Is resistance training with external loads superior to unloaded exercise in the management of chronic low back pain? A systematic review and meta-analysis

Abstract

Introduction: Chronic non-specific low back pain is a leading cause of disability worldwide. Resistance training using external loads is common in rehabilitation, but its added benefit over unloaded exercise remains uncertain, particularly regarding physical and psychological outcomes.
Method: This systematic review and meta-analysis, registered on PROSPERO (CRD42022366975), included randomized controlled trials comparing externally loaded resistance training to unloaded exercise in adults with chronic non-specific low back pain. Primary outcomes were pain intensity and disability. Secondary outcomes included back muscle endurance, maximal strength, fear-avoidance beliefs, and pain catastrophizing. Random-effects meta-analyses were conducted, stratified by follow-up duration.
Results: Thirteen trials including 778 participants were analyzed. At follow-up beyond seven weeks, externally loaded resistance training showed a small but significant reduction in pain compared with unloaded exercise (mean difference –0.52 on a 0–10 scale; 95% CI –0.92 to –0.08). No significant differences were found at short-term or post-washout assessments. Effects on disability were inconsistent. Resistance training improved back muscle endurance and may enhance long-term maximal strength, though wide prediction intervals limit certainty. No meaningful differences were found for psychological outcomes, and pain catastrophizing was reported in only one trial.
Conclusion: Externally loaded resistance training is safe and feasible for chronic non-specific low back pain, but its effects on pain, disability, and psychosocial outcomes are similar to unloaded exercise. Improvements likely reflect exposure, adherence, and therapeutic context rather than load intensity alone. Exercise prescription should remain individualized and embedded within a biopsychosocial framework.

Introduction

Low back pain (LBP) is the most prevalent musculoskeletal disorder (1), affecting approximately 540 million people worldwide at any given time (2). When no specific patho-anatomical cause can be identified, LBP is classified as non-specific (NS-LBP) (3). Among all musculoskeletal conditions, chronic NS-LBP represents the leading cause of disability, as measured in disability-adjusted life-years (4), posing a significant burden on healthcare systems and affected individuals.

Traditionally conceptualized through a biomechanical lens, chronic NS-LBP is now widely recognized as a multifactorial condition in which pain and disability arise from dynamic interactions among biological, psychological, and social factors (5). This paradigm shift has supported the adoption of the biopsychosocial (BPS) model as the best practice for its understanding and management (6).

Exercise is widely recommended as the first-line treatment for chronic NS-LBP, as highlighted by the National Institute for Health and Care Excellence (NICE) guidelines (7).

Beyond its physical effects, exercise can be conceptualized as a multidimensional therapeutic intervention, capable of modulating psychological and behavioral processes (8). Evidence suggests that physical exercise may positively influence factors such as anxiety, mood, fear-avoidance, and pain-related beliefs (9,10), reinforcing its role within the BPS model.

Despite its recognized value, the optimal characteristics of exercise interventions for chronic NS-LBP remain unclear. Current literature shows limited differences in effectiveness between various exercise modalities, with generally modest improvements in pain and disability (11,12). This suggests a need to move beyond broad exercise categories and instead focus on specific intervention components that may drive better outcomes.

One such component is the type and intensity of resistance, or load, used during training. Resistance training (RT) is a form of physical exercise involving internal (e.g., body weight) or external (e.g., free weights, machines, bands) resistance to stimulate skeletal muscle contractions, aiming to enhance strength, power, muscular endurance, and muscle mass (13,14). Although RT is commonly applied in rehabilitation settings, the optimal dosage and resistance parameters for chronic musculoskeletal conditions remain poorly defined (15-17). In research, the term “load” typically refers to external resistance, distinguishing “loaded” from “unloaded” exercises (18).

Loaded exercises may induce adaptations across multiple, interrelated domains. They have been associated with greater gains in muscle strength (19), soft tissue capacity (20), cartilage turnover (21), as well as enhanced neurochemical responses which have been linked to increases in pain thresholds (22,23). Incorporating external load may serve as a graded exposure stimulus, helping patients confront fears, rebuild confidence in movement, and challenge beliefs related to fragility or harm (7). Such mechanisms may be particularly valuable for targeting maladaptive responses such as fear-avoidance, kinesiophobia, and pain catastrophizing, which are known to contribute to the persistence of chronic NS-LBP (24).

Although the addition of external load to an exercise program may influence the adaptations described above, its specific contribution to clinical outcomes remains poorly understood (25,26). Available studies are hindered by inconsistent terminology, insufficiently described loading protocols and inadequate control conditions lacking active comparators (17,27). Therefore, it remains unclear to what extent the inclusion of external resistance affects the multifaceted nature of chronic NS-LBP.

This systematic review and meta-analysis consequently seeks to address the following question: “To what extent does the use of external load in resistance training influence symptoms, function, and psychosocial factors in chronic NS-LBP management?”

The primary objective of this systematic review and meta-analysis was to assess the effects of RT with external resistance compared to unloaded exercises on pain and disability. Secondary outcomes included physical performance metrics (muscle endurance, maximal strength) and psychosocial variables (fear-avoidance, kinesiophobia, and pain catastrophizing).

Methods

This study followed the guidelines outlined in the ‘Cochrane Handbook for Systematic Reviews of Interventions’ (28) and was structured in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) Statement (29). The PRISMA checklist is reported in ‘Appendix A’.

Eligibility criteria

The eligibility criteria were developed based on a research question structured according to the Population–Intervention–Comparison–Outcome (PICO) framework. The population of interest comprised adults (≥18 years) diagnosed with chronic NS-LBP. Eligible Interventions included RT programs that incorporated external loads, either alone or in combination with internal resistance (bodyweight). To isolate the effect of external loads in RT, studies were excluded if the experimental group received multimodal interventions involving additional exercise modalities (e.g., aerobic training, motor control exercises) or therapies (e.g., manual therapy).

For the Comparison group, only exercise interventions without any form of external resistance—hereafter referred to as unloaded exercises (UE)—were considered eligible. Eligible studies had to report on at least one of the primary Outcomes.

