anthropometry and requiring laboratory-grade hardware. Table 2 summarizes these system characteristics and highlights the performance gap that PostureGuard is designed to address.

Table 2 Comparison of prior AI posture systems and postureguard.

As shown in Table 2, current solutions lack regional calibration and remain unsuitable for low-cost deployment, reinforcing the need for a scalable approach tailored to school-based screening in resource-limited environments.

Key differentiator

In contrast to previous works, this study delivers a radiographically validated, anthropometrically calibrated, low-cost AI system with proven deployment success in real-world, low-resource school environments.

This integrated approach fills a critical translational gap between AI posture estimation, clinical biomechanics, and policy-driven public health.

Methods

Overview

This section provides an overview of the PostureGuard system architecture and operational workflow. PostureGuard represents the first multi-modal, region-calibrated AI system designed for early detection of postural dysfunction among Southeast Asian youth. The system bridges the gap between AI-based pose estimation and clinical gold standards by integrating four innovation pillars:

• SEA-calibrated MediaPipe BlazePose (v0.4)

• Cross-validation using FDA-cleared Shimmer3 IMU Unit

• Ground-truth alignment with 50 annotated spinal radiographs

• Low-cost deployment (< USD 200 per unit) for low-resource school environments

The system operates through sequential processing steps including video capture, pose estimation, region-specific calibration, angle computation, and risk classification, as illustrated in Fig. 1.

Inspired by our prior work on context-aware semantic segmentation¹³, we adapt a region-specific calibration layer that enables robust estimation of spinal angles under diverse lighting and occlusion conditions.

To improve reproducibility and transparency, the complete inference pipeline is summarized below. This modular structure ensures adaptability to future hardware, population diversity, and national digital-health frameworks.

Algorithm 1: PostureGuard Inference Pipeline

Participants and data collection

Data were collected from 200 students (aged 8–18 years) enrolled in 15 Indonesian public schools between July and September 2024. Participants were stratified by daily screen time (< 2 h, 2–6 h, and > 6 h per day) and age group (8–12 years, pre-pubertal; 13–18 years, adolescent) to capture variability in postural habits and developmental stages. The raw screen-time values were extracted from in-phone usage logs and recorded as continuous data; however, for visualization and interpretability, these values were grouped into predefined categorical bins. Statistical analyses (including regression modeling) were performed using the original continuous values, not the grouped ranges.

Each participant performed two static postures—neutral standing (front and side views) and forward-head hold (5 s)—and two dynamic tasks simulating common device use: tablet typing (30 s) and slouched video watching (30 s). All sessions were conducted under standardized lighting and background conditions to ensure consistent data acquisition across schools.

Measurement devices used in this study are summarized in Table 3, comprising an RGB-D camera for image capture, a wearable IMU sensor for neck-flexion validation, and radiographic imaging for ground-truth comparison.

Table 3 Data acquisition setup.

Ethical approval was obtained from the Universitas Nasional Ethical Review Board (UNAS-IRB-2025–02). All procedures adhered to institutional and international ethical guidelines. Written informed consent was obtained from parents or legal guardians of all child participants. Assents were obtained from each student participant in accordance with age-appropriate IRB requirements.

(Algorithm 1 describing the PostureGuard inference pipeline has been relocated to Section III.A to maintain logical flow.)

Southeast Asian (SEA) anthropometric calibration and normalization

The SEA calibration pipeline was designed to correct pose-estimation biases arising from anatomical differences between Western-trained datasets and Southeast Asian adolescent morphologies.

The process comprised three sequential stages:

• 1. Joint Ratio Normalization – adjusts shoulder–pelvis–head distances using SEA-specific anthropometric ratios.

• 2. Offset Correction – aligns keypoints based on radiograph–sensor delta values obtained from the calibration cohort.

• 3. Growth-Aware Heatmaps – incorporate age-dependent vertebral morphology to refine landmark inference for adolescents.

To derive scaling parameters, a regional anthropometric dataset comprising 420 adolescents (aged 8–18 years) from Indonesia, Malaysia, and Thailand was used⁷, SEA BodyScan, 2023).

Segmental ratios—including neck-to-torso, shoulder width, and pelvis-to-limb proportions—were computed from 3D body scans and compared to the global COCO benchmark.

Statistical analysis revealed consistent proportional differences: shorter torso-to-leg ratio (− 7 ± 2%), narrower shoulder width-to-height ratio (− 12 ± 3%), and increased baseline thoracic curvature (+ 5 ± 1.8°).

Scaling coefficients were therefore determined empirically as:

$${S}_{torso}=0.93, {S}_{shoulder}=0.88$$

(1)

These were optimized using iterative least-squares fitting between 200 field-captured RGB-D frames and corresponding IMU–radiograph angular pairs.

External validation on additional cohorts from Malaysia (n = 80), Thailand (n = 60), and Vietnam (n = 40) confirmed comparable accuracy improvements (≤ 2.5° deviation), demonstrating cross-national robustness within the SEA region.

