Classifying high-dimensional phenotypes with ensemble learning.

Classification is a fundamental task in biology used to assign members to a class. While linear discriminant functions have long been effective, advances in phenotypic data collection are yielding increasingly high-dimensional datasets with more classes, unequal class covariances, and non-linear distributions. Numerous studies have deployed machine learning techniques to classify such distributions, but they are often restricted to a particular organism, a limited set of algorithms, and/or a specific classification task. In addition, the utility of ensemble learning or the strategic combination of models has not been fully explored.We performed a meta-analysis of 33 algorithms across 20 datasets containing over 20,000 high-dimensional shape phenotypes using an ensemble learning framework. Both binary (e.g., sex, environment) and multi-class (e.g., species, genotype, population) classification tasks were considered. The ensemble workflow contains functions for preprocessing, training individual learners and ensembles, and model evaluation. We evaluated algorithm performance within and among datasets. Furthermore, we quantified the extent to which various dataset and phenotypic properties impact performance.We found that discriminant analysis variants and neural networks were the most accurate base learners on average. However, their performance varied substantially between datasets. Ensemble models achieved the highest performance on average, both within and among datasets, increasing average accuracy by up to 3% over the top base learner. Higher class R2 values, mean class shape distances, and between- vs. within-class variances were positively associated with performance, whereas higher class covariance distances were negatively associated. Class balance and total sample size were not predictive.Learning-based classification is a complex task driven by many hyperparameters. We demonstrate that selecting and optimizing an algorithm based on the results of another study is a flawed strategy. Ensemble models instead offer a flexible approach that is data agnostic and exceptionally accurate. By assessing the impact of various dataset and phenotypic properties on classification performance, we also offer potential explanations for variation in performance. Researchers interested in maximizing performance stand to benefit from the simplicity and effectiveness of our approach made accessible via the R package pheble.

Gains in relative cortical area during growth and their relationship to nutrition, body size, and physical activity.

OBJECTIVES: In studies of growth in the past, low percentage of cortical area (%CA) is commonly attributed to poor general health, due to factors including poor nutrition, low socioeconomic status, or other physiological stressors. What constitutes low relative cortical dimensions has not been established across a diverse range of human skeletal samples. This study examines %CA in a large immature skeletal sample to establish typical variation in humans with consideration of both body mass and subsistence strategy. MATERIALS AND METHODS: Percentage of cortical area was calculated at the midshaft of the humerus, femur, and tibia in seven skeletal samples. Age at death was estimated from dental development, and body mass from bone dimensions. Patterns of %CA with age and log-transformed body mass were examined in the pooled sample and compared among samples using LOESS regression, Welch's ANOVA, and Kruskal-Wallis tests. RESULTS: Across all samples, %CA displays a generally non-linear pattern, but variation in %CA with age was high, particularly in samples with lower levels of %CA. There was no relationship between %CA and age-adjusted body mass. DISCUSSION: The lack of a relationship between %CA and body mass suggests that %CA should not be used as an indicator of mechanical loading. The variation present across samples implies that appositional bone growth is affected by physiological stress in varying ways. Without a deeper understanding of what is "typical" for long bone development, it is impossible to draw conclusions about individual or population level health.

The developmental impacts of natural selection on human pelvic morphology.

Evolutionary responses to selection for bipedalism and childbirth have shaped the human pelvis, a structure that differs substantially from that in apes. Morphology related to these factors is present by birth, yet the developmental-genetic mechanisms governing pelvic shape remain largely unknown. Here, we pinpoint and characterize a key gestational window when human-specific pelvic morphology becomes recognizable, as the ilium and the entire pelvis acquire traits essential for human walking and birth. We next use functional genomics to molecularly characterize chondrocytes from different pelvic subelements during this window to reveal their developmental-genetic architectures. We then find notable evidence of ancient selection and genetic constraint on regulatory sequences involved in ilium expansion and growth, findings complemented by our phenotypic analyses showing that variation in iliac traits is reduced in humans compared to African apes. Our datasets provide important resources for musculoskeletal biology and begin to elucidate developmental mechanisms that shape human-specific morphology.

Effects of osteoarthritis on age-at-death estimates from the human pelvis.