Only randomized controlled trials (RCTs) were included. Studies were excluded if they included individuals with specific LBP (e.g., fractures, radicular pain, radiculopathies, spinal stenosis, or axial spondylarthritis) or with a history of spinal surgery. Finally, while studies not available in English or Italian were excluded, no restrictions were applied regarding publication date or methodological quality.

Information sources and search strategy

Two independent reviewers (MR and DF) each developed and conducted a separate literature search, working in a blinded manner without any prior agreement on which databases to search. The search strategies were based on the previously defined PICO framework, although no terms were included for the Comparison component to reduce the risk of missing relevant studies.

As recommended by the Cochrane Handbook (28), three core electronic databases were searched: PubMed, Cochrane Library, and EMBASE. Filters for “Randomized Controlled Trial” in PubMed and “Trial” in the Cochrane Library were applied to restrict the results to study types relevant to the review’s objective. To enhance comprehensiveness, two multidisciplinary databases (Scopus and Web of Science), a physical therapy-specific database (The Physiotherapy Evidence Database - PEDro) and gray literature sources (Google Scholar) were also queried.

The searches were first conducted in September 2022 and subsequently updated in October 2024 and March 2025. Full search strategies, including the specific databases, search terms, and filters used, are provided in ‘Appendix B’.

Finally, a planned citation search was also performed by screening the reference lists of previously published systematic reviews investigating the effects of exercise on LBP (11,12,25,27,30-35) to enhance the comprehensiveness of the search strategy.

Selection process

The search results were imported into Rayyan software (Online), where duplicate records were manually removed before screening (36).

Two independent reviewers (MR and DF) conducted a blinded, independent screening of titles and abstracts. Studies meeting the eligibility criteria at this stage were exported to an Excel file, and their full texts were retrieved. The reviewers then independently assessed the full texts in a blinded manner. Any discrepancies during the screening process were resolved through discussion between MR and DF or, if necessary, with the involvement of a third reviewer (AP). Inter-rater agreement was not calculated.

Data collection process and data items

Two independent reviewers (MR and DF) extracted data from each included study. To standardize the process, a structured synoptic Excel spreadsheet was developed, in line with the recommendations of the Cochrane Handbook of Systematic Reviews (28).

The extracted data included: (i) Study characteristics: lead author, publication year, sample size, duration, adverse events and follow-up. (ii) Population details: sex, age, height, weight, BMI, baseline pain, baseline disability, duration of symptoms and baseline level of activity. (iii) Intervention characteristics: exercise type and name, type of resistance, total duration of the intervention, weekly frequency, session duration, number of sets and repetitions, baseline and peak intensity and progression parameters. (iv) Outcomes: mean and standard deviations at every follow-up, as reported in the studies or, when not available, derived from other summary statistics using the conversion methods recommended in the Cochrane Handbook (28).

Outcome measures

Pain intensity and disability were the primary outcomes of the systematic review. Pain intensity was assessed based on the Visual Analog Scale (VAS) and Numeric Pain Rating Scale (NPRS), while disability was evaluated using the Oswestry Disability Index (ODI) and Roland and Morris Disability Questionnaire (RMDQ).

Secondary outcomes included measures of physical performance and psychosocial variables. For physical performance, eligible measures comprised back muscle endurance, as assessed by the Biering-Sørensen test (BST), and maximal strength, expressed in terms of muscle force (newtons, N) or joint torque (newton-meters, N·m). Psychosocial outcomes included fear-avoidance beliefs related to physical activity (Physical Activity subscale of the Fear-Avoidance Beliefs Questionnaire, FABQ-PA), kinesiophobia (Tampa Scale of Kinesiophobia, TSK) or pain catastrophizing (Pain Catastrophizing Scale, PCS). The FABQ-PA subscale was selected as a relevant measure of fear-avoidance beliefs due to its specific focus on movement and exercise (37), its established construct validity (37,38), and its documented association with disability and less favorable clinical outcomes in individuals with chronic NS-LBP (39,40).

Study risk of bias assessment and certainty of evidence grading

Two independent reviewers (MR and DF) conducted a blinded risk of bias assessment for each primary and secondary outcome using the revised version of the Cochrane Risk of Bias Assessment Tool (RoB2) (41). Risk-of-bias plots were generated using the Robvis tool (42). The Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) system was used to assess the certainty of evidence at each follow-up for each outcome. The grading process was conducted using ‘GRADEpro’ (43).

Effect measures

The effect size used for the meta-analysis was calculated as the standardized mean difference (SMD) with a 95% confidence interval (95%CI) for outcomes reported in non-convertible units or assessed using different rating instruments. When all studies employed the same measurement scale for a given outcome, the mean difference (MD) with a 95%CI was used instead. Pain intensity scores reported on 100-point scales were rescaled to 0–10, and BST times were converted from seconds to minutes to ensure comparability across studies.

To explore treatment effects within each study, we also calculated within-group changes using pre- and post-intervention data (SMDs with 95%CI), where available. In addition, we computed between-group differences at post-intervention using SMDs based on post-treatment means and standard deviations (SDs). These results are presented descriptively in ‘Appendix D’, along with indications of statistical significance, to facilitate interpretation of individual study findings.

Unit-of-analysis issues in trials with shared comparison groups were addressed by splitting the shared group’s sample size evenly, while retaining the original means, SDs, and participant counts, in accordance with the Cochrane Handbook guidelines (28).

Synthesis methods

To generalize the findings beyond the included studies, an unconditional inference model was employed. The restricted maximum likelihood estimation method (REML) was used to estimate between-study variance in the random-effects model. Post-intervention results were categorized based on the duration of the exercise program. ‘Short training programs’ (STP) were defined as those with follow-up data collected within 6 weeks or less, while ‘extended training programs’ (ETP) were defined as those with follow-up data collected after at least 7 weeks. This cutoff was based on evidence suggesting that programs lasting at least 7 weeks may be more effective for managing chronic pain (44). Additionally, results were analyzed, where applicable, at the final follow-up to assess long-term differences between loaded and unloaded interventions. For inclusion in the ‘Post-Washout’ (PW) follow-up analysis, data had to be collected at least 6 months after the study’s initiation, following a washout period (i.e., a treatment-free interval to minimize carryover effects). When multiple eligible follow-ups were reported, data from the most distant follow-up were used. Heterogeneity was assessed using Cochran’s Q test (χ²), I² statistic, and tau². These metrics were reported descriptively, without applying fixed thresholds for interpretation. In addition, 95% prediction intervals (95%PI) were calculated and used as the primary indicator of the expected dispersion of true effects across comparable future settings (45,46). Meta-analyses were conducted using R version 4.2.3 for Mac OS (47), with the ‘meta’ package and its “metagen” function (48). The complete analysis output (RStudio report) is publicly available on the Open Science Framework (OSF) at Online.