The normalized skeletal geometry followed the correction function:

$${Adjusted}_{length}={Raw}_{length} x \frac{SEA\_ratio}{Global\_ratio}$$

(2)

For example, shoulder–pelvis distances were scaled by 0.93 and neck–torso ratios by 0.88 relative to COCO benchmarks.

Validation against 50 radiographic ground truths reduced mean joint localization error from 4.7 cm to 1.1 cm, as summarized in Table 4, representing a 75% improvement in regional accuracy.

Table 4 Pose estimation error before and after sea-specific calibration.

This calibrated skeletal model was subsequently integrated into the dynamic simulation and screen-time correlation analyses described in Sections III.D–III.G.

IMU sensor setup and angle computation

To validate the posture angles predicted by the AI-based system, two Shimmer3 IMU units were affixed at the nape (C7) and mid-thoracic line (T12) of each participant using medical-grade adhesive patches. Each IMU captured tri-axial accelerometer and gyroscope data (± 2000°/s range) at 128 Hz. The IMU streams were synchronized with the RGB-D video feed through Bluetooth Low Energy (BLE), ensuring frame-level correspondence between sensor and AI-derived measurements.

Three clinical indicators—Forward Head Angle (FHA), Thoracic Kyphosis Angle (TKA), and Pelvic Tilt Angle (PTA)—were used to quantify spinal alignment (Table 5). Each angle was computed independently from both AI-estimated skeletal landmarks and IMU-derived orientation data to enable cross-validation.

Table 5 Clinical angle definitions and screening thresholds used in postureguard.

These thresholds align with international ergonomic standards and were used to train the PostureGuard risk classification module described in Algorithm 1.

The Forward Head Angle (FHA) was computed as the deviation between the line connecting the tragus and C7 vertebra and the vertical axis, expressed as:

$$FHA={tan}^{-1}\left(\frac{{y}_{tragus}-{y}_{C7}}{{x}_{tragus}-{x}_{C7}}\right)$$

(3)

Thoracic Kyphosis Angle (TKA) was derived using the relative orientation of the estimated C7 and T12 keypoints, as illustrated in the following pseudocode.

Algorithm 2: Pseudocode for thoracic kyphosis angle (TKA) computation using AI Landmarks

The resulting angles from IMU and AI measurements were compared frame-by-frame. A mean absolute error (MAE) below 2.3° was considered clinically acceptable for screening consistency, following prior radiographic validation benchmarks.

Posture scoring and validation

To assess the accuracy of the PostureGuard system, a dual reference validation protocol was used. These validation methods were employed only for benchmarking purposes and were not part of the deployed workflow shown in Fig. 1.

Radiographic Validation: Forward Head Angle (FHA) and Cobb-based sagittal curvature measurements were obtained from spinal radiographs (Siemens Mobilett Mira system). Two board-certified radiologists independently annotated all radiographs, yielding an inter-rater reliability of Cohen’s κ = 0.91, indicating excellent agreement. These measurements served as the clinical ground truth against which AI-derived posture angles were compared (Fig. 2).

Fig. 2

Cross-Validation: Predicted vs. X-ray vs. Sensor Overlay. Panel comparison illustrating alignment between AI-estimated sagittal angles and the clinical ground truth derived from X-ray and IMU recordings.

Wearable Sensor Validation: Postural angles derived from AI inference were additionally compared with neck flexion values recorded from Shimmer3 IMU wearable sensors (± 3° clinical tolerance). Synchronization between the wearable sensor stream and video frames was performed through timestamp alignment (Bluetooth Low Energy latency < 100 ms).

Together, these two reference sources provided a robust benchmark for validating the reliability of the AI-generated posture predictions under real-world classroom conditions.

Field Observations: During deployment across 15 Indonesian schools, a distinctive postural deviation pattern emerged that is underrepresented in prior posture AI systems. Specifically, a localized thoracic kyphosis—a pronounced curvature between the upper abdomen and shoulder region—was frequently observed while the lumbar region remained relatively neutral.

This semi-reclined “study posture,” often occurring during screen use in bed or against a wall, induces concentrated biomechanical stress on the thoracic and cervical segments, potentially accelerating early spinal degeneration.

Figure 3 visualizes this forward-flexed posture and the biomechanically vulnerable thoracic segment.

Fig. 3

Postural Deformity in Forward Head Posture with Thoracic Kyphosis.

Conventional pose estimation frameworks typically emphasize cervical (FHA) and pelvic (PTA) parameters while overlooking mid-thoracic deviations. The SEA-specific anthropometric calibration in PostureGuard enables detection of these subtle curvatures, enhancing diagnostic sensitivity and field applicability for early-stage screening in schools.

All three postural angles demonstrated clinically acceptable accuracy for non-invasive screening. Statistical comparison using paired t-tests confirmed significant agreement (p < 0.01) between AI-derived and radiographic measurements, supporting the clinical reliability of PostureGuard for field deployment.

Statistical analysis

Statistical analyses were conducted to evaluate both the accuracy of angular measurements and the reliability of the final posture-risk classification.