OBJECTIVES: This study examined the degree of error in age-at-death estimates when osteoarthritis (OA) is present in three separate pelvic joint areas: (1) the pubic symphysis, (2) the auricular surface, and (3) the acetabulum from a modern known-age European cemetery sample of adults. MATERIALS AND METHODS: Age-at-death ranged from 17 to 79 years (x̄ =50.9 years; n = 252). OA in the pelvic joints was evaluated using standard ranked categorical scoring. Composite OA scores were derived through principal component analysis. Blind age assessments and all analyses were performed separately by region. Error between adult age groups (young, middle, old) and between OA severity groups (low, middle, high) was evaluated using one-way ANOVAs with post-hoc testing, ordinary least squares regression, and transition analysis with a cumulative probit model. Ages-at-transition were compared with Nphases2. RESULTS: Three significant results emerge. First, OA severity has an effect on the accuracy of age estimates from os coxa joints in this sample. Second, this influence is most significant for different age cohorts in each joint region, demonstrating that varied rates of arthritic trait progression occur between the auricular surface, pubic symphyses, and acetabulum. Third, those with OA appear to be aging faster, a consistent trend among the os coxa regions. CONCLUSIONS: These results have significant consequences for understanding the rate of bone remodeling in relation to disease, aging, and the evaluation of skeletal age indicators.

The relationship of age, activity, and body size on osteoarthritis in weight-bearing skeletal regions.

This study examined the simultaneous impact of multiple underlying factors on OA expression in weight-bearing joints of the vertebrae and lower limb of a modern European skeletal sample (Lisbon and Sassari). OA was evaluated using standard ranked categorical scoring; composite OA scores derived through principal component analysis. Body size was calculated from postcranial measurements; torsional strength (J) of the femoral midshaft was calculated from three-dimensional surface models, size standardized and used as a proxy for activity. A standard multiple regression was applied. In all regions, the linear combination of age, body mass, stature, and J was significantly related to differences in OA. Across all joints, age was the strongest predictor; neither body size, nor activity variables demonstrated a statistical relationship with OA at the lumbar or knee; J demonstrated a negative correlation with pelvic OA. Variation in OA can be explained by age, stature, body mass, and structural adaptation related to habitual use. The negative correlation between femoral torsional strength with OA suggests that long-term, repetitive physical work capacity in childhood may be protective against OA development later in life. The multifactorial aetiology of OA requires incorporating multiple lines of evidence to interpret individual or population health from bone samples.

Bilateral Asymmetry in the Human Pelvis.

Asymmetry of the human axial skeleton has received much less attention that of the limb skeleton. Pelvic morphology is subject to multiple selective factors, including bipedal locomotion and obstetrics, among others, as well as environmental factors such as biomechanical loading. How these various factors influence or restrict asymmetry of the pelvis is unknown and few studies have investigated levels and patterns of pelvic asymmetry. This study examines percentage directional (%DA) and absolute (%AA) asymmetry in 14 bilaterally paired dimensions of the pelvic canal, non-canal pelvis, and femur in female (n = 111) and male (n = 126) skeletons from nine geographically dispersed skeletal samples. Directional asymmetries were uniformly low for all measures and lacked any consistent patterning across the variables, while %AA was highest in the pelvic canal, particularly the posterior aspects. Few sex differences and no population differences were found for %DA and %AA; however the latter was correlated with coefficients of variation across the 14 variables in both sexes. While sample mean %DA were low, standard deviations of the canal variables were high and the majority of individuals in both sexes displayed %DA values >±0.5, suggesting asymmetry is common, if not directionally consistent. Biomechanical loading of the pelvic girdle may influence asymmetry of both the canal and non-canal aspects of the pelvis; however it is unlikely that these asymmetries negatively affect obstetric function, given the prevalence for %DA found in this study. Anat Rec, 300:653-665, 2017. © 2017 Wiley Periodicals, Inc.

Principal component analysis in the evaluation of osteoarthritis.