Post-hoc sub-group and sensitivity analysis

To explore potential sources of heterogeneity in the primary outcomes, post hoc subgroup analyses were performed based on: (i) methodological quality, classified using the RoB2 tool as high or low risk of bias; (ii) baseline pain intensity, categorized according to the Pain Monitoring Model (49) as low (<2/10), acceptable (2–5/10), or high (>5/10); and (iii) the type of resistance used in the intervention, comparing programs using only external resistance versus those combining internal and external resistance. Results of these analyses are presented in ‘Appendix F’.

Sensitivity analyses were conducted to assess the robustness of the findings and to examine the influence of individual studies on overall estimates and heterogeneity. These analyses were performed only for meta-analyses that included data from at least five RCTs, to ensure sufficient statistical power and stable estimation of heterogeneity parameters (50). Following Viechtbauer (2010), a leave-one-out approach was applied, iteratively excluding each study from the model and assessing changes in the pooled effect size and heterogeneity parameters (I², tau², and Q) (50). A study was considered influential if its removal resulted in: (a) a substantial reduction in heterogeneity (≥25% reduction in tau² or I²), or (b) a meaningful change in the direction or magnitude of the pooled effect. Studies meeting one or both of these criteria were excluded in subsequent analyses to evaluate whether they had a disproportionate impact on the overall results. This approach allowed for a cautious and transparent assessment of the stability of the conclusions in the presence of statistical heterogeneity.

Reporting bias assessment

Publication bias was assessed using contour-enhanced funnel plots, which illustrate the relationship between the size of the studies and effect sizes (51). For meta-analyses with at least 10 comparisons (52), Egger’s regression test was used as a quantitative measure of reporting bias. The “funnel.meta” and “metabias” (47) functions in RStudio (47) were used to evaluate the reporting bias.

Results

Registration and Protocol

The review protocol was registered in PROSPERO (CRD42022366975), but several amendments were made during the review process. Due to the limited number of available studies and incomplete reporting of intervention data, the planned meta-regression analyses, as specified in the original protocol, could not be performed. Consequently, modifications were made to both the title and statistical analyses in this manuscript.

Study selection

The search strategy identified a total of 1710 records, of which 1609 were retrieved from databases and 101 from citation search. After removing 555 duplicates using Rayyan, 1155 unique records remained. These were screened based on titles and abstracts, resulting in the exclusion of 1082 records for not meeting the inclusion criteria. Overall, 73 records were identified for retrieval. Of those, one (53) could not be retrieved despite attempts to contact the authors. The remaining 72 full-texts were screened for eligibility. Among them, 12 studies identified from database searches (54-65) and 3 from citation searching (66-68) met the eligibility criteria, totaling 15 papers included in the review (Fig. 1). However, Michaelson et al. (62) and Mannion et al. (68) reported supplementary data from the same participant samples described in Aasa et al. (54) and Mannion et al. (67), respectively. These reports were therefore integrated into the original studies, resulting in a final count of 13 unique RCTs included in the meta-analysis.

A total of 57 full-text articles were excluded following the eligibility assessment. Four records were excluded as they were conference abstracts (69–72). An additional four studies were excluded due to the absence of participant randomization; three were published studies (73-75), and one was an unpublished thesis (Costa K. Effects of a trunk strengthening program on pain perception, strength, and flexibility in patients with non-specific low back pain. Doctor of Physiotherapy thesis, Bond University; 2010. Online. Last accessed on 22/09/2025). Four studies were excluded because they enrolled participants with specific forms of LBP (76-79). Eighteen studies were excluded for failing to incorporate external resistance into their exercise protocols (80-97), and nine were excluded because they did not isolate resistance training from other concurrent interventions (98-106). Thirteen studies were excluded for employing either passive control conditions or applying resistance training to both study groups (107-119). Finally, three studies were excluded due to language restrictions (120-122), and two were excluded because the outcomes assessed were not relevant to the review question (123-124).

FIGURE 1 -. Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) flow diagram of study selection.

Description of Study Populations

This systematic review included 13 studies published between 1999 and 2022 involving a total of 778 subjects with chronic NS-LBP. Among them, 395 individuals participated in an RT program incorporating external loads. Participants in the Resistance Training Group (RTG) had an average age of 39.3 years (range: 20.2-61.5 years) and a mean symptom duration of 3.96 years (range: 3.9 months to 13 years). The Unloaded Exercise Group (UEG) had a similar average age of 39.3 years (range: 20.8-57.2 years) and a mean symptom duration of 3.21 years (range: 4.1 months to 9.7 years). The average height and weight were 168.5 cm (range: 159-183 cm) and 74.4 kg (range: 61.7-88.4 kg), respectively, in the RTG, and 167.5 cm (range: 156-182 cm) and 72.3 kg (range: 60.3-86.2 kg), respectively, in the UEG. Perceived pain levels during activities averaged 5.4/10 (range: 2.9-8) in the RTG and 5.7/10 (range: 2.7-7.4) in the UEG. Gender distribution was reported in 10 of the 13 included RCTs. In RTG, 198 men (61.9%) and 122 women (38.1%) were enrolled, while in UEG, 218 men (64.1%) and 122 women (35.9%) were included. Detailed participants’ characteristics are presented in Table 1.

Description of Exercise Interventions

Both RTG and UEG followed equivalent training periods and weekly frequencies. Training duration typically lasted for eight weeks, ranging from three (64) to sixteen weeks (59). Participants trained an average of three sessions per week, ranging from one (65) to seven sessions per week (66), including both supervised and unsupervised formats. The median number of sessions across studies was 12 (interquartile range: 12-16) and was comparable between groups.

