Physical activity in the treatment-resistant depression and non-remitted depression: a systematic review of randomized controlled trials

Identification, selection, and eligibility criteria of studies

For the assessment and selection of the scientific articles, the following PICOS (Participant, Intervention, Comparator, Outcome, Study design) question was utilized:

Two authors (S.M-M and J.E-O) were responsible for retrieving selected articles having in mind the following inclusion criteria: (a) Treatment-resistant or non-remitted depression, (b) physical activity or exercise, (c) clinical trials or experimental trials, (d) studies published in English, (e) scientific studies but not such as a book, magazines, online websites, reports, guidelines or recommendations or thesis. Scientific studies were excluded if they did not meet the eligibility criteria for the PICOS questions. The following exclusion criteria were: (a) suffered depression and responded to treatment, (b) included PA interventions combined with other strategies, (c) the intervention was not an adaptation of PA (d) had less than 18 of age, (e) there was not a comparison between pre-and post-intervention, and (f ) it was not an RCT. These comparator categories were predefined before study selection to reduce conceptual heterogeneity. This approach has been previously adopted in systematic reviews of exercise interventions, where participant and therapist blinding is inherently unfeasible.

Data sources and search strategy

A systematic literature search was conducted in June 2024 using the following online databases: PubMed, Cochrane Central Register of Controlled Trials, Scopus, SPORTDiscus, and Web of Science. The search was developed by synonyms of treatment resistant and physical activity or exercise and was performed by the followings search strategies: In PubMed (“major depression” OR “depression disorder” OR “depressive disorder” OR “treatment resistant depression” OR “non-remitted major depressive” OR “poorly responsive depressive” OR “treatment-resistant patients” OR “treatment-resistant major depressive”) AND (“physical activity” OR exercise); in Cochrane, Web of Science and SPORTDiscus “major depression” or “depression disorder” or “depressive disorder” or “treatment resistant depression” or “nonremitted major depressive” or “poorly responsive depressive” or “treatment-resistant patients” or “treatment-resistant major depressive”) and (“physical activity” or exercise); and in Scopus “major depression” OR “depression disorder” OR “depressive disorder” OR “treatment resistant depression” OR “nonremitted major depressive” OR “poorly responsive depressive” OR “treatment-resistant patients” OR “treatment-resistant major depressive” AND ““physical AND activity” OR “exercise””.

Data screening and extraction

After duplicates were removed, the articles were revised by two independent authors (S.M-M and J.E-O) based on the titles and abstracts generated by the search strategy. After this procedure, any article that remained questionable was analyzed more thoroughly by reading the full text. The chosen articles were read and assessed using eligibility criteria. In case of discrepancies, the authors would reach a consensus. The following articles were revised using Rayyan Intelligent Systematic Review Software, which was developed specifically to expedite the initial screening of abstracts and titles using a semi-automated process. Likewise, in the publications, the details of the investigation appear to show that screening was conducted according to eligibility criteria.

A data extraction form was designed and piloted before its use. Information was extracted on the method (i.e., design, participants, intervention, outcome measures) and on the results (i.e., sample size, incidence rates, and number of events). When information was not available in the published trials, the corresponding author was requested to provide details. The data extraction was carried out by S.M-M and J.E-O (Data extraction form-supplementary data).

Outcome parameters

The more significant outcome for this analysis was the improvement of depression by PA or exercise in TRD and NRD.

Quality appraisal

For assessing included scientific articles, the Physiotherapy Evidence Database (PEDro) scale (de Morton, 2009) and Oxford’s Evidence Levels were used (Howick et al., 2009). The PEDro scale rates RCT on a scale from 0 (low quality) to 11 (high quality) related to scientific rigor (de Morton, 2009). Each measure is rated “yes” or “no”. Items related to blinding of participants and assessors were considered in the context of exercise-based interventions, where blinding is often not feasible. Therefore, these items were not used to penalize study quality scores, and results were interpreted accordingly. This approach is consistent with previous systematic reviews of exercise-related research. For this reason, the maximum score on the modified PEDro 8-point scale was 7 (the highest), since the first item is not included in the total score. The qualitative ratings were adjusted to those used in previous exercise-related systematic reviews as follows: 6–7 = “excellent”; 5 = “good”; 4 = “moderate”; and 0–3 = “poor”. Oxford’s Evidence levels range from 1 to 5, with 1a being systematic reviews of high-quality RCT, 1b individual RCT with a narrow confidence interval, 2a systematic review of cohort studies, 2b individual cohort study, 3a systematic review of a case-control study, 3b individual case-control study, 4 case-series, and 5 being expert opinions (Howick et al., 2009). Two researchers (S.M-M and J.E-O) inspected the methodology of each study independently. Disagreements regarding the appraisal of methodological quality were resolved through discussion between the reviewers.