OBJECTIVES: The purpose of this study is to demonstrate advantages of principal component analysis (PCA) as a standardized procedure in the evaluation of osteoarthritis (OA) in a skeletal series to: (1) compute aggregate scores for joint complexes that accurately capture pathological expression, (2) reveal which variables describe the most sample variation in OA, (3) enable inter- and intra-sample comparison of results, and (4) formulate predictive models from component-based arthritic scores. MATERIALS AND METHODS: The sample (144 males, 145 females) is drawn from a large skeletal cemetery collection of modern Europeans of known sex, age, and occupation. OA data was collected using standard ranked categorical scoring. PCA was conducted separately on lumbar spine, pelvis, and knee regions to generate composite OA scores from eigenequations of the first and second principal components (PC). RESULTS: Results demonstrate that as severity in OA increases, so does the distribution of OA within the joint surface. In each region, PCA produced the same general pattern with eburnation scoring driving significant changes in composite OA scores, representing earlier to later stages of cartilage degeneration. The distribution of arthritic traits determined by PCA produced an OA score that quantifies the expression of joint changes in varied biological joint structures from most moveable to least mobile, the final stage being joint fusion. OA scores are most highly variable in the lumbar region for both males and females, as compared to the pelvis and knee. CONCLUSIONS: PCA is a simple, non-parametric method of extracting relevant information from complex OA datasets and summarizes variation based on correlated multi-attributes to reveal a simplified structure of OA expression. Multivariate techniques like PCA should be used to describe discrete OA samples, and are useful to compute population-specific representative measurements for idiopathic joint OA in a skeletal sample.

Patterns of directional asymmetry in the pelvis and pelvic canal.

OBJECTIVES: The human pelvis is unique among modern taxa for supporting both parturition of large brained young and obligate bipedalism. Though much work has focused on pelvic development and variation, little work has explored the presence or absence of asymmetry in the pelvis despite well-known patterns of asymmetry in other skeletal regions. This study investigated whether patterns of directional asymmetry (DA) could be observed in the pelvis or pelvic canal. METHODS: Seventeen bilaterally paired osteometric measurements of the os coxae (34 measures in total) were taken from 128 skeletons (female n = 65, male n = 63) from recent human populations in five geographic regions. Paired sample t-tests and Mann-Whitney U-tests were used to investigate DA. RESULTS: Results from a pooled sample of all individuals showed that the pelvis exhibited a left-bias in DA. In contrast, the pelvic canal exhibited a pattern in which the anterior canal exhibited a right-bias and the posterior canal exhibited a left-bias. Neither sex nor populational differences in DA were observed in the pelvis or pelvic canal. CONCLUSIONS: The varying patterns of asymmetry uncovered here accord with prior work and may indicate that loading from the trunk and legs place differing stresses on the pelvis and canal, yielding these unequal asymmetries. However, this is speculative and the possible influence of genetics, biomechanics, and nutritional status on the development of pelvic and canal asymmetries presents a rich area for future study. Additionally, the potential influence of pelvic canal asymmetry on obstetric measures of pelvic capacity merits future research. Am. J. Hum. Biol. 28:804-810, 2016. © 2016Wiley Periodicals, Inc.

Shape variation in the human pelvis and limb skeleton: Implications for obstetric adaptation.

OBJECTIVES: Under the obstetrical dilemma (OD) hypothesis, selection acts on the human female pelvis to ensure a sufficiently sized obstetric canal for birthing a large-brained, broad shouldered neonate, while bipedal locomotion selects for a narrower and smaller pelvis. Despite this female-specific stabilizing selection, variability of linear dimensions of the pelvic canal and overall size are not reduced in females, suggesting shape may instead be variable among females of a population. Female canal shape has been shown to vary among populations, while male canal shape does not. Within this context, we examine within-population canal shape variation in comparison with that of noncanal aspects of the pelvis and the limbs. MATERIALS AND METHODS: Nine skeletal samples (total female n = 101, male n = 117) representing diverse body sizes and shapes were included. Principal components analysis was applied to size-adjusted variables of each skeletal region. A multivariate variance was calculated using the weighted PC scores for all components in each model and F-ratios used to assess differences in within-population variances between sexes and skeletal regions. RESULTS: Within both sexes, multivariate canal shape variance is significantly greater than noncanal pelvis and limb variances, while limb variance is greater than noncanal pelvis variance in some populations. Multivariate shape variation is not consistently different between the sexes in any of the skeletal regions. DISCUSSION: Diverse selective pressures, including obstetrics, locomotion, load carrying, and others may act on canal shape, as well as genetic drift and plasticity, thus increasing variation in morphospace while protecting obstetric sufficiency.

Discernment of mortality risk associated with childbirth in archaeologically derived forager skeletons.

