Abstract
High-heeled shoes of excessive height can severely injure shoe users, causing various conditions such as musculoskeletal pain, osteoarthritis, and hallux valgus. While a physician can estimate an appropriate heel height by assessing calcaneal varus deformation, this professional assistance is typically unavailable to consumers during shoe selection. This study presents a novel system for evaluating the maximum safe heel height for individual female wearers. The system utilizes images of the Achilles tendon, medial and lateral longitudinal arches, and plantar pressure distributions as inputs. Trained on these datasets using Basic CNN, VGG16, and MobileNetV2 architectures, the system achieved an accuracy of 0.88. A subsequent statistical algorithm was applied to refine the AI model's outputs, resulting in a mean absolute error (MAE) of 1.21 cm—less than the typical height increment for commercial high-heeled shoes.
Introduction
High-heeled shoes are a staple of fashion and professional attire, yet wearing heels that exceed an individual's physiological limits can lead to severe long-term health issues. Clinically, a physician determines a "safe" height by observing calcaneal varus—the specific way the heel bone tilts or deforms as it is elevated. However, most consumers lack access to objective, automated tools that can translate these complex clinical observations into practical recommendations. This research addresses this gap by developing an AI-powered measurement platform designed to automate the clinical assessment process.
System Architecture and Methodology
The proposed system consists of a custom hardware-software stack designed to capture both visual and tactile data from the wearer's foot at various elevations.
Hardware Components
As shown in the system architecture, the setup includes:
- The Measurement Platform: A motorized lifting mechanism that elevates the foot to simulate different heel heights. It is equipped with 42 force-sensing resistor (FSR) sensors to map plantar pressure and six USB webcams to capture foot angles from multiple perspectives.
- Mainboard and Host: A central processing unit (Mainboard) manages the sensors and motor, while a host computer processes the visual and pressure data.
AI and Statistical Modeling
The researchers evaluated three Convolutional Neural Network (CNN) architectures—Basic CNN, VGG16, and MobileNetV2—to classify foot positions as either "wearable" (safe) or "non-wearable" (unsafe). To address potential noise in AI classifications (e.g., a model incorrectly labeling a 7cm height as safe while labeling a 6.5cm height as unsafe), a mode-based statistical algorithm was implemented. This smoothing technique analyzes adjacent height increments to identify the true "critical point" of instability.
Data Collection and Insights
The study involved 175 female subjects, with 100 passing medical screening to ensure a baseline of healthy foot physiology. A total of 4,200 data points, comprising images and pressure maps, were collected. Professional physicians provided the ground truth labels by manually evaluating each subject on the platform.
Several key insights emerged during the training process:
- The Achilles Advantage: Images of the Achilles tendon alone yielded the highest accuracy (0.88), suggesting that tendon deformation is the most reliable visual proxy for calcaneal varus.
- The "Matting" Paradox: Contrary to standard practices where background removal (matting) improves model focus, non-matted images performed better here. The background likely served as a consistent reference point for the AI to perceive subtle changes in foot position.
- Pressure vs. Vision: Plantar pressure data proved less effective than visual images for classification, likely due to idiosyncratic variations in foot size and weight distribution among subjects.
Key Results
The Basic CNN architecture, utilizing Achilles tendon images, emerged as the superior model with an accuracy of 0.88 and an Area Under the Curve (AUC) of 0.96. The model demonstrated a high recall of 0.92, ensuring it effectively identifies safe heights without being overly restrictive.
Crucially, the final system achieved a Mean Absolute Error (MAE) of 1.21 cm. Given that commercial high heels typically increase in height by approximately 2.54 cm (1 inch) increments, an error margin of ~1.2 cm is small enough to provide practically useful recommendations for consumers.
Limitations and Future Work
While promising, the system's current iteration has limitations. The dataset was restricted to females with foot lengths between 22 and 26 cm, and the computer vision model requires a bright, shadow-free environment. Future research will focus on generalizing the model to a wider range of demographics and foot sizes, as well as exploring video-based dynamic analysis to capture gait stability in real-time.
Reference: Lee, S. H., Lin, B. S., et al. (2021). "Artificial Intelligence-Based Assessment System for Evaluating Suitable Range of Heel Height." IEEE Access, vol. 9, pp. 38374-38385.
Summary updated