Clustering and classification of regional peak plantar pressures of diabetic feet

Abstract

High plantar pressures have been associated with foot ulceration in people with diabetes, who can experience loss of protective sensation due to peripheral neuropathy. Therefore, characterization of elevated plantar pressure distributions can provide a means of identifying diabetic patients at potential risk of foot ulceration. Plantar pressure distribution classification can also be used to determine suitable preventive interventions, such as the provision of an appropriately designed insole. In the past, emphasis has primarily been placed on the identification of individual focal areas of elevated pressure. The goal of this study was to utilize k-means clustering analysis to identify typical regional peak plantar pressure distributions in a group of 819 diabetic feet. The number of clusters was varied from 2 to 10 to examine the effect on the differentiation and classification of regional peak plantar pressure distributions. As the number of groups increased, so too did the specificity of their pressure distributions: starting with overall low or overall high peak pressure groups and extending to clusters exhibiting several focal peak pressures in different regions of the foot. However, as the number of clusters increased, the ability to accurately classify a given regional peak plantar pressure distribution decreased. The balance between these opposing constraints can be adjusted when assessing patients with feet that are potentially "at risk" or while prescribing footwear to reduce high regional pressures. This analysis provides an understanding of the variability of the regional peak plantar pressure distributions seen within the diabetic population and serves as a guide for the preemptive assessment and prevention of diabetic foot ulcers.

Figure 1

Masking of foot regions to extract regional peak pressures. Illustration provided on sample plantar pressure data. MTH = metatarsal head.

Figure 2

Success rate of clustering and classification, i.e. the percentage of feet consistently classified between two independent groups (training and testing sets). Upper and lower bounds represent one standard deviation above and below the average success rate for 20 trials.

Figure 3

Calculated cluster pressures (2 ≤ k ≤ 10, k = number of clusters), showing the relative average peak pressures for each foot region visualized with a linear color gradient, from yellow (low pressure) to red (high pressure). The number of feet grouped into each cluster is provided above the average peak pressure map for each cluster. Appendix provides actual mean and standard deviation for the clusters.

Figure 4

Statistical comparison of regional peak pressures following one-way ANOVA tests between clusters. Comparisons tested the differences in regional peak pressures between clusters, at a prescribed number of clusters (2 ≤ k ≤ 10). Darker regions of the foot indicate statistically significant differences in the peak pressure between two clusters (at the intersection of the row and column for the clusters of interest). Refer to Figure 3 for cluster numbering (from 1 to k, going from left to right).

Figure 5

Actual pressure distribution for a total number of clusters set to seven. A representative pressure distribution is shown for each cluster based on proximity to cluster centroid (by Euclidean distance). See Figure 3 for corresponding cluster centroids for k=7. This clustering will likely be useful for footwear design due to its specific identification of various high pressure patterns with reasonable accuracy (also see Figure 2).