Post-diagnostic health behaviour scores and risk of prostate cancer progression and mortality

Study sample

CaPSURE is a longitudinal observational cohort of 15,310 men with biopsy-proven prostate cancer. Men diagnosed between 1999 and 2018 at any of 43 participating urology practices across the US were eligible. Participating urologists provided data on clinical and pathological features, treatments, and clinical follow-up. Additional details on CaPSURE are reported elsewhere [15]. The study was conducted in accordance with the Belmont Report and U.S. Common Rule under local Institutional Review Board approval, with all participants providing written informed consent.

The CaPSURE Diet and Lifestyle (CDL) sub-study—consisting of a comprehensive lifestyle questionnaire and full-length food frequency questionnaire (FFQ)—was administered at three time points between 2004 and 2016; a total of 2891 men participated in at least one administration. For the subset of men who completed more than one questionnaire (n = 443), only the first administration (closest to diagnosis date) was used. We excluded men with last clinical follow-up or documented progression prior to completing their first CDL questionnaire (n = 551). Consistent with the recommended approach to address implausible energy intakes [16], we excluded men with extreme (<800 kcal/day or >4200 kcal/day) or unknown caloric intake (n = 153) and/or missing ≥70 FFQ items (n = 20). Finally, we excluded men without a discernable clinical T-stage (n = 100) or with a clinical T-stage >T3a (n = 8) and those with death from unknown cause (n = 3). These exclusions left us with a sample size of 2056 men for our primary analyses of prostate cancer progression. Following the exclusion of men with documented progression prior to completing the first questionnaire, the subsequent exclusions resulted in the loss of 23 events, 2 of which were PCSM. For PCSM analyses, men who were excluded due to documented progression prior to completion of the CDL questionnaire were included—death could not occur prior to completing the questionnaire—resulting in a sample size of 2447 men.

Diet and lifestyle questionnaire

Dietary intake was self-reported on a validated [17,18,19] semiquantitative FFQ, wherein men reported how frequently they consumed a standard unit or portion size of approximately 140 different items. The nine frequency options ranged from never or less than once per month to six or more times per day. FFQ data were sent to the Nutrition Department at the Harvard T.H. Chan School of Public Health, which calculated total intake of nutrients, including total caloric intake and grams of whole grains, fibre, and alcohol. Nutrient intake was calculated by multiplying the nutrient value in the specified portion size of each item on the FFQ by its frequency of intake and then summing across all items. Nutrient values were obtained from the US Department of Agriculture databases [20] supplemented with other sources.

The survey asked men if they had smoked 20 packs of cigarettes or more in their lifetime. If they responded “yes”, they were asked to report additional details regarding their smoking history. Men who responded “no” were considered never smokers.

Men completed a validated physical activity questionnaire which asked them to report their average weekly time spent doing nine types of aerobic and resistance training activities over the prior year [21]. Ten frequency options could be selected, ranging from 0 minutes to 11 or more hours per week. Participants were also asked about their regular walking pace and ability/frequency of climbing stairs.

Other information collected on the survey included height and weight [used to calculate body mass index (BMI; kg/m2)]; education level; a brief medical history, including family history of prostate cancer; and a detailed history of the use of vitamins and supplements.

Behavioural scores

Six a priori scores were evaluated, as described below. All scores were oriented such that increasing values reflected healthier behaviours. Please see Supplemental Tables S1 and S2 for additional details.

2021 score

The 2021 (post-diagnostic) Score was based on an extensive literature review conducted in 2021, summarising behaviours following a prostate cancer diagnosis associated with risk of recurrence, progression, and/or PCSM [6]. To determine the factors for inclusion in the score, we searched PubMed using the terms “prostate cancer” and “progression or mortality” in combination with terms describing individual lifestyle factors. Factors considered for the score included those that (1) exhibited a statistically significant association with metastases or PCSM in at least one study and (2) were corroborated by at least one additional study with an association in the same direction, whether or not statistically significant. In total, we identified seven such factors—smoking status [22,23,24,25,26,27,28,29,30,31]; BMI [32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48]; physical activity [49,50,51,52,53]; and intake of saturated fat [54,55,56], whole milk [57, 58], wine [59, 60], and processed meat [61, 62]. The three non-dietary factors demonstrated the strongest evidence in the literature review. We examined two versions of the 2021 Score, one without (“2021 Score”) and one with (“2021 Score + Diet”) the dietary components. The points per behaviour component ranged from 0 to 1 (see Supplement Table S1), with the points for the four dietary components (whole milk, alcohol, red and processed meat, saturated fat) averaged using the arithmetic mean to create a single dietary sub-score ranging from 0 to 1. This approach was consistent with the operationalization of the ACS recommendations into the ACS Score [13]. The point values were based on where the risk associated with prostate cancer outcomes appeared to change in the literature. The points for each component were summed to create the total 2021 Score (range: 0–3) and 2021 Score + Diet (range: 0–4) for each participant.

