Studies were eligible when they enrolled runners of any level, including military runners required to run occupationally, and delivered a gait retraining intervention directed at cadence or step rate, footstrike pattern, or impact loading, with or without real-time feedback. Comparator conditions could be no intervention, usual care, an alternative retraining strategy, or another active treatment. Eligible outcomes were pain, self-reported function, injury occurrence or recurrence, or running biomechanics plausibly linked to injury, namely vertical loading rate, tibial acceleration, footstrike angle, cadence, and joint loading. Primary intervention designs of any kind were included, from randomized trials to case series and case reports, because the field still relies heavily on small uncontrolled designs. Reviews, protocols, editorials, conference abstracts, and single-session laboratory experiments on healthy runners without a clinical or follow-up outcome were excluded. The publication window was 2018 to 2026, and running-related injury was defined following the consensus description of pain attributed to running that restricts running for at least seven days or requires medical consultation.( 12)

The eligibility criteria can be restated in PICOS form. Population: runners of any level with or without a running-related injury. Intervention: gait retraining directed at cadence or step rate, footstrike pattern, or impact loading. Comparator: any or none. Outcome: pain, function, injury occurrence or recurrence, or injury-related running biomechanics. Study design: any primary intervention design. This breadth was deliberate, because restricting the review to randomized trials alone would have reduced the evidence to fewer than a dozen studies and would have discarded the clinical series in which much of the current practice is grounded.

Three databases were searched: Scopus, PubMed, and OpenAlex. Scopus was searched first with field-restricted syntax, and the core query was:

`TITLE-ABS-KEY("gait retraining" AND (running OR runner OR runners) AND (injury OR injuries OR pain)) AND PUBYEAR > 2017 AND PUBYEAR < 2027`

A second Scopus query targeted step rate and cadence, `TITLE-ABS-KEY(("step rate" OR cadence) AND (running OR runner) AND (injury OR injuries OR pain))`, and a third targeted footstrike and impact loading, `TITLE-ABS-KEY((footstrike OR "foot strike" OR "impact loading") AND running AND (retraining OR feedback) AND (injury OR pain))`, both with the same publication window. PubMed was searched with the terms `gait retraining AND running AND (injury OR injuries OR pain)` and `(step rate OR cadence OR footstrike) AND running AND (injury OR pain) AND retraining`, and OpenAlex with `gait retraining running injury` and `cadence step rate running injury prevention`. No language filter was applied at the search stage.

Duplicate records were removed automatically, and the remaining records were screened by title and abstract. The 27 records judged potentially eligible were then assessed in detail, using open-access full text where available and the abstract and bibliographic record otherwise.

For each included study the following were extracted: first author and year, country, design, sample size, population, intervention target and dose, comparator, outcomes, and the main result with its effect estimate where reported. Where a study reported pain on several instruments, the primary scale was preferred and the others recorded. Extraction relied on the abstract and bibliographic record for 22 of the 27 studies, because open-access full text was available for only five, and no field was completed without textual evidence. A single reviewer performed extraction, which is reported as a further deviation from the two-reviewer standard.

Randomized trials were appraised with the Cochrane Risk of Bias 2 tool( 13) across the domains of randomization, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. Non-randomized studies were appraised with the Risk of Bias in Non-randomized Studies of Interventions instrument,( 14) and case reports against the Joanna Briggs Institute case report criteria. Judgments were made from the available reports and were deliberately conservative; where a domain could not be resolved from the abstract, the study was not credited with low risk.

The search strings, the databases queried, and the number of records returned by each query were recorded so that the search can be reproduced, and this level of documentation follows the expectations for search reporting rather than a summary statement of databases alone. The choice of 2018 as the lower bound was set by the review window, not by a judgment that earlier work is irrelevant, and the upper bound of 2026 reflects the publication year of the target volume.

A narrative synthesis was planned because the interventions, populations, and outcome instruments were too heterogeneous for a defensible meta-analysis. Pain was measured with the visual analogue scale, the Anterior Knee Pain Scale, the Lower Extremity Functional Scale, and the Numeric Pain Rating Scale, while loading was reported as tibial acceleration, vertical loading rate, or joint stress, on treadmills and overground. Results were grouped by outcome domain, and divergences were examined against design, comparator, dose, and measurement as candidate explanations. No meta-analysis, heterogeneity statistic, vote counting, or certainty-of-evidence rating is reported because these were not computed.