In addition to the aforementioned tools, the Cochrane Risk of Bias 2 (RoB 2) tool was used to assess the methodological quality of parallel-group randomized trials (Higgins et al., 2011). Five domains were evaluated: bias arising from the randomization process (Domain 1), bias due to deviations from intended interventions (Domain 2), bias due to missing outcome data (Domain 3), bias in measurement of the outcome (Domain 4), and bias in the selection of the reported result (Domain 5). For each domain, a judgment of ‘low risk’, ‘some concerns’, or ‘high risk’ was assigned. The overall risk of bias for each study was determined following Cochrane guidelines, with each study rated as ‘high risk’ if at least one domain was rated as ‘high risk’. Two reviewers (S.M-M and J.E-O) independently assessed the included studies, and discrepancies were resolved by consensus. For substudies derived from the TREAD trial, the published protocol was consulted to verify the pre-specification of outcomes (Trivedi et al., 2006).

Grading the certainty of evidence (GRADE)

The certainty of evidence for the primary outcome (improvement in depressive symptoms) was assessed using the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) framework (Guyatt et al., 2011) The GRADE approach rates the certainty of evidence across four levels: high, moderate, low, and very low. Certainty was downgraded based on five domains: 1) Risk of bias: limitations in study design or conduct (e.g., lack of blinding, inadequate allocation concealment); 2) Inconsistency: unexplained heterogeneity across studies; 3) Indirectness: differences between the populations, interventions, or outcomes of interest and those studied; 4) Imprecision: wide confidence intervals or small total sample size; 5) Publication bias: likelihood that negative or null results remain unpublished.

Heterogeneity evaluation

Heterogeneity across studies was assessed in a structured manner across three dimensions: clinical (population, intervention, comparator, outcome), methodological (study design), and statistical (effect measure, effect size, and Heterogeneity index [I²] when estimable). The results of this assessment are presented in Table 4 and were used to determine the feasibility of a meta-analysis and to justify the narrative synthesis of the findings.

Selection of studies

The database searches identified 10777 references (see Extended data, (Etxaniz-Oses et al., 2023)). 3189 from PubMed, 1219 in Cochrane, 5698 in Web of Science, 567 in Scopus, and 104 in SPORTDiscus. After eliminating duplicates (n = 3973), a total of 6804 articles were removed by title, abstract, study design, population, or type of intervention. Consequently, 20 articles were assessed for eligibility, and the assessment was done in full text. Finally, 11 articles were included in the systematic review. These were published between 2002 and 2017, showing the importance of PA in TRD and NRD Flow diagram, Supplementary data). Between 2018 and 2024, a total of 1,973 articles were identified, of which 1,239 were excluded due to study type, 463 due to the study population, and 271 due to the study design. No articles were included from that period.

Principal results and characteristics of studies

Physical activity as a predictor of various factors of mental illness, as well as producing benefits in depression was the highlight of the articles (Carta et al., 2008; Greer et al., 2014, 2016; Mather et al., 2002; Mota-Pereira et al., 2011; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011).

A total of 11 studies met the inclusion criteria. These comprised three independent RCTs ((Mota-Pereira et al., 2011), n = 33; (Mather et al., 2002), n = 86; (Carta et al., 2008), n = 30) and one principal RCT ((Trivedi et al., 2011), n = 126) with eight associated sub-studies ((Greer et al., 2014), n = 39; (Greer et al., 2016) n = 106; (Rethorst et al., 2013b), n = 73; (Rethorst et al., 2013a), n = 122; (Toups et al., 2011), n = 70; (Toups et al., 2017), n = 119; (Rethorst et al., 2016), n = 122) reporting secondary outcomes from the same patient cohort (total unique participants across TREAD-related studies: n = 126). Thus, the total number of unique participants across all studies was 292 (241 from independent RCTs + 126 from TREAD, with eigth sub-studies reporting overlapping data from the TREAD cohort). When counting each sub-study separately (including overlapping TREAD participants), the aggregate participant count across all 11 reports is 567.

