The modifying role of physical activity in the cross-sectional and longitudinal association of health-related quality of life with physiological functioning-based latent classes and metabolic syndrome

SAPALDIA is a population-based cohort with associated biobank initiated in 1991. In SAPALDIA1, 9′651 adults (18–62 years) were randomly recruited from eight study areas in Switzerland representing the country’s geographic and cultural diversity [22]. In the subsequent decades, two follow-ups were carried out including 8′047 subjects in SAPALDIA2 (2001/2002) and 6′088 in SAPALDIA3 (2010/2011) [23]. All three assessments comprised questionnaires (based on validated instruments as for example IPAQ for physical activity) and health examinations of increasing complexity over time.

The present analysis ultimately involved 1′124 of the SAPALDIA4 55+ participants who provided complete information on all relevant variables. For 902 of these subjects complete information was also available from their SAPALDIA3 assessment for longitudinal analyses (Fig. 1). To assess bias due to loss-of-follow-up, the baseline characteristics at SAPALDIA1 of participants who reached the age of 55+ at the time of the health assessments, stratified by participation status at SAPALDIA4 55+ examination, can be seen in Additional Table 1.

The SAPALDIA cohort study complies with the Declaration of Helsinki. For each survey, ethics approval was granted by the regional ethics committees and participants provided written informed consent prior to participation.

For SAPALDIA4, cohort participants were invited to answer multiple self-administered questionnaires on paper or online. The questionnaires consisted of information on socio-demographic characteristics, lifestyle, psychosocial factors, disease symptoms, diagnoses, medications, as well as health and social service use. The SAPALDIA4 55+ assessment additionally included a questionnaire specifically addressing aging-related determinants and risk factors.

In contrast to the self-administration of SAPALDIA4 questionnaires, SAPALDIA3 questionnaires were applied in the context of in-person interviews. Relevant questionnaire information for this analysis were questions on physical activity levels (frequency of moderate and vigorous physical activity according to WHO guidelines) as well as the covariates: sex, age, and education level (Low = Primary School (≤ 9 years), Middle = Secondary school, middle school or apprenticeship (≤12 years), High = Technical College or University (≥12 years)). We chose for this population, education levels as a proxy for socio-economic status, as this measure has proven in previous studies to well represent the socio-economic status [24].

We further used information on health service utilization as secondary endpoint – next to HRQoL -, which was also assessed in the SAPALDIA4 questionnaire. We defined three health care utilization variables: Visit to physician(s); to hospital(s); combination of visits to physician(s) and hospital(s). The information were reported for the 12 months preceding the survey (yes vs. no for each). We treated these response variables as binary because nearly all participants reported 0 or 1 visit in the specified period.

Furthermore, we used information on doctor’s diagnosis reported in the SAPALDIA4 questionnaires for sensitivity analysis. Specifically, we defined a variable capturing the following cardiovascular diseases: Myocardial Infarction, Angina Pectoris, Heart Insufficiency, Claudicatio Intermittens and Stroke.

Questionnaires at SAPALDIA3 and SAPALDIA4 also included the 36-Item Short-Form Health Survey (SF-36), a widely used HRQoL assessment tool that was validated in large population-based surveys as well as in clinical settings [25, 26]. The construct as well as the criteria validity and reliability were thoroughly tested in the development of this tool and are therefore given [27, 28].

The questionnaire is designed to provide a summary of physical and mental health scores, based on eight domains. The physical component comprises physical functioning (PF), bodily pain (BP), role-physical (RP) and general health perception (GH). The mental component reflects the vitality (VT), social role functioning (SF), role emotional (RE) and mental health perception (MH). Scores for each subscale range from 0 to 100, and higher scores indicate better HRQoL [29]. Three domains of the SF-36 (social role functioning, role-physical & role emotional) showed only very few distinct values in our sample and the proportion of subjects at SAPALDIA4 with the perfect score 100 ranged from 72 to 85%. As a consequence, we did not consider these variables. As these variables represent minor sub-scales of the SF-36, their exclusion still allowed us to assess MCS and PCS as HRQoL outcomes.

The health examination consisted of measurements of weight (SECA877 flat scale; SAPALDIA3&4), height (SECA206, wall-fixed measuring system; SAPALDIA3&4), hip and waist circumference (SECA201 ergonomic measuring tape, SECA, Reinach, Switzerland; SAPALDIA3&4), bio-impedance analysis (BIA) using two different devices (Helios, Forana GmbH, Frankfurt, Germany (SAPALDIA3&4 Tanita MC-780MA, TANITACorporation, Tokyo, Japan; SAPALDIA4), blood pressure and heart rate measurements (Omron MC6 or Omron 705-IT, Anandic Medical Systems AG, Bern Switzerland; SAPALDIA3&4); non-fasting blood glucose (only measured at SAPALDIA3); diagnostics for HbA1c and triglycerides (SAPALDIA4: point-of-care diagnostics from capillary blood: Afinion AS100 Analyzer; ALERE, Wädenswil, Switzerland; SAPALDIA3: analysis of venous blood) [30, 31].