The searches identified 768 records, of which Scopus contributed 402, PubMed 166, and OpenAlex 200. After removing 355 duplicates, 413 records were screened by title and abstract. A total of 386 records were excluded at this stage: 80 were reviews or secondary syntheses, 58 were acute laboratory experiments without a clinical or follow-up outcome, 8 were study protocols, and the remainder were observational studies or interventions other than gait retraining. The remaining 27 reports were sought, 0 could not be retrieved, and 27 were assessed for eligibility. No report was excluded at the full-assessment stage, leaving 27 studies in the review.

The screening decision at title and abstract was conservative, in that any record whose eligibility could not be determined from the abstract was retained rather than discarded, and the final assessment therefore rested on more information than the initial screen. Because the automated deduplication removed records across sources before screening, the screened set of 413 is unique, and the flow of records from identification to inclusion closes arithmetically without manual adjustment.

The 27 included studies comprised 12 randomized or quasi-randomized trials and 15 non-randomized designs, published between 2018 and 2026. Sample sizes ranged from 1 to 320 with a median of about 20, and the largest studies were the 320-runner injury prevention trial and a 220-runner wearable feedback trial. Eight studies addressed patellofemoral pain or patellofemoral loading, five enrolled military populations, and the remainder covered plantar fasciitis, exercise-related lower leg pain, chronic exertional compartment syndrome, chronic knee pain, and injury prevention in novice and recreational runners. Geographically the evidence is concentrated in North America, with 15 studies from the United States and Canada, and the remainder from Brazil, Australia, New Zealand, the United Kingdom, Ireland, the Netherlands, Belgium, Hong Kong, and China. The intervention targets were distributed across footstrike conversion, cadence or step rate increase, and impact loading reduction, and most studies combined a target cue with real-time visual, auditory, or haptic feedback (Table 1).

Two studies deserve individual attention because they define the upper and lower bounds of the evidence. The 320-runner trial by Chan and colleagues is the only large randomized test of injury prevention and remains the reference point for the field, while the 220-runner trial by Van Hooren and colleagues is the largest test of wearable-delivered feedback and its mixed intention-to-treat result illustrates how delivery fidelity can erode an effect that is present among adherent users.(15,16) Between these poles sit the trials of patellofemoral pain, which are small but comparatively well controlled, and the military case series, which are large but uncontrolled.

The three intervention families are not cleanly separated in practice. Several programs combined targets, pairing a footstrike conversion with an instruction to increase step rate, or a cadence cue with a soft-landing instruction, which makes it difficult to attribute effects to a single mechanical variable. Feedback was delivered visually on a monitor, as audio or music, as haptic vibration, or through instrumented insoles and socks, and the dose ranged from a single two-hour session to supervised programs of eight weeks or more. This heterogeneity is itself a finding: the field has not converged on a standard retraining protocol.

No study was rated at low risk of bias overall. Eight studies, all randomized trials, were rated as raising some concerns, and the remaining 19 were rated at high risk, driven by absence of a control group, small samples, short or absent follow-up, and reliance on participant self-report (Table 2). Within the randomized trials, the most common concern was the impossibility of blinding participants to their assigned running pattern, followed by incomplete adherence, which affected the interpretation of the wearable trials in particular. Within the non-randomized studies, the absence of any comparator and the use of pre-post designs meant that observed improvements cannot be separated from natural recovery or regression to the mean. The contrast between the randomized and non-randomized bodies is the central methodological fact of this literature: the trials that can support causal inference are few, while the uncontrolled studies that dominate the field can only describe change.

