PropertyValue
is nif:broaderContext of
nif:broaderContext
is schema:hasPart of
schema:isPartOf
nif:isString
  • This was a secondary data analysis of the “Reach for Health” clinical trial carried out at the University of California, San Diego (UCSD). The original study was approved by the UCSD IRB board, project #101977. All subjects in the Reach for Health study provided written consent. The National Institutes of Health ClinicalTrials.gov identifier is NCT01302379. Our study sample comprised 333 early-stage breast cancer survivors enrolled in a weight-loss intervention. The study protocol and design have been previously published [25]. Briefly, the study enrolled breast cancer survivors, who were postmenopausal at cancer diagnosis, were either overweight or obese at study entry, and had completed primary breast cancer treatment (surgery with or without chemotherapy and radiation). 83% were white; 11% were Hispanic. More information on demographics, lifestyle, clinical factors, coping, sleep, mood, physical factors, and biomarkers is provided in Table 1. The current analysis used baseline information to develop network models. Table data removed from full text. Table identifier and caption: 10.1371/journal.pone.0202923.t001 Characteristics of the Reach for Health cohort of overweight postmenopausal breast cancer survivors (N = 333). + To convert insulin pg/mL to pmol/L multiply by 0.172++Weartime adjusted We obtained participants’ medical records including tumor characteristics (Cancer Stage, hormone receptor status) and years from cancer diagnosis to study entry (YrsDXRND). During clinic visits, participants’ height and weight were measured and used to calculate BMI. Physical activity (PA) and sedentary behavior (SB) were determined by objective 7-day, minute-level triaxial accelerometer counts. Specifically, PA was the average (across days) of total counts per minute per day, thus representing a measure that captured total volume of activity; moderate vigorous physical activity (MVPA) was the average of minutes per day with counts ≥ 1952; SB was the average of minutes per day with counts < 100. Accelerometer-derived measures were adjusted for device wear-time. Demographic information and other study measures were obtained through self-report or questionnaires. The Neighborhood Environment Index (Neighborhood) derived from the NEWS scale [28] was used to measure walkability. It has a range from 0 to 6, with higher scores indicating more walkable neighborhood. Sleep quality was evaluated based on the PROMIS scale [29]. In the current analysis, we used two subscales, the sleep disturbance (sleep1), and the sleep impairment (sleep2) subscales. These subscales were normed to mean 50 with standard deviation of 10. Higher scores indicated worse sleep. Quality of life assessment, both mental (QOLm) and physical (QOLp), used the SF-36 scale [30]. QOLm and QOlp scores from 0 to 100, with higher scores reflecting better quality of life. The Monitor-Blunter (MB) scale assessed participants’ coping mechanism. It ranged from -16 to 16, with higher scores indicating more monitor than blunter. Fasting plasma CRP and insulin concentrations were measured using immune-based assays (Meso Scale Discovery), as described previously [25]. We fit a Bayesian network to examine multivariate relationships between demographics, clinical factors, health behaviors and health outcomes. We disallowed implausible edge directions while learning the network structure. Specifically, we disallowed QOLp and QOLm to be the parent nodes of any other variable in the network; and we disallowed age, education, cancer stage, years between diagnosis and study entry, and neighborhood to be the child nodes of any other variable. We applied bootstrap resampling to learn a set of 500 network structures. We then averaged these networks in an attempt to reduce the impact of locally optimal (but globally suboptimal) networks on learning and inference. The averaged network is a more robust model with better predictive performance than choosing a single, high-scoring network [24]. To quantify stability of inferred edges, we computed arc strength and direction strength. Arc strength was calculated as the frequency of an edge occurring between two variables across the 500 bootstrapped network structures; similarly, directional strength was assessed as the frequency of the observed direction re-occurring in the set of learned network structures in which the relevant edge occurred. The averaged network was created using the arcs whose strength exceeded a threshold, which was computed by searching for the arc set “closest” to the arc strength computed from the original data [24]. Conditional independencies were inferred using Markov blankets and related Bayesian network theory. We used Bayesian information criteria (BIC) and posterior model probabilities to compare fit of candidate networks. The BIC was computed as logLik(M)– 0.5*k*log (n), where logLik(M) is the log-likelihood of model M, k is the number of parameters in M, n is the sample-size. This is the classic definition rescaled by -2; hence, in our calculations, higher BIC scores indicate better fit. We also calculated the Bayes factor, which is the ratio of the posterior probabilities (given the observed data) of the first to the second model, as another metric to compare the two models. The log of the Bayes factor can be approximated as the difference in the BIC scores as defined above [31]. Finally, we used logic sampling [24] to study how small perturbations got propagated through the network. In other words, using Monte Carlo simulations, we evaluated how changes in one part of the network could influence distributions in another part of the network, and thus potentially predict the impact of manipulating specific variables. Biomarkers were log-transformed to better approximate Gaussian assumptions. Models were fitted using the R package bnlearn [32].
rdf:type