2015 score

Our team previously developed the 2015 (pre-diagnostic) Score to identify the risk of developing lethal prostate cancer among healthy men, based on the evidence available circa 2014 [10]. The six components—smoking status, BMI, physical activity, fatty fish intake, tomato intake, and processed red meat intake—were each scored as 0 or 1 based on cut-points associated with risk as reported in the literature at the time of score creation. The sub-scores were then summed to create the total 2015 Score (range: 0–6). Components of the 2015 Score were identified based on the existing literature available in 2015.

ACS score

To create a primary and an alternative ACS Score, we expanded on the operationalization of the ACS Nutrition and Physical Activity Guidelines for Cancer Survivors developed by McCullough et al. [13]. Each of the three components—BMI, physical activity, and dietary—were scored from 0 to 2 and then summed to create the primary ACS Score (range: 0–6). The dietary component included total servings and variety of fruits and vegetables, red and processed meat intake, and whole grain intake. We expanded to include strength training when assigning physical activity points, consistent with the guidelines. The “ACS Score + Alcohol” additionally included alcohol intake, scored from 0 to 2 (with the highest score for moderate alcohol intake: >0 to 2 servings/day), reflecting the inclusion of alcohol in the ACS recommendations for cancer prevention but not cancer survival (alternative score range: 0–8).

WCRF/AICR score

The WCRF/AICR Cancer Prevention Recommendations were operationalized based on published scoring guidelines [11, 12] and included BMI; physical activity; and intake of alcohol, sugar-sweetened beverages, fruits/vegetables, fibre, red and processed meat, and adapted ultra-processed foods (range: 0–7).

Outcome

The primary outcome was time to prostate cancer progression, defined as biochemical recurrence, secondary treatment, bone metastases, or PCSM, as applied previously [9, 53, 63]. Given the small number of PCSM events (n = 73) in this cohort, PCSM was evaluated as a secondary outcome.

Biochemical recurrence was defined as two consecutive prostate-specific antigen (PSA) readings ≥ 0.2 ng/mL following radical prostatectomy or a rise of 2.0 ng/mL above post-radiation nadir on two consecutive PSA readings; the date of recurrence was recorded as the date of the second elevated PSA. Secondary treatment was defined as any treatment started at least 6 months following primary treatment. Bone metastases included prostate cancer progression to bone, advancement to TNM stage M1b, a positive bone scan, and radiation to treat bone metastases. Cause of death was determined by the registry data coordinating centre and through confirmation by either the vital statistics official death certificate from the state in which the death occurred or by the National Centre for Health Statistics National Death Index [64]. Deaths were attributed to prostate cancer if the death certificate included ICD-9 code 185 [(metastatic) malignant neoplasm of prostate] as the primary or secondary cause of death.

Time to progression was measured from completion date of the CDL questionnaire to the date of progression (first event of biochemical recurrence, secondary treatment, bone metastases, or PCSM). For men with documented non-PCSM progression (i.e., recurrence, secondary treatment, or bone metastasis failure events), the censoring interval (i.e., window in which the event occurred) was bound by the last normal clinical visit (left limit) and the clinical visit documenting evidence of progression (right limit). For men who died from prostate cancer, the left and right limit were both date of death. Men without documented progression or PCSM were censored at their last date of follow-up or death (other cause); thus, the left limit of their censoring interval was defined by the last clinical follow-up date or date of death (non-PCSM), respectively, and the right limit was undefined (i.e., censored). Clinical follow-up was last consistently assessed across all CaPSURE sites on January 31, 2019; 26 men had a last known clinical follow-up date beyond this date and were administratively censored on that date.

Statistical analysis

Parametric survival models with a Weibull distribution were used to accommodate interval censoring associated with uncertainty in actual date of prostate cancer progression [65]. Because the date of death is known for PCSM (i.e., interval censoring was not an issue), we utilised Cox proportional hazards models rather than parametric survival methods when assessing the PCSM outcome. Proportional hazards assumptions were assessed visually by plotting the scaled Schoenfeld residuals against follow-up time.