The evidence on pain converges. In runners with patellofemoral pain, the three techniques tested by dos Santos and colleagues, forefoot landing, a 10% step rate increase, and forward trunk lean, all increased the Anterior Knee Pain Scale at six months, and pain and the Lower Extremity Functional Scale improved regardless of group.(19) A pilot trial found a large between-group difference favouring gait retraining over foot orthoses for the Anterior Knee Pain Scale and worst pain, with a number needed to treat of 2.(18) Two later randomized trials extended this pattern: impact and cadence retraining both reduced running pain more than no intervention at six months, with mean differences of -3.2 (95% CI -5.1 to -1.3) and -2.9 (95% CI -4.8 to -1.0) points,(37) and a telehealth protocol produced a greater pain reduction than standard physical therapy, a mean difference of 2.52 (95% CI 0.91 to 4.12).(39) Feasibility work in the same population reported large reductions in average and worst pain, with effect sizes of 1.7 and 2.0 respectively,(21) and increased joint variability after retraining coincided with clinically meaningful pain relief.(23) An acute crossover study in 68 symptomatic runners added a within-session observation: forefoot striking and a 10% step rate increase produced the most responders, and 42% of runners reported an immediate pain reduction with at least one modification.(34)

The magnitudes deserve emphasis because they are not marginal. A number needed to treat of 2 for a clinically meaningful improvement means that, for every two runners treated with retraining rather than orthoses, one additional runner benefits, a figure that would be regarded as highly efficient in most rehabilitation contexts.(18) Effect sizes for pain in the feasibility work reached 1.7 to 2.0, which exceeds the threshold conventionally considered large, and the between-group pain difference in the telehealth trial, 2.52 points on a 10-point scale, exceeds the minimal clinically important difference for running pain.(21,39) These are treatment effects in the range expected of established active therapies, not subtle signals.

Beyond the knee, the same direction holds. Foot strike retraining in service members with chronic knee pain improved the Lower Extremity Functional Scale by 10 points and the Numeric Pain Rating Scale by 2.0 points at one month,(33) and two patients with patellofemoral chondral defects showed improved pain and function after converting to a non-rearfoot strike.(38) A randomized trial in runners with plantar fasciitis reported large improvements in pain (d = 0.92) and foot function after visual biofeedback toward a softer landing,(41) and contact-time feedback added to home exercises produced greater function and recovery than exercises alone at six weeks.(31) A case report extended the reach of the intervention to chronic exertional compartment syndrome, in which six weeks of forefoot retraining resolved symptoms and reduced compartment pressures enough to avoid fasciotomy.(35) A pilot study in previously injured cadets reported improved function on the Single Assessment Numeric Evaluation and the Patient-Specific Functional Scale after a ten-week program.(29)

Even so, one trial qualifies this convergence. Esculier and colleagues randomized runners with patellofemoral pain to education alone, education plus exercise, or education plus gait retraining and found no group-by-time interaction for symptoms or function, even though gait retraining increased step rate by 7.0% and reduced loading rate by 25.4%.(17) This reading, that the biomechanical change did not translate into additional symptom benefit, contrasts with the trials that used no active comparator and is the clearest instance of a control-group effect operating in this literature.

Taken together, the symptom studies form a convergent picture that spans three continents and at least four diagnostic groups, from patellofemoral pain to plantar fasciitis, lower leg pain, and chronic exertional compartment syndrome. The convergence is notable precisely because the individual designs are weak; a series of uncontrolled studies pointing in the same direction across different injuries carries more weight than any single one would alone, although it does not reach the standard of a pooled estimate with a confidence interval. The one dissenting result, from the education-controlled trial, is best read not as a failure of retraining but as evidence that education is itself an effective intervention against which retraining must demonstrate added value.( 17)

The injury evidence diverges and is thinner. The strongest result comes from a randomized trial of 320 novice runners, in which a two-week soft-landing program reduced 12-month injury occurrence to 16% against 38% in controls, a hazard ratio of 0.38 (95% CI 0.25-0.59).( 15) This is the only large trial to measure injury as a primary outcome, and its effect size is substantial.

This contrasts with two trials in recreational runners. A randomized trial of 220 runners using real-time insole feedback found no difference in injury rate in the intention-to-treat analysis (hazard ratio 1.11), although the as-treated analysis, limited to participants who used the device correctly, showed a lower injury rate (hazard ratio 0.53), and one third of participants were reclassified because they had used incorrect settings.(16) A second trial of 128 rearfoot strikers found similar one-year injury rates between rearfoot (37%) and non-rearfoot runners (30%), even though 80% of participants converted their footstrike after a single training session.(28) That study nonetheless attributed a higher knee injury risk to rearfoot strike, with a relative risk of 5.64 (95% CI 1.90 to 16.8), a claim that rests on a secondary comparison and warrants caution.(28) The pattern is therefore one of a large preventive effect in injury-naive novice runners and no overall effect in mixed recreational cohorts, with adherence emerging as the plausible mediator. No study in this review measured recurrence of a specific injury as a primary outcome, which means the question of recurrence, as distinct from first injury, remains answered only indirectly through the symptom studies.