Training intensity and Resistance modalities

The RTG program intensity was quantified using various metrics. Five studies used an estimated percentage of One Repetition Maximum (1RM) (54,55,58,59,66). One study employed repetition in reserve (RIR) (57). One study used an estimated percentage of maximal voluntary isometric contraction (66). Resistive modalities included: free weights (54,57,59,63,66), weight-stack machines (58,59,63,65,67), elastic bands (55,56) and isokinetic machines (60,61,64). Seven studies used multi-joint exercises (54-57,59,63,66), while six targeted the lumbopelvic muscles (58,60,61,64, 65,67).

TABLE 1 -. Characteristics of the included participants

Study	G	n	Male/Female, n	Age, Years	Weight, Kg	Height, Cm	Baseline pain	Baseline Disability	Type of LBP	Adverse events*, n	Duration of Symptoms, Years
Aasa et al., (54)	RTG	35	15/20	42 ± 10	74 ± 13	174 ± 8	43 (0-100) ± 24	7 (0-24) ± 4	C	2 (related)	5.98 ± 5.96
	UEG	35	16/19	42 ± 11	78 ± 15	172 ± 10	47 (0-100) ± 28	7 (0-24) ± 5	C	0	6.52 ± 3.64
Cai et al., (66)	RTG (a)	28	14/14	28.9 ± 5.3	61.7 ± 12.6		3.5 (0-10) ± 1		C	1 (unrelated)	1.19 ± 0.67
	RTG (b)	28	14/14	26.1 ± 4.1	61.7 ± 10.8		3.4 (0-10) ± 0.9		C	1 (unrelated)	1.24 ± 0.63
	UEG	28	14/14	26.9 ± 6.4	60.3 ± 12.1		3.6 (0-10) ± 1.1		C	2 (unrelated)	1.32 ± 0.52
Calatayud et al., (55)	RTG	42		52 ± 11	76 ± 19	164 ± 10	6.2 (0-10) ± 2	4.3 (0-24) ± 2	C
	UEG	43		50 ± 12	72 ± 14	165 ± 7	6.3 (0-10) ± 2	5.1 (0-24) ± 3	C
Castro et al., (56)	RTG	12	0/12	61.5 ± 7.5	73.2 ± 13.3	159± 13	6.75 (0-10) ± nr	13.8 (0-24) ± nr	C
	UEG	13	0/12	57.2 ± 10	67.3 ± 11	156 ± 7	6.1 (0-10) ± nr	13.2 (0-24) ± nr	C
Gibbs et al. (57)	RTG	32	19/13	35.6 ± 12.4			5.6 (0-10) ± 2.1	18.8 (0-100) ± 7.8	C	1 (related)	4.42± 3.68
	UEG	32	17/15	33.5 ± 11.9			5.8 (0-10) ± 2.2	15.4 (0-100) ± 8.7	C	1 (related)	4.79 ± 5.62
Helmhout et al. (58)	RTG	71	69/3	37 ± 11	85 ± 12	183± 8		8.3 (0-24) ± 4.8	SA, C	1 (related)
	UEG	56	54/2	35 ± 11	86 ± 11	182 ± 7		7.9 (0-24) ± 4.4	SA, C	0
Kell et al., (59)	RTG	9	6/3	40.1 ± 8.7	88.4 ± 22.4	174 ± 8	5.4 (0-10) ± 0.9	40.4 (0-100) ± 2.4	C		2.29 (range 0.5 - 8)
	UEG	9	5/4	36.7 ± 8.9	81.7 ± 11.5	173 ± 10	5.1 (0-10) ± 0.8	39.8 (0-100) ± 2.3	C		2.29 (range 0.5 - 8)
Mannion et al., (67)	RTG	49	22/27	43.7 ± 10.1	70.3 ± 13.4	172 ± 9	4.2 (0-10) ± 1.8	8 (0-24) ± 5.1	C	0	13 ± 10
	UEG	50	23/27	45.2 ± 9.7	68 ± 12.3	170 ± 11	4.1 (0-10) ± 1.8	7.7 (0-24) ± 4.7	C	2 (related)	9.7 ± 9.1
Nambi et al., (60)	RTG	20	20/0	21.1 ± 1.4	66.8 ± 1.5	167 ± 14	7.2 (0-10) ± 0.4		C		0.325 ± 0.06
	UEG (a)	20	20/0	22.1 ± 1.3	66.8 ± 1.5	166 ± 15	7.3 (0-10) ± 0.3		C		0.35 ± 0.05
	UEG (b)	20	20/0	21.4 ± 1.4	66.4 ± 1.4	165 ± 13	7.4 (0-10) ± 0.6		C		0.367 ± 0.06
Nambi et al., (61)	RTG	15	15/0	20.2 ± 1.6	65.6 ± 1.4	168 ± 14	7.3 (0-10) ± 0.5		C		0.342 ± 0.05
	UEG (a)	15	15/0	21.3 ± 1.2	66.2 ± 1.4	165 ± 18	7.1 (0-10) ± 0.6		C		0.342 ± 0.03
	UEG (b)	15	15/0	20.8 ± 1.6	65.5 ± 1.5	168 ± 15	7.3 (0-10) ± 0.6		C		0.358 ± 0.04
Santos et al., (63)	RTG (a)	14		49.1 ± 7.1	78.2 ± 17.8	161± 7	7 (0-10) ± 1.7	14.8 (0-24) ± 3.2	C		6.2 ± 5.4
	RTG (b)	15		43.3 ± 12.1	79.1 ± 12.4	163± 7	8 (0-10) ± 1.3	16.8 (0-24) ± 5	C		3.7 ± 2.4
	UEG	14		46.7 ± 15.3	86.2 ± 11.4	160 ± 6	6.53 (0-10) ± 1.2	16.1 (0-24) ± 5.4	C		5 ± 3.88
Sertpoyraz et al., (64)	RTG	20	4/16	38.8 ± 7.8			4.9 (0-10) ± 0.9	16.6 (0-100) ± 8.1	C		2.4 ± 2.1
	UEG	20	5/15	38.3 ± 7.4			5.4 (0-10) ± 1.3	18.8 (0-100) ± 7.8	C		3.8 ± 4.3
Smith et al., (65)	RTG (a)	16					30.1 (0-100) ± 17.2	39.2 (0-100) ± 14.7	C
	RTG (b)	17					28.7 (0-100) ± 17.4	35.7 (0-100) ± 12.6	C
	UEG	13					26.8 (0-100) ± 9	32.7 (0-100) ± 5.9	C