The supplemental data and Table 1 included the main characteristics of the accepted studies in this systematic review. First, it is important to highlight that eight of the 11 selected articles were from the same research project in the TRD or NRD population (n=275). Additionally, we have added a table that specifies the different definitions used to characterize the populations analyzed in each included article (Definitions of populations in the studies – Supplementary data). Therefore, no variability was observed in the population and study design (i.e., two groups comparing different volumes of exercise doses) (Greer et al., 2014, 2016; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011). Nevertheless, the other three articles were not part of the same research project (Carta et al., 2008; Mather et al., 2002; Mota-Pereira et al., 2011). Concerning the type of PA assessed, while two articles conducted a novelty concurrent exercise intervention (i.e., including endurance and resistance training in the same session) (Carta et al., 2008; Mather et al., 2002), an endurance training approach was observed in the other articles’ intervention, such as only aerobic continuous exercise (treadmill or cycle-ergometer) (Greer et al., 2014, 2016; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011), or weight-bearing strengthening exercise (Mota-Pereira et al., 2011). None of the control groups performed any particular exercise, but pharmacological treatment (Carta et al., 2008) or health-educational talks (Mather et al., 2002).

In terms of the weekly frequency and time (volume) of PA, variability was shown in the studied articles (see Extended data, (Etxaniz-Oses et al., 2023)). First, there was only one intervention, five days per week (Mota-Pereira et al., 2011), whereas the rest of the articles were performed two days per week. Second, the duration of the session ranges from 45 to 60 minutes. In addition, the period of PA or exercise intervention to analyze the changes ranged from 10 (Mather et al., 2002) to 12 weeks (Greer et al., 2014, 2016; Mota-Pereira et al., 2011; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011); but on the other hand, there was one article with a period of eight months of assessment (Carta et al., 2008), and only one study presented a follow-up assessment after 24 weeks (Mather et al., 2002). All the reviewed studies used moderate intensities for exercise and continuous endurance training controlled by a heart rate monitor (Carta et al., 2008; Greer et al., 2014, 2016; Mota-Pereira et al., 2011; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017) and rate of perceived exertion (Mota-Pereira et al., 2011). However, in two studies, the intensity was not mentioned (Carta et al., 2008; Mather et al., 2002), and another survey reported endurance training on a treadmill with an absolute exercise intensity of 5 km/h (Mota-Pereira et al., 2011). In this sense, the rest of the articles performed two different exercise doses (i.e., low dose: 4 kcal/kg/week (KKW) equivalent to 75 min/week, and high dose: 16 KKW equivalent to 210 min/week (Greer et al., 2014, 2016; Rethorst et al., 2013a, 2013b, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011).

Overall, after PA intervention, in all but one of the reviewed articles (Rethorst et al., 2016), depression symptoms and functioning parameters improved, in both TRD and NRD. The Hamilton Rating Scale for Depression (HRSD), Inventory of Depressive Symptoms-Clinical rated (IDS-C), Clinical Global Impression (CGI), and Geriatric Depression Scale (GDS), were the most widely used scales for depression screening (Carta et al., 2008; Mather et al., 2002; Mota-Pereira et al., 2011; Rethorst et al., 2013a, 2013b). In addition, quality of life, as assessed with instruments such as the SF-36 and the World Health Organization Quality of Life (WHOQOL) also improved significantly (Greer et al., 2016). Further, sleep quality (Rethorst et al., 2013a), pro-inflammatory levels (Rethorst et al., 2013b), and some biomarkers of depression (Toups et al., 2011) showed better values after PA intervention. In addition, executive function and working memory improved with increasing exercise volume. Still, psychomotor speed, attention, visual memory, spatial planning, instrumental and expressive roles, and cognitive function improved irrespective of the type or dose of exercise participants undertook (Greer et al., 2014, 2016). The remission of the participants also improved and was achieved in response to antidepressants with the PA as an adjuvant program on the different traits of TRD and NRD. Thus, in one of the RCTs, it was shown that 29.5% of the participants achieved remission (Trivedi et al., 2011), and the remission was analyzed with different scales, such as HRSD, IDS-C, CGI, and GDS (Mather et al., 2002; Mota-Pereira et al., 2011).

Rethorst et al. (2016) examined whether the atypical depression subtype predicted exercise response in 122 participants from the TREAD study. The authors did not find a significant improvement in depressive symptoms associated with the atypical subtype (small effect, p = 0.0735). However, a non-significant trend toward a greater response was observed in the atypical depression group assigned to the high-dose exercise condition. The authors concluded that atypical depression could be a potential moderator of exercise response, but that these findings should be considered hypothesis-generating due to limited statistical power.