Biomechanical change is the most consistent outcome, and the magnitude is large where loading was targeted. A controlled pre-post study reduced tibial shock by 32% and vertical loading rates by 25% to 27%, with retention at one year.(20) Haptic feedback reduced resultant tibial acceleration by 50% immediately and 41% at four weeks, with smaller but retained transfers to overground running.(24) Music-based biofeedback reduced peak tibial acceleration by 25.5% without changing cadence,(32) and field-based feedback in high-impact runners reduced axial peak tibial acceleration by 29% in the field and 33% in the laboratory, with the change retained at one month.(40) In service members, footstrike conversion reduced average vertical loading rate by 29%(30) and by 56% in a clinical series,(33) and a structured cadence and soft-landing program increased cadence and shifted a predominantly heel-strike cohort toward a forefoot pattern.(36) A pilot in previously injured cadets reported that all nine participants converted from rearfoot to non-rearfoot strike with reductions in loading rate.(29)

Footstrike and cadence produce different loading responses, and this is where the intervention families diverge. Forefoot strike retraining reduced the vertical average load rate by 49.7% with long-term retention, whereas cadence retraining reduced it by only 16% and only at six months, despite similar increases in cadence in both groups, of 7.2% to 173 steps per minute in the cadence group and 6.1% to 172 steps per minute in the forefoot group.(25) Midfoot strike retraining reduced footstrike angle with a large effect (d = 1.65) but changed vertical loading rates inconsistently, and only 40% of participants adopted the pattern for more than 80% of footfalls.(26) Twelve weeks of forefoot strike retraining reduced peak patellofemoral joint stress by 13.3%,(27) and acute cueing studies confirmed that higher cadence and lower vertical oscillation both reduce loading rates, with lower oscillation reducing peak vertical ground reaction force more and higher cadence reducing braking impulse.(22)

The choice between families therefore carries a measurable consequence. A clinician seeking the largest reduction in loading would select a forefoot strike program, whereas a clinician seeking a smaller but perhaps more tolerable change would select cadence, and the two have not been compared on clinical outcomes in the same trial. This gap is material, because footstrike conversion is more demanding and its uptake varied between studies, with the crossover study reporting the highest responder rate for forefoot striking and the midfoot trial reporting that only 40% of participants adopted the pattern.(26,34) Even so, no study reported serious adverse events from retraining, and the most frequent unintended effect was transient calf soreness after forefoot conversion.( 18)

Across the loading studies, the typical reduction in impact-related variables fell between 25% and 50%, with the largest changes observed for tibial acceleration feedback and the smallest and slowest for cadence-only programs. This graded pattern is itself informative: the intervention that targets the measured variable most directly produced the largest change in that variable, whereas cadence, which acts indirectly through stride length, produced a smaller and later effect. The hierarchy of effect sizes mirrors the hierarchy of mechanistic specificity, and it is consistent with the view that the three families are not equivalent doses of the same treatment.

Retention was the variable that separated the durable programs from the transient ones. The one-year follow-up studies reported that loading reductions persisted after the supervised phase ended,(20) and the footstrike trials reported maintenance of the new pattern at six-month and one-year assessments in most participants.(25,30) By contrast, the cadence effect in one trial appeared only at the six-month assessment, and the single-session footstrike program retained the pattern in only a proportion of runners at one year.(25,28) This distribution suggests that retention is not a property of the target but of the dose, and that programs with several supervised sessions and a gradual withdrawal of feedback outperform one-off instruction.

Several grey zones remain despite the consistency of these mechanical changes. Retention was variable, with some programs maintaining the new pattern at one year and others losing it within weeks, and the dose-response relationship is unresolved, because programs ranged from a single session to twelve weeks and no study compared doses directly. The transfer of a treadmill-learned pattern to overground running was demonstrated in only a few studies, and the long-term durability of the change during unsupervised running is largely unknown. What remains open is whether these mechanical reductions translate into fewer injuries, given that the two trials that measured both found no consistent link.