An obstetric dilemma may have been a persistent characteristic of human evolution, in which the bipedal female's pelvis is barely large enough to accommodate the birth of a large-brained neonate. Evidence in the archaeological record for mortality risk associated with childbirth is rare, especially among highly mobile, immediate return hunter-gatherer populations. This research explores the idea that if excess mortality is associated with first pregnancy, females will outnumber males among young adult skeletons. The sample is of 246 skeletons (119 males, 127 females) representing Later Stone Age (LSA) foragers of the South African Cape. Young adults are distinguished through incomplete maturation of the medial clavicle, iliac crest and vertebral bodies. With 26 women and 14 men in the young category, a higher mortality risk for women is suggested, particularly in the Southern Cape region. Body size does not distinguish mortality groups; there is evidence of a dietary protein difference between young and older women from the Southern Cape. Possible increased mortality associated with first parturition may have been linked to morphological or energetic challenges, or a combination of both. Exploration of the sex ratio among young adult skeletons provides a tool for exploring the antiquity of an important evolutionary factor.

Skeletal variability in the pelvis and limb skeleton of humans: does stabilizing selection limit female pelvic variation?

OBJECTIVES: This study tests the hypothesis, a correlate of the obstetric dilemma, that skeletal variability in the human female pelvic canal is limited owing to the action of stabilizing selection. Levels of variation in three skeletal regions (pelvic canal, noncanal pelvis, and limbs) of females and males are compared to each other and between sexes. METHODS: Nine human skeletal samples (total female n = 101; male n = 117) representing diverse populations were included. Osteometric data were collected from the articulated pelvis, os coxa, sacrum, femur, tibia, humerus, radius, and clavicle. Coefficients of variation, adjusted for small sample size (V*), were calculated for variables in separate samples by sex, and mean V*s were taken for the skeletal regions. Size variances were measured as V* of the geometric mean (GM) of the skeletal region variables. Using nonparametric methods, coefficients were compared between sexes and skeletal regions and correlations among V*s were calculated. RESULTS: Females and males do not differ in levels of variation for any skeletal region. The pelvic canal is the most variable region in both sexes, while size variability (GM) is similar among the three skeletal regions. Across the samples, canal and noncanal pelvic regions share patterns of variability in females but not males, while variability of the limb skeleton is independent in both sexes. CONCLUSIONS: The results suggest that stabilizing selection does not limit variability in the female pelvic canal. Biological plasticity may be greater in the canal than that in other skeletal regions.

Bony pelvic canal size and shape in relation to body proportionality in humans.

Obstetric selection acts on the female pelvic canal to accommodate the human neonate and contributes to pelvic sexual dimorphism. There is a complex relationship between selection for obstetric sufficiency and for overall body size in humans. The relationship between selective pressures may differ among populations of different body sizes and proportions, as pelvic canal dimensions vary among populations. Size and shape of the pelvic canal in relation to body size and shape were examined using nine skeletal samples (total female n = 57; male n = 84) from diverse geographical regions. Pelvic, vertebral, and lower limb bone measurements were collected. Principal component analyses demonstrate pelvic canal size and shape differences among the samples. Male multivariate variance in pelvic shape is greater than female variance for North and South Africans. High-latitude samples have larger and broader bodies, and pelvic canals of larger size and, among females, relatively broader medio-lateral dimensions relative to low-latitude samples, which tend to display relatively expanded inlet antero-posterior (A-P) and posterior canal dimensions. Differences in canal shape exist among samples that are not associated with latitude or body size, suggesting independence of some canal shape characteristics from body size and shape. The South Africans are distinctive with very narrow bodies and small pelvic inlets relative to an elongated lower canal in A-P and posterior lengths. Variation in pelvic canal geometry among populations is consistent with a high degree of evolvability in the human pelvis.

Allometry of head and body size in Holocene foragers of the South African Cape.

Opportunities to assess morphological allometry in small-bodied human populations are rare. The foragers of the Later Stone Age of the South African Cape are characteristically small-bodied. Previous studies have shown that during the period of ca. 3500 to 2000 years BP (uncalibrated (14) C dates), the regional population shows transient reduced stature, body mass, and cranial size, a pattern that has been tentatively tied to demographic pressure on resources. This study examines the relationships among cranial size (centroid size) and body size (femoral length, femoral head diameter, and bi-iliac breadth) during the second half of the Holocene (N = 62). Reduced major axis regression indicates negative allometry of cranial centroid size with body size. Residuals (from ordinary least squares regression of cranial centroid size on body size) are regressed on radiocarbon date to examine temporal changes in the relationship between cranial and body size. Cranial and pelvic sizes are most conserved through time, while more ancient skeletons possess shorter femora and smaller femoral heads. The relationship between cranial centroid size and femoral length shows larger and more variable residuals at more recent dates, indicating a greater or more variable disassociation between cranial size and stature relative to more ancient skeletons. A similar, but nonsignificant relationship exists between cranial size and bi-iliac breadth. These results provide insights into the use of aspects of body size and proportionality in the assessment of health in past populations.