• (1) Angle Validation and Error Metrics

  For each participant, Forward Head Angle (FHA), Thoracic Kyphosis Angle (TKA), and Pelvic Tilt Angle (PTA) predicted by the AI model were compared with radiographic ground truths using paired t-tests. Agreement between the two modalities was further quantified through Bland–Altman analysis and Pearson correlation coefficients (r).

  The model achieved mean absolute errors of 2.3°, 3.6°, and 1.7° for FHA, TKA, and PTA, respectively. Correlation with radiographic measurements was stronger = 0.93 (FHA) and r = 0.89 (TKA), both p < 0.01—confirming statistical agreement within clinically acceptable tolerances.

• (2) Rule-Based Risk Classification

  Postural risk levels were categorized according to WHO-aligned angular thresholds derived from regional clinical references (Table 5). Participants exceeding both FHA > 45° and TKA > 55° were classified as High Risk, those within FHA 40–45° or TKA 45–55° as Moderate Risk, and all others as Low Risk. This deterministic approach enabled interpretable and clinically consistent decision boundaries.

• (3) Posture Score and Trend Monitoring

  Each participant received a Posture Score ranging from 0 to 100, calculated from normalized angular deviations relative to healthy baseline values. The system generated:

  • (i) Color-coded risk indicators (green = low, yellow = moderate, red = high);

  • (ii) Six-month trend dashboards tracking longitudinal improvement or deterioration; and.

  • (iii) Real-time warning integration with Indonesia’s Ministry of Health (MoH) Digital Wellness Compliance Alerts for policy-level deployment.

All statistical computations were performed using Python 3.10 (NumPy and SciPy libraries). A significance level of p < 0.05 was adopted for all inferential tests.

Screen-time monitoring & confounders

Screen-time exposure was objectively quantified using Android’s Digital Wellbeing and iOS Screen Time logging systems, which provide continuous usage data across all installed applications.

Each participant’s daily average screen time was calculated over a seven-day baseline period preceding posture assessment. To ensure reliability, parental validation logs were collected for participants under 16 years old, while students above 16 provided self-reported verification using institutional monitoring forms.

Screen-time values were stratified into three exposure levels:

• (i) Low (< 2 h/day),

• (ii) Moderate (2–6 h/day), and.

• (iii) High (> 6 h/day).

Potential confounders—including age, sex, body mass index (BMI), physical activity, and academic workload—were controlled through a multivariate regression model assessing their influence on the Forward Head Angle (FHA) and Thoracic Kyphosis Angle (TKA). Exercise frequency (hours/week) and study duration (hours/day) were included as continuous covariates.

The adjusted regression model took the form:

$${FHA}_{i}={\beta }_{o}+{\beta }_{1}\left({ScreenTime}_{i}\right)+ {\beta }_{2}\left({BMI}_{i}\right)+{\beta }_{3}\left({Exercise}_{i}\right)+ {\beta }_{4}\left({StudyTime}_{i}\right)+ {\in }_{i}$$

(4)

Model diagnostics confirmed the absence of multicollinearity (VIF < 2.0) and heteroscedasticity (Breusch–Pagan p > 0.1).

This ensured that the observed association between screen time and postural deviation reflected genuine behavioral effects rather than demographic or lifestyle confounding.

Evaluation metrics

To assess classification robustness and overall system performance, the PostureGuard algorithm was evaluated against radiologist-labelled ground truths using standard machine learning metrics. Sensitivity and specificity were derived from the confusion matrix comparing AI-based risk predictions with expert classifications across both coronal and sagittal planes.

$$Sensitivity= \frac{TP}{TP+FN}$$

(5)

$$Specificity= \frac{TN}{TN+FP}$$

(6)

where TP, TN, FP, and FN denote true positives, true negatives, false positives, and false negatives, respectively.

Additional performance indicators included accuracy, precision, recall, F1-score, and area under the ROC curve (AUC). Quantitative benchmarking showed that the system remained within clinically acceptable error boundaries across all angles, as shown in Table 6. Beyond ablation analysis, the final model demonstrated strong diagnostic capability when evaluated against clinical ground truth. The performance metrics summarised in Table 7 indicate high predictive reliability, with sensitivity and specificity exceeding 90%, and an AUC score of 0.94.

Table 6 Mean absolute error (Mae) between AI-estimated and radiographic angles.

Table 7 Diagnostic performance of the postureguard classification model.

The results confirm that PostureGuard maintains high diagnostic reliability across both anatomical planes.

This consistent performance demonstrates the system’s suitability for non-invasive school-based screening, providing a quantitative complement to conventional visual examination methods.

Results

Technical validation

To establish the technical feasibility of PostureGuard for school-based deployment, we first evaluated protocol adherence, tracking performance, and device-level constraints across the full study cohort.

