Abstract
Forward Head Posture (FHP), commonly known as "tech neck," is an increasingly prevalent musculoskeletal disorder among adolescents, driven by prolonged electronic device usage. Traditional diagnostic methods are often expensive, manual, or difficult to scale for large populations. This study introduces a low-cost, non-contact automated screening pipeline that combines computer vision-based landmark extraction with evolutionary feature selection. By utilizing the OpenPose algorithm to extract 2D skeletal keypoints and a genetic algorithm to optimize geometric features, the researchers developed an Extra Tree Classifier (ETC) that achieves 82.4% accuracy. This approach highlights the value of interpretable feature engineering in medical machine learning.
Introduction
Forward Head Posture (FHP) is defined by the cranium protruding forward beyond the body's sagittal plane. In some adolescent demographics, prevalence rates have reached as high as 63%. If untreated, FHP can lead to chronic muscle imbalances, restricted cervical mobility, and permanent spinal alterations. Despite its clinical significance, screening remains a challenge. Current gold standards—such as head-mounted trackers or manual photogrammetry—are either too costly for school-wide screenings or too time-consuming for clinical efficiency. There is a pressing need for an objective, automated, and non-contact tool to identify FHP in its early stages.
Methodology
The proposed pipeline transitions from raw image data to clinical classification through four distinct stages:
1. Landmark Extraction
The researchers employed the open-source OpenPose algorithm to identify 18 skeletal keypoints from 2D profile images of subjects. These landmarks include the nose, ears, eyes, and shoulders, providing a digital skeleton of the subject's posture.
2. Geometric Feature Engineering
Rather than training a deep learning model on raw pixel data, the study focused on calculating specific anatomical angles and distances. Using the Right Acromion Point (RAP) as a primary vertex, several key angles were defined to quantify the relationship between the head, neck, and torso.
3. Feature Selection via Genetic Algorithm (GA)
To refine the model and eliminate noise, a Genetic Algorithm was applied. The GA iteratively evolved subsets of features—including angles, distances, and demographic data like BMI—to find the combination that maximized classification accuracy. This process identified a core set of predictive features: NAH, ENA, PAH (Pupil-Acromion-Horizontal), and EAH (Ear-Acromion-Horizontal).
4. Classification
The optimized feature set was used to train ensemble models, with the Extra Tree Classifier (ETC) emerging as the top performer. The ETC’s randomized branching thresholds provided superior generalization compared to standard Random Forests or Gradient Boosting machines.
Results and Discussion
The study utilized a dataset of 1,651 Chinese adolescents, with ground truth established through expert-annotated clinical photogrammetry.
Model Performance
The Extra Tree Classifier achieved an overall accuracy of 82.4% (95% CI, 75.7% - 87.9%), with a sensitivity of 80.6% and a specificity of 85.5%. Notably, the Positive Predictive Value (PPV) reached 90.2%, making it a highly reliable tool for initial screening.
The Impact of Feature Selection
The ablation study underscored the importance of the Genetic Algorithm. Using all available features resulted in a lower accuracy of 76.4%. The GA-selected subset provided a 6% absolute improvement by filtering out non-contributory variables.
As shown in the figure above, angles like NAH and ENA show distinct distributional differences between FHP and normal subjects, validating their selection as primary predictors.
Conclusion
This research demonstrates that "revitalizing" classical machine learning through rigorous feature engineering and evolutionary optimization can outperform black-box deep learning models in specific medical tasks. The resulting pipeline offers a scalable, low-cost solution for FHP screening in schools and clinics. Future work will focus on expanding the age range of the dataset and exploring 3D pose estimation to capture more complex spinal curvatures.
Summary updated