Blood pressure measurements at SAPALDIA3&4 were taken after the participant was seated for at least 10 min. Two measurements were taken, with a break of 3 min between measurements. The blood measurements for glycemia and triglycerides were taken in a non-fasting state. From the anthropometric measurements, body mass index (BMI) (SAPALDIA3&4) and waist to hip ratio (SAPALDIA3 & 4) were derived.

Five categorical variables at SAPALDIA3&4, which reflect cardio-metabolic physiological functioning were considered for identifying latent classes. At both time points (SAPALDIA3 & 4) variables with evidence-based thresholds or recommendations were categorized accordingly, namely systolic blood pressure [32] (diastolic blood pressure was omitted due to its high correlation with systolic blood pressure), BMI [33], HbA1c [34] and triglycerides [35]. For percentage body fat, sex-specific tertiles were calculated in the absence of a reference for categorization (Table 1).

The metabolic syndrome at SAPALDIA3&4 was defined based on the Joint Interim Statement (JIS) [36]. The JIS defines the metabolic syndrome if any three of the following values are present: Blood glucose ≥5.6 mmol/L (medication or diagnosis for elevated blood glucose as alternate indicator); waist circumference ≥ 94 cm in men & ≥80 cm in women; systolic blood pressure ≥ 130 or diastolic blood pressure ≥ 85 (antihypertensive medication or diagnosis as alternate indicator); hdl < 1.0 mmol/L in men & < 1.3 mmol/L in women; triglycerides ≥1.7 mmol/L (medication or diagnosis for dyslipidemia as alternate indicator). At SAPALDIA4 we used Hba1C instead of blood glucose (Hba1C ≥5.7%, medication or diagnosis for elevated blood glucose as alternate indicator), as blood glucose was not measured in this follow-up.

In a first step [1] LCA was carried out to empirically classify cardio-metabolic physiological functioning variables, separately for SAPALDIA3 and SAPALDIA4. Subjects were characterized based on their values of the five predictor variables. LCA was used as an explorative tool (unconstrained LCA) without a priori expectation about the number of classes.

In order to detect the appropriate number of classes and maximize model fit, we started with a one-class model and increased the number of latent classes up to six. The final model was selected by examining the Bayesian information criterion (BIC) and the Akaike Information Criterion (AIC) in the first place. These indices have shown to be useful for determining the appropriate number of classes for LCA [37]. We checked whether there was a good discrimination between the final definition of latent classes and the five predictor variables. Two additional model fit indices, the adjusted BIC and the consistent AIC, were considered along with the proportions of the single classes to further support the decision on the final model. We assessed cross-sectional associations of the derived latent classes at SAPALDIA4 with socio-demographic characteristics of the study population by using multinomial logistic regression models with latent class membership as outcome variable and sex, age, educational level and study area as simultaneous predictor variables.

To assess cross-sectional and longitudinal associations of the latent classes with HRQoL domains and to compare them to according MetS-HRQoL associations, we performed quantile regression models with adjustment for individual predictor variables (e.g. sex, age and educational level) as well as for HRQoL at baseline in the case of longitudinal associations. Quantile regression models were chosen to address the problem of left-skewed distribution of HRQoL measures.

In order to assess confounding and effect modification by physical activity in the above associations, we first ran the models with and without adjustment for physical activity and second created composite variables classifying participants based on a combination of latent classes or MetS with physical activity status.

Finally, we used logistic regression models to assess cross-sectional associations of the composite variables with health service utilization.

For sensitivity analyses we reran analyses by excluding subjects with self-reported diagnoses of a cardiovascular disease at SAPALDIA4 (Myocardial Infarction, Angina Pectoris, Heart Insufficiency, Claudicatio Intermittens, Stroke).

We performed all analyses using Stata 15 (Stata Corporation, College Station, Texas).

The descriptive characteristics of the study population can be seen in Table 2. Women and men were equally distributed among the subjects at both time points (SAPALDIA3&4). The mean age at SAPALDIA3 was 60.0 ± 7.8 (range: 47–80 years) and at SAPALDIA4 67.4 ± 7.9 (range: 55–88 years). Nearly two-third of the subjects showed medium education levels, which refers to having completed middle school. The mean score of the overall HRQoL domain (GH) showed descriptive difference by sex and with advancing age. With declining education the score decreased. The HRQoL GH scores mostly decreased in a dose-dependent manner with increasingly poorer physiological functioning variables (blood pressure, glycemia, body fat, BMI and triglycerides). Physical activity displayed higher HRQoL GH scores for the sufficiently active subjects. Physical activity and physiological parameters generally decreased from SAPALDIA3 to SAPALDIA4, and therefore with increase in age, while the prevalence of the metabolic syndrome increased.