Adherence to exercise interventions and dropout rates varied across studies. In the three independent RCTs, Mather et al. (2002) reported the highest adherence to supervised group exercise (67% mean attendance) and no dropouts. Mota-Pereira et al. (2011) reported a high adherence rate (91%) defined as completing ≥50% of prescribed walks per week, with a 6% dropout rate. In the TREAD study (Trivedi et al., 2011), adherence was dose-dependent: the low-dose group (4 KKW) had a median adherence of 99.4%, while the high-dose group (16 KKW) had a median adherence of 63.8%. The overall dropout rate during the 12-week acute phase was 17.2% (21/122), with main reasons including inconvenience/time constraints (n = 10), non-compliance (n = 7), and desire for alternative treatment (n = 5). In TREAD substudies, attrition was higher when additional procedures were required (e.g., blood draw for biomarkers: 42-44% dropout; cognitive testing: 20% dropout from eligible participants). (Carta et al., 2008) did not report adherence or dropout rates. These findings highlight that adherence to high-dose exercise is challenging in TRD/NRD populations, and that dropout rates are substantial, particularly in studies requiring additional assessments.

A meta-analysis was not conducted because several studies did not report the statistical data required for quantitative synthesis (e.g., standard deviations or effect estimates for remission/response). For this reason, a meta-analysis can actually assess heterogeneity, which can be taken into consideration in the final quality of evidence and interpretation.

Risk of bias

Table 2 summarizes the PEDro scale and Oxford’s Evidence levels for the included articles in this systematic review. All the articles except one (Carta et al., 2008), were classified as excellent. Inadequate concealed allocation to groups and the lack of similarity in the comparison groups at baseline were the methodological limitations in the article, which was considered a “good” quality description. All papers were RCT, and for this reason, Oxford’s Evidence levels were 1b in six articles (Greer et al., 2016; Mather et al., 2002; Rethorst et al., 2013a, 2016; Toups et al., 2017; Trivedi et al., 2011) and five 2b (Carta et al., 2008; Greer et al., 2014; Mota-Pereira et al., 2011; Rethorst et al., 2013b; Toups et al., 2011). All studies performed comparisons of both intra-group (pre- and post-PA-intervention) and inter-group (exercise vs. control group) when presented. Further, inclusion and exclusion criteria were defined, and blinding eligibility was presented.

Considering the RoB 2 tool across the 11 included studies, five main trials (Mather et al., 2002; Mota-Pereira et al., 2011; Rethorst et al., 2013a, 2016; Trivedi et al., 2011) presented an overall risk of bias rated as ‘some concerns’, primarily due to lack of participant blinding (unavoidable in exercise interventions) and non-pre-specified post-hoc analyses. The remaining six studies were rated as having ‘high risk’ of overall bias (Risk of Bias RoB2, Supplementary data).

Summary of findings

The quality of the evidence varied according to the outcome assessed. For the primary outcome of depressive symptoms, the quality was moderate (Mather et al., 2002; Mota-Pereira et al., 2011; Trivedi et al., 2011), based on three clinical trials with moderate risk of bias. For the secondary outcomes, the quality was low for sleep quality (Rethorst et al., 2013a) and very low for psychosocial functioning, cognition, biomarkers, quality of life and positive valence symptoms, due to serious methodological limitations, inconsistency across studies, and indirectness of the measurements (Carta et al., 2008; Greer et al., 2014, 2016; Rethorst et al., 2013b, 2016; Toups et al., 2011, 2017) (Table 3).

Heterogenity

As shown in Table 4, clinical, methodological, and statistical heterogeneity were assessed across the three studies reporting depressive symptom outcomes. Marked heterogeneity was observed across populations (i.e., adults with TRD vs NRD), interventions (i.e., individual walking vs group-based exercise vs dose-comparison designs), and outcome measures (i.e., HAMD17, HRSD, and IDS-C₃₀). Statistical heterogeneity between the studies by Mather et al. (2002), and Mota-Pereira et al. (2011) was low (I² = 0%). However, the (Greer et al., 2014, 2016; Rethorst et al., 2013b, et al., 2013a, 2016; Toups et al., 2011, 2017; Trivedi et al., 2011) studies could not be included in this pooled estimate due to the use of a different effect measure. Furthermore, this study was not independent of the other TREAD-derived sub-analyses included in this review. Consequently, we opted for a narrative synthesis rather than a meta-analysis. On the other hand, the study by Carta et al. (2008) presented the following characteristics: Clinically, it included women aged 40-60 years with TRD, a group-based exercise intervention (2 sessions/week of 60 minutes for 32 weeks) versus TAU, and the outcome was quality of life (WHOQOL-Bref physical domain). Methodologically, it was a parallel-group RCT (2:1 allocation ratio) with unreported randomization, no allocation concealment, no intention-to-treat (ITT) analysis, and no reporting of losses to follow-up. Statistically, the effect measure was the mean difference, showing an improvement of +2.2 points in the exercise group (p < 0.03). As this was a single study, I² could not be calculated.