Adverse events related to the exercise intervention

Reporting of adverse events was heterogeneous and predominantly descriptive. Overall, the RTG did not demonstrate a higher incidence of events compared with the UEG. Aasa et al. (54) and Helmhout et al. (58) each reported one withdrawal in the RTG due to symptom aggravation, with no corresponding events in the UEG. In contrast, Mannion et al. reported two dropouts in the UEG due to LBP flare-ups, with none in the RTG (67). Gibbs et al. observed one transient flare-up in each group, which was managed through temporary reduction of load or repetitions (57). Cai et al. documented three musculoskeletal injuries, equally distributed across groups and deemed unrelated to the training interventions (66). The remaining RCTs did not provide explicit information regarding adverse events associated with the interventions. Taken together, adverse events were infrequent, mild in nature, and evenly distributed between groups. The number of adverse events reported in each group is summarized in Table 1.

Progressive overload strategies

While studies using isokinetic machines (60,61,64) did not incorporate progressive overload, all other RT programs implemented progressive overload methods.

Aasa et al. (54), Helmhout et al. (58), and Smith et al. (65) increased the load by 2.5 kg or 5% when participants exceeded the target repetitions. Cai et al. (66) and Kell et al. (59) adjusted load based on planned 10RM reassessments.

Gibbs et al. (57) focused on progressing movement complexity for four weeks before increasing intensity to 1RM in the final week. Castro et al. (56) and Calatayud et al. (55) progressively increased elastic resistance every two weeks, starting from 20RM and reaching 10RM. Santos et al. (63) applied linear periodization, gradually increasing intensity while decreasing training volume. Further details on the training programs used in the RTG and the UEG are presented in ‘Appendix C’.

Results of individual studies

The results of the individual studies categorized into STP, ETP and PW follow-ups, along with the calculated within-group improvements, are presented in Appendix D. Notably, none of the included studies reported post-intervention scores for kinesiophobia, and only one study (57) provided data on pain catastrophizing.

Short Training Program (STP) Follow-up

At the STP follow-up, improvements in pain intensity (59,60,61,63), disability (59,63), and muscle endurance (63) were generally greater in the RTG than in the UEG. Most studies reported within-group improvements in pain and disability, though gains in muscle endurance and strength were smaller. One UE program (63) demonstrated a significant improvement in VAS scores despite a slight decline in BST performance.

Extended Training Program (ETP) Follow-up

At the ETP follow-up, pain intensity (66,59,65) and disability (54,55,57-59) improved more in the RTG, whereas muscle endurance and strength showed similar changes in both groups. Calatayud et al. (55) reported a mean BST performance change of 44.39 seconds, although pain and disability improvements were only a minimal clinically important difference (MCID) (125). Gibbs et al. (57) found reductions in pain catastrophizing on the PCS (0-52) in both groups at post-intervention. RTG improved from 16.0± 10.3 to 8.5 ± 9.6. UEG improved from 18.0 ± 11.7 to 8.9 ± 8.5.

Post-Washout (PW) Follow-up

At the PW follow-ups, primary outcomes generally continued to improve. Aasa et al. (54) reported a slight worsening in pain intensity and disability, but muscle endurance and strength improved. Helmhout et al. (58) found that the UEG experienced a slight decline in maximal strength, with no change in disability at 8 and 14 months. Regarding PCS scores, Gibbs et al. (57) reported a slight increase in pain catastrophizing in both groups: RTG from 8.5 ± 9.6 to 9.8 ± 10.1. UEG from 8.9 ± 8.5 to 10.8 ± 10.5.

Between-Group Comparisons

Seven studies (54-58,64,66) found no significant differences between RTG and UEG after treatment. Five studies (59,60,63-65) concluded that RT led to superior outcomes. Conversely, Nambi et al. (61) reported that UE was superior to RT.

Risk of bias in studies

The detailed ROB2 assessment is provided in Appendix E, including a summary and traffic light plots for each outcome. None of the ROB2 domains were rated as having a high risk of bias. However, five RCTs (58-60,65,66) were judged to have a high overall risk of bias due to the cumulative impact of multiple domains rated as “Some Concerns.”

Primary outcomes meta-analyses

Data on pain intensity were reported in 12 RCTs, while data on disability were available from 10 RCTs. One study (56) was excluded from the quantitative synthesis due to the unavailability of summary statistics, despite attempts to contact the authors.

Pain intensity

At the STP follow-up (6 RCTs, n = 282), the MD was -0.48/10 (95%CI: [-1.15, 0.18]; p = 0.15), indicating no clear average between-group difference. The 95%PI (-3.25 to 2.44) suggests a wide range of possible effects, including the possibility of both meaningful pain reduction and no benefit. Heterogeneity statistics were I² = 97% and tau² = 1.02 (Fig. 2 – Forest Plot 1a). At ETP follow-up (7 RCTs, n = 474), the pooled MD was −0.52/10 (95%CI: [−0.92, −0.08]; p = 0.022), suggesting a small average reduction in pain favoring RTG. The 95%PI (–1.88-0.83) includes the null, indicating that future studies may still find no effect in some settings. Heterogeneity statistics were I² = 67% and tau² = 0.28 (Fig. 2 – Forest Plot 1b). At PW follow-up (5 RCTs, n = 345), the pooled MD was −0.36/10 (95%CI: [−1.18, 0.47]; p = 0.40), again showing no clear average difference. The 95%PI (–3.25-2.54) indicates a broad distribution of possible effects across studies, consistent with the variability captured by I² = 97% and tau² = 1.09 (Fig. 2 – Forest Plot 1c).