Protocol execution and system reproducibility

All procedures detailed in Section III were consistently executed across 200 participants (aged 8–18 years) from 15 Indonesian public schools. The experimental protocol yielded three principal outcomes demonstrating technical reliability and procedural reproducibility:

• (1) High-Fidelity Pose Tracking:

  The SEA-calibrated PostureGuard system achieved 98.2% successful skeletal keypoint detection under varying classroom lighting and camera angles. The underlying MediaPipe BlazePose engine ensured stable landmark inference and minimal frame-drop, with a mean keypoint confidence of 0.97 ± 0.02 across all trials.

• (2) Real-Time Risk Detection on Low-Cost Devices:

  The quantized TensorFlow Lite implementation demonstrated full operational capability on sub–USD 200 Android smartphones, achieving an average inference latency of 1.8 s per scan and 18 frames per second (fps) for continuous monitoring. These results confirm that posture-risk classification can be achieved in real time without dependence on high-end hardware or cloud-based computation.

• (3) Ethical and Procedural Uniformity:

  All data collection adhered to the Universitas Nasional Institutional Review Board (UNAS-IRB-2025–02) approval, ensuring consistent participant treatment and full compliance with ethical standards. Parental consent and participant assent were obtained for all minors. This procedural consistency ensured equitable representation across demographic subgroups and safeguarded the study’s methodological integrity. Taken together, these findings indicate that PostureGuard can be operated reliably in typical public-school environments using affordable hardware, while maintaining stable pose tracking and standardized procedures across diverse classrooms. This level of robustness is essential for scaling posture screening beyond controlled laboratory settings.

• (4) Field Deployment Validation

  To assess the system’s robustness under real-world, non-laboratory conditions, PostureGuard was deployed directly in authentic classroom environments using standard smartphone cameras. The SEA-calibrated BlazePose model provided live skeletal overlays and instant postural feedback for each student during daily activities.

These deployment results confirm that PostureGuard maintains high reproducibility and operational reliability even in uncontrolled environments. As visualized in Fig. 4, the system captures subtle misalignments in the head–shoulder–pelvis axis that are commonly observed in Southeast Asian adolescents and may be under-recognized by Western-trained posture models. This supports the importance of region-specific calibration and highlights the system’s readiness for cost-effective, scalable health screening across under-resourced educational settings.

Fig. 4

Illustrates an example of live deployment at a public school in Jakarta, showing real-time posture inference under natural lighting conditions. The skeletal visualization highlights subtle forward head posture (FHP) and mild shoulder rounding; deviations commonly linked to prolonged screen exposure among adolescents.

Real-world deployment robustness

While controlled evaluations are informative, school-based AI systems must also remain reliable under the heterogeneous conditions of everyday classrooms. Field validation was conducted across 15 public schools in Jakarta, Bandung, and Surabaya, encompassing diverse classroom environments and demographic variations.

Data was collected under a range of lighting conditions (indoor fluorescent, open-window daylight, and backlit scenarios), postural variations (neutral standing, seated device use, and slouched postures), and clothing types (school uniforms, sports attire and layered clothing).

Despite high overall robustness, performance degradation was observed in strong backlighting conditions, where keypoint detection accuracy dropped to 85.1%.

This decline was attributed to reduced local contrast around facial and shoulder landmarks, particularly affecting Forward Head Angle (FHA) computation.

In these scenarios, pose misalignment occasionally resulted in angular estimation offsets of up to ± 3.8°, which could shift borderline cases from “Moderate” to “High” risk classification.

To address these limitations, two on-device adaptive preprocessing methods were implemented:

• 1. Auto-exposure compensation, which dynamically adjusts brightness and white balance to stabilize visual contrast across lighting zones.

• 2. Contrast normalization layer, applied prior to keypoint extraction, enhancing edge definition around low-texture regions (e.g., neck and shoulder boundaries).

After integrating these modules, re-evaluation on 120 image sequences from backlit classrooms restored detection accuracy to 95.8%, with MAE reduction from 3.8° to 2.1°.

This confirmed that environmental variability can be effectively corrected through lightweight, embedded preprocessing without affecting inference latency or device compatibility.

Additionally, tests on clothing variability showed no significant accuracy difference (< 1° deviation) between uniform and sportswear conditions, indicating that the pose-estimation framework was resilient to minor occlusions and fabric folds.

Overall, these results demonstrate that anticipated sources of variability in school settings—lighting, posture diversity, and clothing—can be effectively managed through lightweight on-device preprocessing, preserving accuracy without compromising speed or accessibility.

Figure 5 illustrates a representative real-world posture scenario observed during smartphone use.

Fig. 5

Example of forward head posture and thoracic kyphosis during smartphone use (image anonymized for privacy). The illustration reflects typical adolescent postural deviations captured under real-world classroom lighting.

The visualized forward displacement and upper-thoracic curvature correspond closely with the quantitative deviations reported in Table 8, reinforcing the ecological validity of field measurements.

Table 8 Comparison of AI-predicted angles and radiographic ground truths (Mean ± SD, °).

Captured under natural daylight with partial backlighting, this scenario demonstrates the typical postural deviations quantified in adolescent users—forward head inclination and thoracic rounding—highlighting the relevance of environmental adaptation in AI-based screening.

