Short Physical Performance Battery and all-cause mortality: systematic review and meta-analysis

Appropriate articles were found using the Medical Subject Headings (MeSH) strategy and searching in MEDLINE, the Cochrane Library, Google Scholar, and BioMed Central. The search strategy was created by RP. The terms searched were: ((short physical performance battery) OR (SPPB) OR (lower limb strength) OR (standing balance) OR (walking speed) OR (chair stand)) AND ((mortality) OR (death)).

Only articles published in the English language and in peer-reviewed journals were selected. The research was carried out between July 2015 and January 2016. Independent reviewers (RP, GC) analyzed the titles and abstracts of the articles and determined which of them warranted the examination of the full text. Studies included in the analysis had to have the following characteristics: (1) observational (non-randomized) study; (2) inclusion of more than 50 subjects; (3) reporting the stratification of patients/population according to SPPB cut-points; (4) presenting data on all-cause mortality in relation to the value of SPPB expressed as hazard ratio (HR) or odds ratio (OR). Duplicate, interventional, or animal studies were excluded. Both reviewers agreed to the final number of studies included in the present analysis.

Independent reviewers (GC, RP, and SV) completed the database, which contained information about the journal, year of publication, authors, baseline characteristics of study population, follow-up length, SPPB cut-points, and source of mortality data. The primary endpoint was all-cause mortality. Additional analyses were performed after stratification of studies according to the following criteria: (1) mean age of the study population (≤75 years versus >75 years); (2) setting (general population versus outpatients versus hospitalized patients); (3) geographical region (North America versus Europe versus Asia); (4) follow-up length (≤1 year versus >1 year and ≤5 years versus >5 years). To obtain standardized data, the authors of all the selected papers (n = 24) were contacted. Of the 22 authors contacted (two were corresponding authors for two studies), one was not able to provide the requested data, one refused to participate, and five never replied to the inquiry. A total of 15 authors (68%) gave complete available data for 17 of the studies originally selected (71%) (see Fig. 1). Authors were asked to complete a table summarizing baseline characteristics of their studies (mean age, sex, hypertension, cardiovascular disease, cerebrovascular disease, diabetes) and to stratify the population into four SPPB score categories (0–3, 4–6, 7–9, 10–12) according to the cut-points provided by Guralnik and colleagues in their original work [4]. The reference group for the analyses comprised participants ranging between 10 and 12 on the SPPB score. In addition, authors were asked to calculate the odds ratio (OR)/hazard ratio (HR) for all-cause mortality in SPPB groups with values 0–3, 4–6, and 7–9 compared to the group 10–12 as reference, and to perform multivariate analyses adjusted for age, sex, and body mass index (weight/height²).

Two unblinded reviewers (RP and SV) independently evaluated the quality of the included studies using prespecified electronic forms (piloted over the first three cases) and a modified version of the Newcastle-Ottawa Scale (NOS) for cohort studies [34] (Additional file 1: eTable 1). Because of the design of the studies considered, we did not consider the section for “Comparability” and question 2 in the section “Selection” (“selection of the non exposed cohort”). Discrepancies between reviewers were solved by consensus. No study was excluded on the basis of this analysis. The same reviewers independently analyzed references of all the evaluated articles to avoid the eventual exclusion of additional studies.

Continuous variables were reported as mean (± standard deviation) or median (interquartile range). Categorical variables were expressed as number and percentage (%). Point estimates and standard errors were extracted from individual studies and combined by the generic inverse variance method [35], computing risk estimates with 95% confidence intervals according to logarithmic transformation of the hazard measures. Considering the high likelihood of between-study variance, a random effect model was used. Statistical heterogeneity was assessed using Cochran’s Q test. This statistic was complemented with the I² statistic, which quantifies the proportion of total variation across studies that is due to heterogeneity rather than chance. A value for I² of 0–25% represents insignificant heterogeneity, 26–50% low heterogeneity, 51–75% moderate heterogeneity, and >75% high heterogeneity [36]. The chi-square test was used to test differences between subgroups. To estimate the percentage of deaths that could be attributed to poor physical function, the percentage attributable risk (%AR) was calculated [37]. Finally, a random effect meta-regression analysis was performed to assess the effect of some potential confounding factors (age, sex, previous history of cardiovascular disease, previous history of cerebrovascular disease, diabetes, hypertension) on the results. Publication bias was appraised by graphical evaluation of funnel plots and through Begg and Mazumdar rank correlation, Egger’s regression intercept, and Duval and Tweedie trim and fill [36]. Statistical analyses were conducted using ProMeta software (Internovi, Cesena, Italy) and RevMan 5 (the Cochrane Collaboration, the Nordic Cochrane Centre, Copenhagen, Denmark).