At both time points (SAPALDIA3 & 4) the main values of the BIC and AIC reached its minimum in the model consisting of three classes. The adjusted BIC as well as the CAIC supported these results (Additional Tables 2 & 3). By considering these goodness-of-fit indices the model with three latent classes was chosen to be the final model. The model with three classes also generated a relatively even distribution at SAPALDIA4 (N₁ = 340 N₂ = 455, N₃ = 329) within the study sample, which increases statistical power, while the SAPALDIA3 sample showed fewer subjects in the third class (N₁ = 412, N₂ = 352, N₃ = 137).

The proportion of the different classes and the estimated class-specific probabilities for the predictor variables are shown in Additional Table 4 (SAPALDIA4) and Additional Table 5 (SAPALDIA3). The by-latent-class distribution of variables included in deriving the classes was comparable between SAPALDIA3 and SAPALDIA4.

Class 1, labeled “Healthy” consisted of individuals having mainly normal physiological functioning measured such as blood pressure, glycemia, triglycerides and percentage body fat or BMI. This class included 30% of participants at SAPALDIA4 and 46% of participants at SAPALDIA3.

In Class 2, labeled “At risk”, most individuals were overweight and had intermediate percentages of body fat. The prevalence of high blood pressure, hyperglycemia, and dyslipidemia in this group were intermediate compared to those observed for classes 1 and 3. Most subjects at SAPALDIA4 were categorized into this class (42%), while at SAPALDIA3 this class consisted of 29% of the participants.

Class 3, labeled “Unhealthy” had an overrepresentation of individuals that were obese, in contrast to the other classes that contained no obese participants. Accordingly, participants in class 3 were more likely to have a high percentage of body fat, to be hypertensive, to have high glycemia and triglyceride values. In the SAPALDIA4 sample 29% was assigned to this class, while in the SAPADALDIA3 sample only 15%.

In regards to the socio-demographic distribution of the latent classes at SAPALDIA4 (Table 3) we observed that females were less likely to belong to the Unhealthy or the At risk class. The likelihood of being categorized as “Healthy” decreased with age and lower educational level. Persons aged 75+ or with low educational level were particularly likely to be categorized as Unhealthy. The descriptive data of sex, age and education between the different classes can be seen in Additional Table 6.

The results of the quantile regression analyses (Table 4A) showed that, compared with the reference class (Healthy) median HRQoL scores were lower in the Unhealthy class, i.e., with differences of − 3.06 (95%CI: (− 5.51, − 2.14) in the GH domain, of − 5.64 (− 6.98, − 4.30) in the PF domain and of − 9.14 (− 13.91, − 4.38) in the BP domain. The At risk class only had borderline significant scores in the PF domain (− 1.01 (− 2.05; − 0.03)), compared to the reference class. The association results of the latent classes with HRQoL did not substantially differ when adjusting for physical activity in the model, although they tended to become somewhat weaker. We observed statistically significant positive associations for being physically active (i.e. meeting WHO guidelines) with all HRQoL domains. Coefficients ranged from 3.92 to 7.50 for the comparison of meeting vs. not meeting WHO physical activity guidelines. The median scores and descriptive differences of the latent classes (without adjustment of physical activity) can be found in Additional Fig. 1.

Participants with MetS showed associations in the same direction with HRQoL domains as the Unhealthy group (Table 4B). Neither MetS nor the class “Unhealthy” were associated with the MH domain. The associations did not differ substantially after adjustment for physical activity. As in the model with the latent classes, physically active subjects had higher HRQoL scores in all domains.

Figure 2 shows the results of the quantile regression models of for the composite variables cross-categorizing participants based on their physical activity and their latent class or MetS, respectively, and their association with HRQoL domains. We observed (Fig. 2A) that compared to the reference group (Healthy & active) all categories had coefficients in the direction of lower HRQoL scores, irrespective of the domain considered. Moreover, within latent class categories the inactive groups consistently exhibited lower scores on average than the active one. The lowest HRQoL scores were observed in the unhealthy and physically inactive group compared to the healthy and physically active group for the GH, PF, and BP domain. This group even exhibited statistically significantly lower scores in the MH domain compared to the same reference group.

The association patterns for the composite MetS/physical activity variables with HRQoL (Fig. 2B) were consistent with the association patterns for the composite variable considering latent classes instead of MetS.