We fit survival models using both continuous scores (per 1-unit change) and tertiles of scores. All models were clustered by CaPSURE clinical site with robust standard errors used to calculate confidence intervals (CI). Simple models were adjusted for time between diagnosis and participants’ first CDL questionnaire (continuous) and age at diagnosis. A directed acyclic graph (DAG) was developed to reflect our understanding of the complex relationship of interest; variables identified in the DAG were included as covariates in the fully adjusted models, as appropriate [66, 67]. Fully adjusted models were additionally adjusted for clinical T-stage (T1, T2, T3a), Gleason score (<7, 7, >7), diagnostic PSA level (≤6 ng/mL, >6 to 10 ng/mL, >10 ng/mL), primary treatment (radical prostatectomy, radiation, hormonal therapy, watchful waiting/active surveillance, other), family history of prostate cancer in a brother or father (yes, no), self-identified and physician-reported race (white, non-white), selenium supplement use (non-user; <140 µg/day; ≥140 µg/day; user with unknown daily dosage), total caloric intake (continuous, kcal/d), and the following variables if not part of the score of interest: whole milk intake (≤4 servings/week, >4 servings/week), wine intake (3–14 servings/week, <3 or ≥14 servings/week), alcohol intake (non-drinker, >0–2 servings/day, >2 servings/day), red and processed meat intake (quartiles), tomato intake (continuous, servings/day), dark fish intake (continuous, servings/day), and smoking (never, quit ≥10 years prior, quit <10 years prior, current). We further considered adjustment for comorbidities (diabetes, stroke, prior myocardial infarction, or other heart disease; yes/no) but the magnitudes of the estimates changed very little with adjustment, so these variables were not included in the final models.

We assessed potential interaction between each of the scores and age at diagnosis (<65 years, ≥65 years) and, separately, stage at diagnosis (T1, T2–T3a) by adding interaction terms with the scores in the models and using Wald tests. Given statistically non-significant Wald tests and small magnitudes of estimated interaction regression coefficients, interaction terms were not included in the final models. We examined goodness-of-fit of the survival models using Cox–Snell residual plots. Across all scores, goodness-of-fit was best in the fully adjusted models, with decreasing fit in the tails. Fully adjusted models for progression were also run using exponential distributions, which produced Cox–Snell residual plots that demonstrated poorer fit than Weibull models and thus were not reported.

Sensitivity analyses

First, we were concerned about confounding due to PSA surveillance after diagnosis (i.e., men with healthier behaviours may be more likely to be monitored via PSA tests, potentially creating a positive correlation between healthy lifestyle habits and risk of progression). To address this, Poisson regression was utilised to compare the number of PSA visits to tertile of each of the six scores, with the lowest tertile (i.e., the least healthy group) as the reference. Total follow-up time was used as an offset in these models.

Second, whereas our primary analyses used time of the CDL questionnaire completion as time zero—which necessitates excluding men who experienced an event prior to the survey—sensitivity analyses re-assigned time zero as time of diagnosis. These analyses assumed that the responses on the CDL questionnaire were consistent with what would have been measured at the date of diagnosis. Men excluded from our primary analyses due to documented progression prior to CDL questionnaire were included in these sensitivity analyses, resulting in an analytic sample of 2447 men. For this approach, we first assessed whether there was an interaction between year of diagnosis and each of the behavioural scores by adding an interaction term with the scores in the models and using Wald tests; no evidence of interaction was found.

Third, we were interested in understanding how competing events (i.e., deaths due to causes other than prostate cancer) impacted our primary results. Methods to address competing events in the presence of interval censoring are not readily available or accessible. Thus, we ran Cox proportional hazards models for progression and compared these results to Fine-Gray analyses accounting for other deaths as a competing risk. Proportional hazards assumptions were assessed visually by plotting the scaled Schoenfeld residuals against follow-up time.

Lastly, missingness in the covariates resulted in a loss of events in our fully adjusted models. Specifically, men with missing data for any of the score components were excluded from the primary analysis for that score: n = 60 for 2021 Score, n = 60 for the 2021 Score with Diet, n = 83 for 2015 Score, n = 40 for ACS Score, n = 70 for the ACS Score with Alcohol, and n = 43 for WCRF/AICR Score. To understand the impact of this missingness on our primary results, we performed sensitivity analyses utilising multiple imputation to handle missing data [68], which assumes that data are missing at random. We assessed the plausibility of this assumption by summarising participant characteristics by missingness status for each of the six scores. We performed multiple imputation via chained equations using the chained command in Stata to first generate 25 imputed datasets. We then fit survival models across all 25 imputed datasets and pooled the results using Rubin’s Rules [69]. Our imputed models included fully observed variables (CaPSURE clinical site, age at diagnosis, BMI, days of follow-up, total energy intake, tomato intake, days from CDL return to the left interval of follow-up time, race, diagnostic T-stage, and family history of prostate cancer) and variables with incomplete values (diagnostic PSA and Gleason score; total alcohol, whole milk, dark fish, total wine, and red and processed meat intake; each of the scores; smoking status; and primary treatment).

All statistical analyses were performed using Stata version 17 (StataCorp, College Station, TX) using a two-sided alpha level of 0.05 to assess statistical significance.

留言 (0)

沒有登入
gif