Disability

At STP follow-up (3 RCTs, n = 206), the pooled SMD was −2.04 (95%CI: [−3.92, −0.16]; p = 0.033), suggesting a large average reduction in disability favoring the RTG. The 95%PI (−4.43-0.35) indicates that future studies could observe effects ranging from very large improvements favoring the RTG to moderate effects favoring the UEG. Heterogeneity statistics were I² = 93% and tau² = 4.18 (Fig. 2 – Forest Plot 2a). At the ETP follow-up (7 RCTs, n = 505), the pooled SMD was –0.40 (95%CI: [-1.00, 0.20]; p = 0.19), indicating that the average improvement in disability was similar between groups. The 95%PI (–1.68-0.88) suggests that future studies may find results ranging from large improvements favoring either the RTG or the UEG. Heterogeneity statistics were I² = 82% and tau² = 0.63 (Fig. 2 – Forest Plot 2b). At the PW follow-up (4 RCTs, n = 331), the pooled SMD was 0.06 (95%CI: [−0.16, 0.28]; p = 0.59), suggesting that the degree of improvement in disability was comparable between groups. The 95%PI (–0.43 to 0.55) indicates that future studies are likely to observe small to moderate effects in either direction. Heterogeneity statistics were I² = 0% and tau² = 0 (Fig. 2 – Forest Plot 2c).

FIGURE 2 -. Primary Outcomes Forest Plots: Effects of Loaded and Unloaded Exercise on Pain and Disability. ROB2: Risk of Bias Tool 2; CI: Confidence Interval; PI: Prediction Interval; LL: Lower Limit; UL: Upper Limit; IV: Inverse Variance; W: Weight; MD: Mean Difference; SMD: Standardized Mean Difference; VAS: Visual Analog Scale; ODI: Oswestry Disability Index; RMDQ: Roland and Morris Disability Questionnaire; RTG: Resistance Training Group; UEG: Unloaded Exercise group

Post hoc Subgroup Analyses

Subgroup effects were observed only in the STP meta-analysis for disability, where greater treatment effects were found among participants with high baseline pain (Q = 59.55; df = 2; p < 0.0001), and in interventions combining internal and external resistance (Q = 49.02; df = 1; p < 0.0001). Full subgroup results are reported in Appendix F.

Sensitivity analyses were performed only for pain STP, ETP and PW follow-ups and the disability ETP follow-up, as these were the only meta-analyses including data from more than five RCTs. In the pain STP analysis, the studies by Nambi et al. (60,61) were identified as influential, as their exclusion resulted in a statistically significant pooled effect (MD = −0.53/10; 95%CI: [−1.06, −0.05]; p = 0.0478) and a marked reduction in heterogeneity statistics (I² = 58%, tau² = 0.24). At the PW follow-up, Nambi et al. (61) were again identified as influential due to the reduction of the I² and tau² statistics to zero and the emergence of a significant pooled effect (MD = −0.31/10; 95%CI: [−0.55, −0.07]; p = 0.01). The study by Kell et al. (59) was influential in both the ETP pain and disability analysis. In the pain ETP analysis, its exclusion reduced tau² from 0.276 to 0.091 and I² from 67.3% to 37.8%, and the pooled effect became non-significant (MD = −0.33/10; 95% CI: [−0.69, 0.02]; p = 0.687). In the disability ETP follow-up, tau² dropped from 0.634 to 0.081 and I² from 82.1% to 68.9%, although the effect direction and significance remained unchanged (SMD = −0.169; 95%CI: [−0.44, 0.10]; p = 0.223). These findings suggest that the identified trials contributed disproportionately to between-study heterogeneity and, in some cases, to the statistical significance of the pooled effects.

Secondary outcomes meta-analyses

Although meta-analyses were planned for all secondary outcomes, pooling for the PCS was not feasible, as data were available from only one study (57). Consequently, quantitative syntheses were conducted only for back muscle endurance (4 RCTs), maximal strength (3 RCTs), and FABQ-PA (2 RCTs).

Back Muscle Endurance

At STP follow-up (2 RCTs, n = 56), the pooled MD was 0.50 minutes (95%CI: [0.26, 0.74]; p < 0.001), indicating a statistically significant improvement in the performance of the BST, favoring the RTG. The 95%PI (–1.80 to 2.80) suggests that future studies could observe large effects favoring either treatment. Heterogeneity statistics were I² = 38% and tau² = 0.018 (Fig. 3 – Forest Plot 3a). At ETP follow-up (3 RCTs, n = 173), the pooled MD was 0.47 minutes (95%CI: [0.09, 0.85]; p = 0.015), again indicating a statistically significant improvement in the BST performance favoring the RTG. The very wide 95%PI (–3.80 to 4.73) reflects the considerable uncertainty about the generalizability of the pooled effect. Heterogeneity statistics were I² = 69% and tau² = 0.075 (Fig. 3 – Forest Plot 3b). At PW follow-up, Aasa et al. (54) also reported a significant effect on the BST favoring the RTG (MD = 0.35 minutes; 95%CI: [0.03, 0.67]).

Maximal strength

At STP follow-up (2 RCTs, n = 150), the pooled SMD was 0.00 (95%CI: [–0.30, 0.30]; p = 0.998), indicating that the average change in maximal strength was similar between the RTG and UEG. The 95%PI (–1.96-1.96) suggests that future studies may observe large effects on maximal strength, favoring either the RTG or the UEG. Heterogeneity statistics were I² = 0% and tau² = 0 (Fig. 3 – Forest Plot 4a). At the ETP follow-up (2 RCTs, n = 176), the pooled SMD was 0.22 (95%CI: [–0.70, 0.26]; p = 0.38), again suggesting that average changes were comparable between RTG and UEG. Heterogeneity statistics were I² = 20% and tau² = 0.033 (Fig. 3 – Forest Plot 4b). At PW follow-up (2 RCTs, n = 176), the pooled SMD was 0.40 (95%CI: [0.08, 0.71]; p = 0.013), suggesting a moderate and statistically significant effect favoring the RTG. Heterogeneity statistics were I² = 0% and tau² = 0.0 (Fig. 3 – Forest Plot 4c). The 95%PI was not calculated for ETP and PW due to the limited number of comparisons included.

Fear-avoidance beliefs related to physical activity

At ETP follow-up (2 RCTs, 2 comparisons), the pooled MD was –0.09/24 (95%CI: [–0.88, 0.71]; p = 0.82), indicating that average FABQ-PA scores were similar between RTG and UEG. Heterogeneity statistics were I² = 0% and tau² = 0 (Fig. 3 – Forest Plot 5a).