The exact coefficients for Fig. 2A and B can be seen in Additional Table 7.

The longitudinal results are displayed in Table 5. We observed statistically significant differences compared to the Healthy & active group of − 4.95 (95% CI: − 7.01; − 2.90) points in effect score for the association of the Unhealthy & inactive group with GH. Furthermore, the Unhealthy & active group scored − 2.47 (− 4.32; − 0.81) points in effect score less in the PF domain compared to the reference group. The At risk & inactive group showed statistically significantly lower scores for the VT domain of − 2.29 (− 4.34; − 0.23) points in effect scores. Similar to Table 4 we detected generally lower HRQoL scores in inactive participants in the Unhealthy group, but the same pattern was not as consistently observed in the Healthy and At risk groups.

In Table 5B we detected no differences in HRQoL scores by physical activity status among the participants without MetS. In contrast, among persons with MetS, HRQoL scores were consistently lower for inactive participants. The difference reached statistical significance for GH, and PF. In the MetS group the sufficiently active and inactive, respectively, had − 0.43 (95%CI: − 2.38; 1.53) and − 4.83 (− 7.89; − 1.78) lower scores for GH as well as − 1.79 (− 2.80; − 0.79) and − 2.48 (− 4.46; − 0.50) lower scores for PF, when compared to active subjects without MetS. The Mets & active group showed statistically significant differences, compared to the reference group (NoMetS & active) in the MH domain of 1.92 (0.31: 3.53).

Table 6A displays the association of composite variables with self-reported health service use in the year before SAPALDIA4. While no associations were observed with physician’s visits, participants in the Unhealthy & inactive group (OR 2.69 (1.52; 4.74)) and At risk & inactive (2.09 (1.14; 3.81)) group exhibited the highest risk for a hospitalization in the previous 12 months compared to participants in the reference group (Healthy & active).

The risk of hospitalization was also strongest for the group MetS & inactive (2.63 (1.64; 4.22)) and No MetS & inactive group (2.26 (1.38; 3.72)), compared to the reference group (Table 6B). Same as for the latent classes, the associations with physician’s visit did not show a pattern of a higher likelihood among inactive persons without MetS.

In the sub-sample of subjects not reporting any cardiovascular diseases we found cross-sectional and longitudinal associations with HRQoL and cross-sectional associations with health service utilization comparable to those in the overall sample for both, composite variables based on latent classes and on MetS, respectively (Additional Table 8 and 9 for HRQoL; Additional Table 10 for health service utilization).

The information on socio-economic status and physical activity were self-reported by the participants, which could lead to measurement errors. The fact that participants at SAPALDIA3 were interviewed, whereas at SAPALDIA4 participants self-administered questionnaires, could have introduced differential misclassification of physical activity. But it is unlikely that this would have altered the main messages of our study, given the consistency of longitudinal and cross-sectional results. All other variables included in the analysis were objectively assessed. Triglyceride and glycemia tests and the type of blood as test source (venous versus capillary blood) differed between SAPALDIA3 and SAPALDIA4. While this could again be the source of differential misclassification over time, the consistency of the cross-sectional and longitudinal results suggests this bias to be minimal. The variables reflecting cardio-metabolic physiological functioning were chosen based on the evidence reported in the literature. Our current LCA may furthermore be sensitive to the chosen categories for the predictor variables. Yet, the consistency in results between latent classes and MetS points to the fact that the chosen cutoffs for categorization are clinically relevant. As in any cohort, we cannot exclude bias due to loss-to-follow-up. Several of the factors leading to or included in the categorization of participant’s physiological functioning influenced participation at follow-up. Persons with healthy habits were more likely to participate in SAPALDIA4. Older persons were less likely to participate at follow-up, in part because they had died. It is therefore likely that the effect of poor health habits including obesity and physiological functioning on HRQoL was underestimated. Given this loss-to-follow-up, the SAPALDIA4 participants are most likely no longer representative of the current general population in Switzerland. Yet, with regard to physical activity the observed distribution in our study sample is highly comparable to the distribution recently observed in the population-based Swiss Health Survey. The generalizability of our findings to other study settings needs testing, particularly in the light of the very high physical activity level of the Swiss population in general [43].

Despite these limitations, the findings of this study strongly support that it is possible to reduce the number of physiological functioning variables using LCA. The latent classes corresponded well with the cluster of the MetS, which underlines the significant interplay of these variables.

The LCA was facilitated by access to data from the SAPALDIA participants, who are deeply characterized on various physiological domains. The population-based design of the study is an important prerequisite for the generalizability of the findings, at least within the Swiss setting, although participation and survivor bias always pose a threat to validity and generalizability of any long-running cohort. Finally, the longitudinal design of this study adds knowledge to possible pathways in the investigated associations.