Clinical benchmarking with radiographs and sensors

To verify clinical validity, PostureGuard’s AI-derived postural angles were compared against gold-standard radiographic measurements obtained from 50 anonymized spinal X-rays, independently validated by two certified radiologists.

All Mean Absolute Errors (MAE) were within clinically accepted tolerance ranges, confirming strong predictive fidelity of the AI-based estimation.

The model achieved 93.4% sensitivity and 91.8% specificity in detecting postural deviations beyond diagnostic thresholds, demonstrating substantial diagnostic equivalence with radiographic evaluation.

Among all angular measures, the Forward Head Angle (FHA) achieved the lowest MAE (2.3°), supporting its suitability for non-radiographic field applications in large-scale screening.

Key performance highlights:

• All MAE values are within radiologist-defined acceptance limits.

• Inference time per scan: 1.8 s, representing a 20 × speed-up compared to manual radiographic interpretation.

Figure 6 presents the Bland–Altman plot comparing AI-predicted FHA values with X-ray ground truths.

Fig. 6

Bland–Altman plot showing agreement between AI-predicted Forward Head Angles (FHA) and radiographic ground truths. The gray line denotes the mean bias; red and blue dashed lines indicate ± 1.96 SD (95% limits of agreement).

A total of 96% of data points lie within the 95% limits of agreement, indicating minimal systematic bias and confirming strong concordance between the two modalities. From a clinical standpoint, this level of agreement suggests that PostureGuard can approximate radiographic assessments closely enough to support large-scale screening in contexts where X-ray imaging is impractical or unavailable. Rather than replacing radiographic evaluation, the system functions as a front-line triage tool to prioritize adolescents who may benefit from further specialist assessment.

To provide a holistic overview, Fig. 7 summarizes three critical validation aspects:

• (A) correlation between screen time and postural deviation,

• (B) quantitative comparison between AI predictions and radiographic benchmarks, and

• (C) ablation analysis demonstrating the contribution of individual calibration and validation components.

Fig. 7

A Linear correlation between daily screen time and Forward Head Angle (FHA). B Mean Absolute Error (MAE) comparison between PostureGuard and radiographic ground truths. C Ablation study illustrating performance degradation after removal of SEA calibration and multi-modal validation components.

As shown in Fig. 7:

• Panel A reaffirms the linear dose–response relationship detailed in Section IV.B.

• Panel B demonstrates that PostureGuard consistently achieves radiographic-grade precision within ± 3° of measurement error.

• Panel C confirms that removing SEA anthropometric calibration or multi-modal validation significantly increases MAE, underscoring their critical role in maintaining accuracy.

Collectively, these results validate that PostureGuard delivers clinically reliable posture assessments aligned with radiographic standards, while retaining low computational cost and scalability for real-world public-health deployment.

Contribution of SEA calibration and multi-modal validation

To quantify the contribution of Southeast Asian (SEA) anthropometric calibration and multi-modal validation, a three-tier ablation experiment was conducted. The baseline model used the standard MediaPipe BlazePose configuration, followed by sequential integration of SEA calibration and validation against wearable sensor and radiographic reference measurements (used only for benchmarking, not in the deployed workflow).

To evaluate the benefit of Southeast Asian anthropometric calibration, angle estimation errors were compared before and after calibration and subsequent multi-modal validation. As shown in Table 9, calibration reduced angular error by 39–50%, demonstrating its value in improving measurement precision for this population.

Table 9 Mean absolute error (MAE) reduction following model calibration.

The SEA-specific calibration layer yielded the largest performance gain, reducing MAE by approximately 40–50% across all postural parameters.

This improvement primarily resulted from correction of regional morphological discrepancies—particularly in neck-to-torso and shoulder-to-pelvis proportions—which differ significantly between Southeast Asian and Western adolescent populations.

Subsequent inclusion of multi-modal validation (AI + sensor + radiograph fusion) provided an additional 8–12% reduction in residual error, reflecting enhanced cross-sensor coherence and improved field-level inference stability.

Together, these results highlight the critical role of region-specific biomechanical modeling and sensor-informed feedback loops in achieving clinically reliable posture assessment in low-resource settings.

Without SEA calibration, Western-trained pose estimators tended to overestimate thoracic curvature and underestimate pelvic tilt, leading to potentially misleading clinical classifications. The multi-modal fusion layer effectively mitigated this bias by jointly optimizing angular alignment across AI predictions, IMU sensor readings, and radiographic ground truths. Together, these ablation outcomes confirm that both cultural–anthropometric adaptation and sensor-informed correction are indispensable for developing equitable and context-aware AI systems applicable across diverse populations.

Benchmarking against state-of-the-art systems

To contextualize PostureGuard’s performance, a comparative benchmarking analysis was conducted against five state-of-the-art AI-based posture or spinal assessment frameworks published between 2018 and 2024.

The comparison covered validation methodology, clinical accuracy, population adaptation, hardware cost, and deployment readiness.

