The presence of occipital hair in the pilonidal sinus cavity—a triple approach to proof

Abstract

Purpose

Hair in the pilonidal sinus is not growing within the sinus cavity, as hair follicles are not present there. Not few pilonidal patients do not have intergluteal hair, which is said to be the causative agent of folliculitis and pilonidal genesis. So, what is the real source of the hair forming the typical pilonidal hair nest?

Methods

A trifold approach was used: First, axial hair strength testing of pilonidal hair and body hair harvested from head, lower back (glabella sacralis), and cranial third of intergluteal fold. Hair strength match was compared clinically. Second, comparative morphological examination by expert forensic biologist of hair from sinus and dorsal body hair. Third, statistical Bayesian classification of every single sinus hair based on its strength was done to determine the most probable region of origin.

Results

Using clinical hair strength comparison, in 13/20 patients, head hair is the stiffest hair, followed by intergluteal hair. Only in 6/20 patients, this is the case with hair from the glabella sacralis. According to comparative morphological comparison, a minimum of 5 of 13 hair nests with possible hair allocation examined contain hair from the occiput. In 5/18 nests, hair could not be determined to a specific location though. Statistical classification with correction for multiple testing shows that 2 nests have hair samples that are at least 100 times more probable to originate from head or lower back than from intergluteal fold.

Conclusion

We saw our null hypothesis that “hair in the sinus cavity is from the intergluteal region” rejected by each of three different approaches. There is strong evidence that occipital hair is present regularly in pilonidal sinus nests. We should start thinking of occipital hair as an important hair source for the development of the pilonidal hair nest.

Appendix. Statistical remarks

Aggregating hair strength measurements

The strength of every hair was measured up to six times in sequence; we used a linear model with the hair as a factor variable and the measurement number as a numerical variable to aggregate multiple measurements into a single strength value for each hair (see the section “Statistical analysis”). The most important assumption of this model, the homoscedasticity assumption, was fulfilled by the data (see Fig. 2, left), but the distribution of the residuals was rather long-tailed and slightly skewed than normal (see Fig. 2, right). Despite this fact, we considered the overall model fit good enough to be used for aggregating multiple strength measurements for hair samples.

Fig. 2

Tukey-Anscombe plot (residuals vs. fitted values) and residual quantiles vs. normal quantiles. Plots for validation of the linear model used for aggregating repeated hair strength measurements

Classification of pilonidal sinus hair to regions of origins

We performed linear discriminant analysis (LDA) to statistically attribute pilonidal sinus hair of each patient to their most probable region of origin (protuberantia occipitalis externa [POE], glabella sacralis [GS], or intergluteal fold [IGF], respectively). This proceeded as follows:

• 1. We fitted one LDA model per patient, using hair strengths from POE, GS, and IGF samples to train the model. The LDA model assumes hair strengths of each region of origin are normally distributed with equal variance, but different means.

• 2. We classified each pilonidal sinus hair of a patient by calculating its posterior probabilities to originate from POE, GS, and IGF, respectively. This calculation was done based on the normal distributions of hair strength fitted in step 1, and based on the prior probabilities, we assigned to each region of origin; we assigned equal prior probability to each region of origin (POE, GS, and IGF; “objective” or “uninformed” prior.

• 3. In order to quantify the uncertainty in the calculated posteriors, we performed non-parametric bootstrap. That means, we resampled hair samples for each patient 999 times (random sampling with replacement), and repeated steps 1 and 2 above for every simulated sample. We used the empirical distribution of the 999 bootstrap posteriors calculated like this to derive 95% confidence intervals for the estimated posteriors.

This analysis has several potentially delicate points:

• Assumption of a model (normally distributed hair strength with equal variance in each region of origin) that may or may not be appropriate. By visual inspection of the distribution of hair strengths in different regions (POE, GS, and IGF), we convinced ourselves that the assumption of this model seems safe.

• Selection bias. In order to reliably fit an LDA model, we would need (random) hair samples from the three regions (POE, GS, and IGF). However, it is not possible to ensure random sampling in practice: hair are chosen and harvested by human researchers. The sampling strategy that was applied for most patients actually contradicted this requirement: in the beginning of the study, hair were sampled based on visual strength, having forensic examination and visual inspection in mind. To estimate the validity of models trained on these non-random hair samples, new samples were collected from three patients (#1, #4, #15), this time making sure selection is as random as possible for a human researcher. We compared these randomly selected hair samples to the older samples selected based on visible strength.

Interestingly, the hair samples collected with the aim of selecting strong hairs from the different regions were on average not stronger than the random samples (Fig. 3). When comparing hair strength from random and non-random sampling with T tests, we got unadjusted p values of 0.018 (patient #1, IGF; mean hair force from randomly sampled hair is larger), 0.045 (patient #15, GS; mean hair force from randomly sampled hair is larger), and above 0.05 in other cases. After adjusting the p values for multiple testing with the Holm correction, the minimal (adjusted) p value was 0.160. We concluded that the problem of non-random selection of hair samples can be neglected in practice. It seems that it is difficult to determine the hair strength by eye, and that the samples of supposedly strong hair are indeed representative for the whole range of hair strengths of the patients, which is good for fitting the LDA models.

Fig. 3

Comparison of hair strength measurements for randomly harvested hair (black) and hair selected by visible strength (red)

• Small sample size. Most models are based on only 6 hair from each region (POE, GS, and IGF); the calculated posteriors are therefore not very precise. We performed the non-parametric bootstrap (see above) in order to quantify the uncertainty. As expected, confidence intervals are large for most hair samples, and most pilonidal sinus hair cannot be statistically attributed to a single region of origin with high probability (see also detailed results in Fig. 3).

Classification results for all pilonidal sinus hair samples are depicted in Fig. 4.

Fig. 4

Posterior probabilities of all classified hair to originate from POE, GS, or IGF, respectively. Numbers in brackets denote the 95% confidence intervals calculated by bootstrap. Red cells refer to a high probability and green cells to a low probability. Patients not shown here did not have a sinus hair sample (patients #7, #12), or no IGF hair sample (patients #2, #14; the null hypothesis that sinus hair stems from IGF cannot be tested in this case). Reading example for patient #1: hair no. 1 is assigned a probability between 43 and 90% of originating from POE, 2 to 27% of originating from GS, and 4 to 36% of originating from IGF