At PW follow-up (2 RCTs, 2 comparisons), the pooled MD was –0.24/24 (95%CI: [–1.58, 1.09]; p = 0.72), again suggesting no clear difference in FABQ-PA scores between groups. Heterogeneity statistics were I² = 0% and tau² = 0 (Fig. 3 – Forest Plot 5b). The 95%PI could not be estimated at either follow-up due to the restricted number of comparisons available.

Reporting biases

Reporting bias was assessed using funnel plots (Appendix G). The plots showed no significant signs of bias (126). Due to the limited number of comparisons, Egger’s test was applied only to the STP meta-analysis for pain, yielding an intercept of 0.20 (t = −0.57, p = 0.58), suggesting no funnel plot asymmetry.

Quality of the evidence (GRADE Assessment)

The overall certainty of the evidence for primary outcomes was rated as ‘very low’, due to methodological limitations and inconsistency across studies. For back muscle endurance, certainty was rated “moderate” at both STP and ETP follow-ups, suggesting probable benefits of externally loaded resistance training over unloaded interventions. At the PW follow-up, the certainty decreased to “very low” due to the limited number of trials and imprecision. For maximal strength, the certainty was consistently rated “low” across all follow-ups, primarily due to the small number of included studies and risk of bias. For fear-avoidance beliefs related to physical activity, the evidence was judged as “low” certainty at both ETP and PW timepoints, reflecting imprecision and methodological concerns. For pain catastrophizing, certainty was rated as “very low” at both time points, as only one study provided data, limiting confidence in the estimate. A detailed evidence quality assessment is provided in Appendix H.

Discussion

Main findings

This systematic review evaluated current evidence on the effectiveness of RT interventions involving external loads in the management of chronic NS-LBP. Thirteen studies were included, comprising 395 participants enrolled in exercise programs utilizing various external load modalities such as free weights, elastic bands, weight stack machines, and isokinetic devices.

FIGURE 3 -. Secondary Outcomes Forest Plots: Effects of Loaded and Unloaded Exercise on Muscle Endurance, Maximal Strength and Fear-Avoidance Beliefs related to Physical Activity. ROB2: Risk of Bias Tool 2; CI: Confidence Interval; PI: Prediction Interval; LL: Lower Limit; UL: Upper Limit; IV: Inverse Variance; W: Weight; MD: Mean Difference; SMD: Standardized Mean Difference; BST: Biering-Sørensen Test; MIBE: Maximal Isometric Back Extension; PT: Peak Torque; n: newton; nm: newton meters; RTG: Resistance Training Group; UEG: Unloaded Exercise group.

Across studies, both RTG and UEG were associated with reductions in pain intensity and disability, but between-group differences were generally modest and inconsistent across follow-ups. At ETP follow-up, the RTG showed a small but statistically significant advantage over the UEG in improving pain intensity (MD = –0.52/10; 95%CI: [–0.92, –0.08]). However, the magnitude of this effect did not exceed the MCID (125), questioning its clinical relevance. For disability, only the STP meta-analysis indicated superiority of the RTG (SMD = −2.04; 95%CI: [−3.92, −0.16]), though this estimate was accompanied by very high inconsistency and was strongly influenced by small-sample trials. Importantly, the 95% PI for both outcomes was wide at every follow-up, suggesting that in future comparable settings, true effects may range from clinically meaningful benefit to no added value of external load. Taken together, these findings indicate that even when significant, the average between-group differences remain small, and the variability across contexts makes strong generalizations premature.

For secondary outcomes, signals in favor of RTG were observed for physical performance measures, although these results should be interpreted cautiously. Back muscle endurance improved more consistently with RTG, with significant effects reported at both STP (MD = 0.50 min; 95%CI: [0.26, 0.74]) and ETP (MD = 0.47 min; 95%CI: [0.09, 0.85]).

The very wide 95%PI calculated for these analyses largely reflects the small number of contributing RCTs. At PW, one trial (54) showed maintenance of these benefits (MD = 0.35 min; 95%CI: [0.03, 0.67]) one year after baseline. For maximal strength, the effects of RT became significant only at PW follow-up (SMD = 0.40; 95%CI: [0.08, 0.71]), potentially reflecting a delayed recovery of physical capacity consistent with the Fitness-Fatigue model (127). Yet, for both ETP and PW follow-ups, 95%PI could not be calculated due to the limited number of comparisons.

In contrast, psychosocial variables were rarely reported. Two RCTs assessed the FABQ-PA (57,67) and only one the PCS (57), without showing any clear between-group difference. The underreporting of psychosocial outcomes likely reflects the historical dominance of biomechanical perspectives in LBP research (5,6). This narrow emphasis limits our ability to determine to what extent training variables, such as load, exert effects beyond neuromuscular adaptations.

Interpretations and Clinical Implications

This meta-analysis examined whether adding external load to RT could enhance the effectiveness of exercise-based treatment for chronic NS-LBP, compared to unloaded approaches. While our findings suggest that external loads may promote measurable neuromuscular adaptations, these benefits do not consistently translate into superior outcomes in pain or disability, nor into clear advantages on psychosocial variables such as fear-avoidance or catastrophizing.

This dissociation reinforces the notion that improvements in physical capacity alone are not sufficient, on their own, to drive meaningful change in chronic NS-LBP—a condition shaped by complex and interacting biological, psychological, and social processes (5,6). The modest and inconsistent effects observed here are in line with accumulating evidence that challenges the assumption of a linear relationship between biomechanical gain and symptom relief (24,128).

Rather than being determined by tissue status or physical capacity, changes in pain and disability may often reflect modifications in pain-related beliefs, behaviors, and emotional responses (129). Although external load may serve as a means of graded exposure or a catalyst for behavioral re-signification (130,131), current data suggest that its effectiveness is not inherently superior. Simply increasing load, without integrating a therapeutic narrative, may be insufficient to influence the multidimensional experience of pain (132,133). It is probable that its effectiveness depends on being embedded within a broader psychological or contextual framework.