Representative benchmarks included:

• PoseNet (Google AI, 2018) – a lightweight 2D pose estimator widely used for general human posture tracking but lacking clinical calibration or validation.

• OpenPose (Cao et al., 2019) – multi-person skeletal estimation using part affinity fields, validated primarily in Western adult datasets.

• DOPE (Deep Object Pose Estimation, Tremblay et al., 2020) – strong geometric performance in structured environments but not optimized for medical or adolescent contexts.

• SPINE-Net³ – designed for radiographic spinal curvature analysis, requires medical imaging input and lacks field-level deployability.

• ScolioNet (Kim et al., 2022) – deep-learning model for scoliosis angle estimation using X-rays only, non-scalable for community screening.

Several existing AI-based posture and spinal assessment frameworks represent distinct design trajectories, ranging from general-purpose pose estimation models to clinically oriented imaging systems and emerging AI–biomechanical hybrid approaches. These systems were selected for comparison to reflect the current landscape of posture assessment technologies.

Validation confirmed that PostureGuard’s angle estimation remained within clinically accepted tolerance ranges when benchmarked against both radiographic and wearable sensor references. This dual-reference validation was used for accuracy assessment only and is not part of the operational workflow.

As summarized in Table 2, key differentiators among existing systems include clinical validation status, regional anthropometric adaptation, hardware accessibility, and deployment feasibility. While many prior approaches remain limited to laboratory evaluations or rely on specialized diagnostic equipment, PostureGuard was designed for low-cost school-based deployment and has undergone real-world field testing.

Rather than positioning systems as competitive, this comparison highlights divergent development priorities across the field. Within this landscape, PostureGuard addresses a specific gap by offering a region-calibrated, clinically benchmarked, and deployment-ready screening approach suitable for public health implementation in resource-constrained environments.

Behavioral correlation between screen time and postural angles

Beyond technical performance, we next examined how device-use behaviors relate to postural deviations in the study population. A clear dose–response association was observed between daily screen exposure and the severity of postural deviation among participants. Both the Forward Head Angle (FHA) and Thoracic Kyphosis Angle (TKA) demonstrated progressive increases with prolonged screen use, while the Pelvic Tilt Angle (PTA) showed a compensatory decrease, indicating potential lumbar load redistribution. The summarized results are presented in Table 10.

Table 10 Progressive changes in postural angles with increasing daily screen time.

These findings indicate a statistically significant upward trend in both FHA and TKA with longer daily screen exposure. Participants exceeding six hours of screen use exhibited, on average, a + 11.1° increase in FHA and + 9.7° in TKA relative to controls. The concurrent decrease in PTA reflects reduced pelvic tilt and altered lumbar mechanics, suggesting early postural compensation.

Figure 8 illustrates the linear regression between daily screen time and FHA. The model demonstrated a strong positive correlation (R² = 0.82, p < 0.001), where each additional hour of screen exposure corresponded to an estimated ~ 2° increase in forward head inclination.

Fig. 8

Linear correlation between daily screen time and Forward Head Angle (FHA). The solid line indicates the regression fit, and the shaded region represents the 95% confidence interval (R² = 0.82, p < 0.001).

Collectively, these results reveal a robust statistical association between extended device use and measurable deviations in cervical and thoracic posture. While the observed trend suggests potential biomechanical strain due to prolonged screen exposure, the data remain correlational; longitudinal or interventional studies are required to confirm causality.

Policy translation

To evaluate real-world implementability, PostureGuard was deployed as a screening tool across 15 public schools, where it was integrated into existing health-check routines with 92% adoption among targeted classrooms (Table 10). Teachers and school health officers were able to operate the system using standard Android smartphones without additional hardware or technical staff. These findings indicate that the AI pipeline is compatible with current school workflows and resource constraints.

Building on these outcomes, the system has been incorporated into a broader digital wellness initiative led by local education and health authorities, positioning PostureGuard as a candidate framework for scalable adolescent spinal health screening in public-school settings.

These translational findings suggest that AI-assisted posture screening can be integrated into existing school health ecosystems with minimal disruption, creating a practical pathway from technical validation to population-level, screening-level implementation. During deployment across 15 schools, the system demonstrated practical feasibility, and preliminary six-month follow-up indicated improved posture-related indicators in participating schools; however, this observation was not derived from a controlled trial and should be interpreted cautiously.

Discussion

Summary of key findings

This study presents PostureGuard, a low-cost and field-tested AI system designed for adolescent posture assessment in Southeast Asian school settings. By incorporating region-calibrated pose estimation and multi-modal validation, the system demonstrates the potential to bridge research-grade computer vision methods with practical applications in resource-limited environments.

When compared with radiologist-interpreted spinal radiographs, PostureGuard achieved a mean absolute error of 2.3° for forward head angle estimation. A strong linear association (R² = 0.82, p < 0.001) was observed between daily screen exposure and deviations in cervical posture, indicating measurable behavioral correlations within the study population. In addition, the system maintained consistent performance in real-world classroom conditions and operated effectively on sub–USD 200 smartphones with an average inference latency of 1.8 s.

