System for Prediction of Recurrent DFU
Table 2: Summary of DFU prediction for four foot temperature asymmetry thresholds.
Asymmetry threshold
2.22°C
2.75°C 3.20°C 3.75°C
Sensitivity (%) 97 90 70 50
Specificity (%) 43 57 37 81
3.1 2.2 1.7 1.1
37 ± 18 36 ± 17 35 ± 16 35 ± 17
Positive predictive value (%) 16.6 19.7 1.79 ± 0.08 Negative predictive value (%) 99.2 98.0 111.7 ± 22.2
Alert frequency (per participant/year)
Alert lead time (days)
Data are means ± SD unless otherwise indicated.
and Accountability Act of 1996– compliant servers
managed by the manufacturer. The data are saved
and processed, and the foot temperature asymmetry is
automatically calculated based on the thermogram.
The study device is legally marketed in the U.S. as a
class I medical device (product code OIZ Daily Assist
Device; 510[k] designation K150557) and has been
cleared by the U.S. Food and Drug Administration
for its intended use of“periodic evaluation of the
temperature over the soles of the feet for signs of
inflammation.”
Analysis Plan
We compared two sub-cohorts: those who developed
at least one DFU during the study and those who
remained ulcer free throughout participation. To make
between-group comparisons over continuous variables,
we used the independent t test with Welch correction
for unequal population variances. For comparisons of
proportions between groups, we used the Fisher exact
test to evaluate independence. For all comparisons, we
set a = 0.05 as the threshold for significance. Given
these direct comparisons, we completed a multiple
logistic regression including all variables that were
significant at the a = 0.05 level to minimize the
influence of multicollinearity, which we anticipated to
46
Pedorthic Footcare Association | www.pedorthics.org
be relevant among several covariate subsets.
Effect sizes for continuous variables were reported
using Cohen’s d statistic, and ORs were used for
proportion effect sizes. These were categorized as
“small,”“medium,” and “large” per the conventions
of Cohen (22,23). Specifically, for comparison of
continuous variables, Cohen’s d values of 0.2, 0.5, and
0.8 were considered small, medium, and large effect
sizes, respectively. For comparison of proportions, ORs
of 1.45, 2.5, and 4.3 were considered small, medium,
and large effect sizes, respectively.
To evaluate classification accuracy, we constructed
a receiver operator characteristic (ROC) curve
that defined the sensitivity and specificity of the
prediction as a function of temperature asymmetry
threshold. False-positive and false-negative rates were
calculated over 2-month samples of participant data.
Reporting these statistics over a 2-month interval
allows for a more clinically meaningful and consistent
interpretation of the results commensurate with a
hypothesized duration between office visits for a high-
risk patient. Another benefit of this approach is that it
implicitly weights the outcomes for each participant
by the quantity of data collected for that participant,
naturally handling participants with censored data
because of developing a clinical contraindication. This