From a clinical perspective, this implies that the value of load should be seen less as a mechanical input, but as a potential behavioral signal—the meaning of which depends on patient interpretation, context, and clinical communication (133). In this sense, the delivery and contextual framing of exercise may be more impactful than its intensity or volume.

Given the current heterogeneity in exercise protocols and outcomes, and the limited added value associated with higher intensities, we believe that providing specific dosage prescriptions (e.g., frequency, sets, repetitions) falls beyond the scope of this meta-analysis. However, our subgroup analyses also suggest that exposure consistency and volume, rather than training intensity, may play a more central role in symptom modulation (134). This supports growing interest in low-intensity, high-frequency strategies, such as “exercise snacks,” which may provide similar benefits while enhancing adherence, reducing perceived threat, and improving safety (135).

Ultimately, load can be a useful tool—but not a universally necessary one. Its clinical relevance depends on how it is integrated into a person-centred, biopsychosocial framework that acknowledges not only tissue adaptation, but also individual experience, meaning, and the therapeutic alliance.

Study limitations

While interpreting the findings of this review, several limitations should be considered.

The search and selection strategy introduced some limitations. Restricting eligibility to studies published in English and Italian may have introduced language bias, potentially excluding relevant evidence in other languages. The use of an RCT filter increased the specificity of the search but may also have reduced its sensitivity, with the risk of omitting relevant trials. Notably, no eligible studies published between 2022 and March 2025 were identified. However, similar or even longer publication gaps had occurred in earlier periods, suggesting that this pattern was more likely related to the specificity of the inclusion criteria. In addition, although independent screening was conducted by two reviewers, inter-rater agreement was not formally calculated, which is acknowledged as a methodological limitation of the study selection process.

The characteristics of the included populations also limit the generalizability of the findings. The age range of included participants (20.2 to 52 years) limits the applicability of these results to younger and older individuals with chronic NS-LBP. In addition, sex distribution was imbalanced, with men comprising the majority of participants (62–64% across groups). This imbalance may in part reflect the specific populations investigated in certain trials rather than systematic recruitment bias: for instance, Helmhout et al. (58) enrolled military personnel, and Nambi et al. (60,61) studied male soccer players—both predominantly male populations. While this contextualizes the skewed sex distribution, it nonetheless constrains the transferability of our findings, particularly to women with chronic NS-LBP.

Considerable clinical and methodological heterogeneity was observed across studies, particularly in terms of exercise type, loading modalities, intensity, progression strategies, and baseline symptom severity. This variability, together with the lack of standardized intervention protocols, may have influenced treatment effects and contributed to statistical inconsistency. In addition, although available data suggest a favorable safety profile for externally loaded RT, with event rates comparable to unloaded exercise and mostly mild, transient adverse events, harm reporting was inconsistent and largely descriptive, preventing reliable estimation of event incidence.

Additionally, psychosocial variables such as fear-avoidance, catastrophizing, and kinesiophobia were rarely reported, limiting insights into the cognitive-affective impact of externally loaded exercise. Lastly, subgroup and sensitivity analyses were conducted post hoc and were not based on predefined hypotheses; thus, their findings should be interpreted with caution. Finally, some deviations from the original protocol, including the omission of planned meta-regression analyses due to insufficient data, limited the possibility of exploring potential dose–response relationships.

Implications for future research

To better support person-centered care, future research should investigate how RT variables interact not only with physical adaptations but also with psychological and behavioral responses. It remains unclear whether higher training loads confer meaningful benefits beyond strength gains, particularly when not embedded within a therapeutic narrative that addresses pain-related fear or perceptions of fragility. Similarly, increased exercise variety or task complexity may act as a form of graded exposure, helping to reshape pain-related beliefs and improve self-efficacy, rather than solely enhancing motor control. Total RT volume—including frequency, intensity, and duration—should be examined in relation to adherence, perceived safety, and emotional responses, not merely physiological outcomes. Future trials should incorporate validated psychosocial variables to clarify how exercise influences the full spectrum of pain experience and whether individualized interventions can optimize outcomes through mechanisms beyond tissue-level adaptations.

Conclusion

This review highlights that both loaded and unloaded resistance training can lead to reductions in pain and disability among individuals with chronic NS-LBP, although the magnitude and consistency of these effects remain modest. Notably, symptom improvement does not appear to be solely dependent on load intensity. While high-load RT (85-100% 1RM) may be safe and effective when appropriately progressed—even in individuals with elevated pain levels—clinical outcomes likely depend more on how exercise is structured, delivered, and interpreted than on mechanical intensity alone. Variables such as total training volume, consistency of exposure, and therapeutic framing may play a more substantial role in driving meaningful improvements. From a biopsychosocial perspective, exercise should not be prescribed merely to restore physical capacity, but also as a behavioral intervention aimed at challenging maladaptive beliefs, reducing fear, and fostering movement confidence. Accordingly, exercise programs should be individualized, progressive, and tailored to the patients’ needs. Clinicians may consider incorporating both internal and external resistance, along with multi-joint exercises and adequate variation, in order to promote neuromuscular adaptation, enhance psychological engagement, and support long-term adherence.

Declaration of generative AI in scientific writing

During the preparation of this manuscript, the authors used ChatGPT (OpenAI, version GPT-4o) for language editing and improving the clarity of some sections. The tool was used solely to improve readability and language clarity. All outputs were reviewed, and the authors take full responsibility for the final content.

Other information

This article includes supplementary material

Corresponding author:

Marco Ranzani

email: m.ranzani@casadicuraigea.it

Disclosures

Conflict of interest: The authors declare that they have no conflicts of interest.

Financial support: No funding was awarded for this project.

Data availability statement: All data extracted and analyzed in this systematic review are provided in the appendices of this manuscript. Their inclusion ensures transparency and facilitates the reproducibility of our findings. Additional information can be obtained by contacting the corresponding author.

Ethical compliance: Ethical approval was not required for this study, as it involved the collection and synthesis of data from previously conducted clinical trials in which informed consent had already been obtained by the original investigators.

Author contribution role: All authors contributed equally to the conceptualization and study methodology. MR and DF played equal roles in the data collection and curation. MR conducted the analyses. MR, DR and AP worked on data interpretation and drafted the manuscript. MR edited the manuscript, and MT reviewed the manuscript.