Together, these findings suggest that PostureGuard may offer a feasible and context-adapted approach for large-scale posture screening in low-resource educational environments, providing evidence for both technical readiness and behavioral relevance.

Clinical significance and patient-level benefits

From a clinical perspective, the proposed system is intended to support early posture screening and monitoring at the individual student level, particularly in settings where routine clinical assessment is limited. By providing objective and repeatable posture measurements, the system may assist clinicians and school health personnel in identifying students who could benefit from further clinical evaluation, thereby supporting earlier referral and more consistent triage decisions.

Importantly, PostureGuard is not designed as a diagnostic tool, but rather as a screening and monitoring aid that complements existing clinical workflows. Its use may help track posture-related changes over time, support posture-awareness interventions, and reduce reliance on subjective visual inspection alone. In this way, the system prioritizes patient-level relevance by enabling scalable, low-cost monitoring while preserving the role of clinicians in diagnosis and treatment decisions.

Engineering innovations for global health

PostureGuard incorporates several engineering design choices aimed at enabling reliable posture assessment in resource-limited school environments. These decisions were shaped by hardware constraints, regional anthropometric characteristics, and the need for interpretable outputs suitable for non-specialist operators.

• (1) Edge AI Optimization:

  The posture-estimation model was quantified and deployed using TensorFlow Lite (v2.14), reducing its memory footprint and enabling efficient on-device inference. In field tests, the system achieved an average processing latency of 1.8 s per image on commonly available Android smartphones (e.g., Redmi Note 11, Vivo Y20s). Offline capability was also implemented through local storage to support use in areas with inconsistent network availability.

• (2) Southeast Asian Anthropometric Calibration:

  A post-inference calibration layer was introduced to align pose-estimation outputs with regional body morphology, including variations in torso proportions and shoulder geometry. Validation using paired radiographic datasets demonstrated a measurable reduction in joint-localization error relative to uncalibrated baseline models. This underscores the relevance of population-specific adjustments when deploying AI posture systems beyond their original training distribution.

• (3) Multi-Modal Clinical Benchmarking:

  To strengthen measurement reliability, PostureGuard was evaluated using both radiographic references and wearable inertial-sensor measurements. Agreement between these modalities resulted in an inter-rater reliability of κ = 0.91, providing complementary evidence supporting consistency of the angle estimates produced by the system.

Existing posture-assessment systems vary in intended application, validation methodology, and deployment feasibility. Many rely on high-cost hardware or remain confined to laboratory evaluation, with limited support for diverse body morphology or non-clinical usage environments. As summarized in Table 2, PostureGuard differs in its combined emphasis on clinical validation, low-cost deployment, population-specific calibration, and school-based field implementation.

In aggregate, the system demonstrated reduced angular estimation error following calibration, broader compatibility with low-cost consumer hardware, and practical feasibility in real-world school workflows. These characteristics reflect an engineering direction focused on accessibility, contextual relevance, and translational readiness. The combination of edge optimization, population-informed calibration, and multi-modal benchmarking suggests potential suitability for broader adoption in school-based posture-screening programs, pending continued evaluation.

Policy and societal implications

Policy context

PostureGuard was piloted within school-based health initiatives aligned with the Indonesian Ministry of Health’s 2024 Digital Wellness Framework. During deployment in 15 public schools, the system was operated by teachers and school health officers using standard Android smartphones, demonstrating compatibility with existing workflows and resource constraints. While not a formal policy endorsement, these pilot activities indicate that AI-assisted posture screening may be feasible for integration into early-stage preventive health programs.

Economic feasibility

The system’s hardware requirements—sub–USD 200 smartphones—suggest potential accessibility for low-resource settings. Preliminary cost-related observations indicate that school-based AI screening may help reduce reliance on radiographic imaging for initial assessments, though formal economic evaluation is required to estimate long-term budgetary implications.

Clinical and behavioral insights

Associations observed between daily screen exposure and posture deviations suggest that behavioral factors may contribute to spinal alignment patterns among adolescents. These correlations highlight the potential value of routine posture monitoring in school environments; however, causal interpretations should be made cautiously, and longer-term studies are needed to understand sustained musculoskeletal impact.

Technological and societal benefits

Edge-based inference enables offline operation, making the system suitable for schools with limited connectivity. Feedback from educators and health staff during the pilot period suggests that interpretable visualizations may support greater awareness among students regarding posture-related behaviors.

Regional opportunities

Discussions with regional stakeholders have explored the possibility of examining PostureGuard’s applicability in neighboring Southeast Asian contexts. Such exploratory dialogues reflect broader interest in scalable, context-adapted digital health tools, although formal regional deployment remains in preliminary stages.