After removal of duplicates, 725 titles were identified by the databases search (Fig. 1). Overall, 529 items were excluded after the first evaluation of the title and abstract, as they failed to meet the prespecified inclusion and exclusion criteria. Of the remaining 196 records examined, 134 were excluded because they focused on other outcomes or on other physical performance measures. An additional 5 were not retained because they were not original papers but reviews, and 31 because they were study protocols. Twenty-six studies were examined as full papers. Two of these were excluded because they were based on the same study sample used in Lai et al. [15]. The corresponding authors of the retained 24 records were contacted [6‐29]. As previously explained, standardized information was obtained for 17 of them [5‐21], and these studies were included in the final qualitative and quantitative analysis (Fig. 1).

A total of 16,534 participants (mean age: 76 ± 3 years, women: 78%) were included in the meta-analysis. Overall, 47% of participants had hypertension, 9% diabetes, and 31% prior diagnosis of cardiovascular disease, while only 5% had a previous diagnosis of cerebrovascular disease (Table 1). Eleven studies were focused on the general population [7, 8, 12‐14, 16‐21]. Of note, the study of Greene et al. focused on current and former injection drug users [13]. Three studies enrolled hospital inpatients [10, 11, 22]. In the study by Volpato et al., the causes of hospitalization were congestive heart failure (HF), chronic obstructive pulmonary disease, pneumonia, or minor stroke [22]. In the study by Chiarantini et al., the cause of hospitalization was decompensated HF [10], whereas in the study by Corsonello et al., patients were admitted to an acute care medical ward or a long-term care/rehabilitation unit [11]. The remaining three studies were based on outpatients presenting heterogeneous clinical conditions (primary care center: cancer, cirrhosis, on waiting list for liver transplant) (Table 1) [6, 9, 15]. Eleven studies were performed in Europe [6, 8‐11, 16‐20, 22], five in America [7, 12, 13, 15, 21], and one in Asia [14] (Table 1). The SPPB score was 0–3 in 9% of the study population, 4–6 in 16%, 7–9 in 33%, and 10–12 in 42%, respectively. Quality assessment of the 17 studies according to the NOS demonstrated an average high quality level with all studies having a score ranging between 5 and 6 (Additional file 1: eTable 1).

Overall, 3237 deaths (19.5%) were recorded. The length of follow-up ranged from a minimum of 1 year to a maximum of 11 years (median 4.5 years) (Additional file 1: eTable 2). As compared to an SPPB score of 10–12, scores of 7–9, 4–6, and 0–3 were associated with greater and progressively increasing risks of all-cause mortality (Fig. 2). After adjustment for age, sex, and body mass index, we found an OR of 1.50 (95%CI 1.32–1.71) for SPPB scores 7–9, OR 2.14 (95%CI 1.92–2.39) for SPPB scores 4–6, and OR 3.25 (95%CI, 2.86–3.79) for SPPB scores 0–3 (Fig. 2). The %AR (95% CI) of death attributed to poor functional performance was 33% (24–42%), 53% (48–58%), and 69% (65–73%), respectively. Heterogeneity, expressed as I²%, was insignificant for the comparison between SPPB scores 0–3 and 4–6 versus 10–12, whereas it was low between SPPB scores 7–9 versus 10–12.

Subgroup analyses demonstrated that after stratification of the studies for age, type of population, geographic area, and follow-up length, the association between SPPB and all-cause mortality remained highly consistent, with no statistical significance of the interaction terms (Table 2). Random effects meta-regression disclosed no significant association between confounding factors (previous cardiovascular disease, cerebrovascular disease, diabetes, hypertension, age, and sex) for SPPB scores 0–3 or 4–6 versus 10–12 and the risk of all-cause mortality (Additional file 1: eTable 3). In contrast, the OR for all-cause mortality with SPPB scores 7–9 was higher in younger populations, diabetics, and men (Additional file 1: eTable 3, eFigure 2).

According to graphical evaluation of funnel plots, Begg and Mazumdar rank correlation, and Egger’s regression intercept, there was no evidence of publication bias (Additional file 1: eTable 4 and eFigure 1A-C).

Our results suffer from those limitations that are inherent to all meta-analytic techniques including particularly heterogeneity in populations and variable endpoint definitions across studies. We could analyze data only from authors who replied to our request and, even if statistical analyses do not show the presence of publication bias, this could not be completely excluded. Secondly, we decided to report SPPB score in classes (0–3, 4–6, 7–9, 10–12) and not as a continuous variable. Finally, we only evaluated the association between SPPB and mortality. Additional studies are needed to show that adoption of SPPB into a prediction model improves discrimination of mortality and to evaluate its clinical utility in the practice setting. Nevertheless, this is a meta-analysis on a large sample, including more than 16,000 patients. Our protocol has been prespecified and registered on a public platform (PROSPERO), and the collaboration between authors allowed us to obtain highly standardized data.