Pelvic dimorphism in relation to body size and body size dimorphism in humans.

Many mammalian species display sexual dimorphism in the pelvis, where females possess larger dimensions of the obstetric (pelvic) canal than males. This is contrary to the general pattern of body size dimorphism, where males are larger than females. Pelvic dimorphism is often attributed to selection relating to parturition, or as a developmental consequence of secondary sexual differentiation (different allometric growth trajectories of each sex). Among anthropoid primates, species with higher body size dimorphism have higher pelvic dimorphism (in converse directions), which is consistent with an explanation of differential growth trajectories for pelvic dimorphism. This study investigates whether the pattern holds intraspecifically in humans by asking: Do human populations with high body size dimorphism also display high pelvic dimorphism? Previous research demonstrated that in some small-bodied populations, relative pelvic canal size can be larger than in large-bodied populations, while others have suggested that larger-bodied human populations display greater body size dimorphism. Eleven human skeletal samples (total N: male = 229, female = 208) were utilized, representing a range of body sizes and geographical regions. Skeletal measurements of the pelvis and femur were collected and indices of sexual dimorphism for the pelvis and femur were calculated for each sample [ln(M/F)]. Linear regression was used to examine the relationships between indices of pelvic and femoral size dimorphism, and between pelvic dimorphism and female femoral size. Contrary to expectations, the results suggest that pelvic dimorphism in humans is generally not correlated with body size dimorphism or female body size. These results indicate that divergent patterns of dimorphism exist for the pelvis and body size in humans. Implications for the evaluation of the evolution of pelvic dimorphism and rotational childbirth in Homo are considered.

Compromised skeletal growth? Small body size and clinical contraction thresholds for the female pelvic canal.

Small body size has been used in bioarchaeological contexts as an indicator of poor health and in clinical contexts as a risk factor for difficult childbirth, under the assumption that small size reflects compromised growth. Compromised growth in females may result in contracted pelvic dimensions and elevated risk of cephalopelvic disproportion. The basis for the use of clinical pelvic contracture thresholds for evaluating health and growth in skeletal populations is examined through comparisons of contracture frequencies and the relationships between pelvic canal size and body size in 11 skeletal samples (total n=195). The small-bodied samples show higher frequencies of contracted inlet and midplane posterior space dimensions. Canal breadth is correlated with femoral head diameter and bi-iliac breadth, but not femoral length. These results suggest that modern clinical standards do not take into account the variation in human body size and shape and the effect this variation may have on obstetric capacity without compromising obstetric function. It is problematic to use small pelvic dimensions, in the absence of evidence of death during childbirth, to infer compromised obstetric function resulting from biological stress, as is assuming that small size represents compromised skeletal growth.

Protection of obstetric dimensions in a small-bodied human sample.

In human females, the bony pelvis must find a balance between being small (narrow) for efficient bipedal locomotion, and being large to accommodate a relatively large newborn. It has been shown that within a given population, taller/larger-bodied women have larger pelvic canals. This study investigates whether in a population where small body size is the norm, pelvic geometry (size and shape), on average, shows accommodation to protect the obstetric canal. Osteometric data were collected from the pelves, femora, and clavicles (body size indicators) of adult skeletons representing a range of adult body size. Samples include Holocene Later Stone Age (LSA) foragers from southern Africa (n = 28 females, 31 males), Portuguese from the Coimbra-identified skeletal collection (CISC) (n = 40 females, 40 males) and European-Americans from the Hamann-Todd osteological collection (H-T) (n = 40 females, 40 males). Patterns of sexual dimorphism are similar in the samples. Univariate and multivariate analyses of raw and Mosimann shape-variables indicate that compared to the CISC and H-T females, the LSA females have relatively large midplane and outlet canal planes (particularly posterior and A-P lengths). The LSA males also follow this pattern, although with absolutely smaller pelves in multivariate space. The CISC females, who have equally small stature, but larger body mass, do not show the same type of pelvic canal size and shape accommodation. The results suggest that adaptive allometric modeling in at least some small-bodied populations protects the obstetric canal. These findings support the use of population-specific attributes in the clinical evaluation of obstetric risk.