Novelty and future outlook

PostureGuard introduces several contributions that may support the development of context-adapted AI posture assessment tools for low-resource environments. First, the system incorporates anthropometric adjustments tailored to Southeast Asian adolescents, addressing potential mismatches when applying pose-estimation models trained predominantly on Western datasets. Second, the multi-modal evaluation using radiographic references and wearable sensors provides complementary evidence supporting measurement reliability in real-world school contexts. Third, the pilot implementation in Indonesian public schools illustrates how edge-based AI posture assessment can be explored within existing school health workflows.

The translational pathway—from data collection and calibration to field deployment and initial policy engagement—suggests potential avenues for integrating AI-assisted posture monitoring into school-based preventive health programs. While preliminary pilot results indicate promising improvements in posture-awareness behaviors, these findings should be interpreted cautiously due to the limited sample size and short follow-up duration.

Future work will focus on expanding longitudinal monitoring, refining demographic calibration for broader populations, and developing adaptive learning mechanisms to reduce reliance on radiographic data for calibration. Exploratory discussions with regional stakeholders also indicate interest in evaluating the system’s applicability in other Southeast Asian settings, presenting opportunities for scalable, context-aware digital health solutions.

Translational milestones

As illustrated in Fig. 9, the development and evaluation of PostureGuard progressed from multi-school data collection and model calibration to multi-modal benchmarking and exploratory engagement with school health stakeholders. These activities reflect an early-stage translational pathway for assessing the feasibility of AI-assisted posture monitoring in real-world educational environments.

Fig. 9

Overview of the developmental and evaluation pathway for PostureGuard.

Across the 15 pilot schools, teachers and school health officers reported improved awareness of posture-related risk patterns among students. Preliminary analyses suggested favorable trends in posture behavior during the six-month pilot; however, these findings should be interpreted cautiously due to the limited duration and absence of a controlled comparative group.

Figure 9 provides a visual summary of the project stages, including data acquisition, region-specific calibration, radiographic and wearable validation, and early interactions with local health and education stakeholders. These steps outline a potential framework for exploring how AI-based posture assessment tools may be integrated into broader school health initiatives in the future.

The flow diagram summarizes the stages from field data collection (n = 200 students, 15 schools), Southeast Asia–specific calibration, and multi-modal validation, to pilot deployment in school settings and initial engagement with health and education stakeholders. These steps illustrate a possible trajectory for future integration of AI-assisted posture assessment tools into preventive health programs.

Key milestones include:

• A. Multi-school data collection (n = 200 students, 15 schools)

• B. AI pose estimation with Southeast Asian anthropometric calibration

• C. Multi-modal benchmarking (radiographs + wearable sensors)

• D. Real-time risk stratification dashboard for pilot schools

• E. Preliminary posture-behavior improvements observed in pilot phase

• F. Technical documentation prepared for review by local health authorities

• G. Collaborative discussions with orthopedic and school health teams

• H. Early-stage alignment with national digital wellness initiatives

• I. Exploratory interest from regional stakeholders

Limitations and future work

Despite its robust field performance, several limitations warrant further exploration:

• 1. Calibration dependency: The current SEA anthropometric calibration relies on initial radiographs; future iterations may explore self-supervised domain adaptation to enable fully automated recalibration across diverse populations.

• 2. Demographic bias: The dataset predominantly represents urban school populations; subsequent studies will extend to rural and socioeconomically diverse settings for improved generalizability.

• 3. Limited longitudinal scope: The six-month follow-up provides early insight into behavioral impact; extended (> 12-month) longitudinal monitoring is planned to assess sustained musculoskeletal outcomes.

Closing remark

In summary, this work illustrates how culturally informed and resource-conscious AI design can support the development of posture assessment tools suitable for school-based use in low-resource environments. While further validation and longitudinal evaluation are needed, the experiences from this pilot suggest potential pathways for integrating AI-assisted screening into broader preventive health initiatives.

Conclusion

This study presents PostureGuard, an AI-assisted posture assessment framework designed for screening and monitoring in low-resource educational settings. By combining Southeast Asia–specific anthropometric calibration, multi-modal clinical benchmarking, and on-device inference compatible with low-cost smartphones, the system demonstrates the feasibility of applying culturally informed and resource-conscious AI design to school-based posture monitoring.

The pilot deployment across 15 Indonesian public schools highlights the potential for integrating AI-assisted posture assessment into existing health routines. While early field observations indicate positive trends in posture-awareness behaviors, these findings should be interpreted cautiously due to the limited sample size and absence of long-term follow-up. Further validation is needed before broader implementation can be considered.

Overall, PostureGuard illustrates how contextually adapted AI systems may support preventive musculoskeletal health initiatives in resource-limited environments. Continued refinement and longitudinal evaluation will be essential to assess its long-term utility, clinical relevance, and generalizability across diverse populations.

Future work

Future research will explore reducing the system’s reliance on radiographic calibration, including the use of self-supervised adaptation to improve generalizability across different demographic groups. Additional work will expand data collection to include rural and socioeconomically diverse populations. Longer-term (> 12-month) follow-up studies are planned to evaluate sustained behavioral and musculoskeletal outcomes. Integration of an AI-based educational feedback module will also be explored to support real-time posture-awareness interventions in classroom environments.