Participants

Individuals with knee OA were recruited from the community using paper-based and online advertisements between October 2018 and October 2021 (including a required research shutdown due to the COVID-19 pandemic). Eligible participants had to be 50 years of age or above and have radiographically confirmed OA predominantly in the medial tibiofemoral compartment assessed using the Kellgren and Lawrence (KL) classification scale [21] from coronal plane radiographs of the tibiofemoral joint obtained during upright standing. Participants also had to demonstrate the following inclusion criteria: greater joint space narrowing in the medial compartment than lateral compartment; a history of knee pain longer than six months; and a minimum average knee pain of 3 out of 10 (0 = “no pain”; 10 = “worst pain imaginable”) in the 1-month period preceding study participation. Any individual with any history of the following were excluded from study participation: lower-limb surgery or joint injections in the preceding 6-months; any injury or dysfunction that impaired standing balance or walking ability in the 12-months preceding study participation; and consistent use of orthotic insoles in the 12-months preceding study participation. All participants provided written informed consent, and the study was approved by the institution’s research ethics board for clinical studies.

Insoles

Eligible participants were referred to a Certified Canadian Pedorthist for final confirmation of study eligibility and to undergo 3D laser volumetric casting of their feet, taken in a non-weightbearing subtalar neutral position. Three pairs of sulcus length orthotic insoles were custom-fabricated for each participant, and finished with an identical neoprene cover. Neutral 3 mm flat control (FLAT) and 5° lateral wedge (WEDG) insoles were made from ethyl-vinyl acetate foam (EVA) (Shore A hardness = 55). One pair of variable stiffness custom contoured arch supports incorporating a lateral wedge (WEDG + V-ARCH) were formed from the volumetric casts using plastazote foam laterally (Shore A hardness 70) and EVA medially (Shore A hardness 20). Every pair of insoles was sent directly to the University; upon receipt, participants were invited to the laboratory for a single testing session.

Data collection

The index side was defined as the one with knee pain and radiographic evidence of osteoarthritis in the case of unilateral symptoms, or the more painful knee in cases of bilateral symptoms and radiographic evidence. Participants first completed questionnaires for the index limb, including a numerical rating scale of average knee pain over the previous week (numerical rating scale (NRS) pain: 0 = “no pain”; 10 = “worst pain imaginable”), as well as the Foot Function Index (FFI) questionnaire (revised, short form) [22]. For the purposes of this study, we only used the foot pain subscale (FFI pain), and all scores were converted to a percentage score (25% = least pain; 100% = most pain), as per guidelines. A number of anthropometric measurements were taken, including height and body mass.

We also measured a number of outcomes shown to be correlated with KAM magnitudes (see Fig. 1):

i)

The six-item Foot Posture Index (FPI) assessment was conducted by a trained assessor to provide a numerical rating of the foot posture of the index foot for each participant [23]. The numerical sum of the six FPI items for the for each participant was used for this study.

ii)

Passive subtalar joint eversion range of motion of the index foot was measured with manual goniometry, which has previously demonstrated acceptable relative intra-rater measurement reliability [24]. Measurements were taken with participants resting in a prone position with the foot and ankle hanging off the end of a plinth, where the calcaneus was passively moved by the assessor until a firm end feel was detected. Peak range was recorded, and the average of three measurements was calculated.

iii)

Frontal plane tibial angle of the index limb was measured during relaxed standing via smartphone inclinometry supplemented with an external alignment device. This technique demonstrates adequate measurement validity compared to motion capture, as well as excellent inter-rater and inter-session measurement reliability [25]. Participants were barefoot and stood with knees extended, feet facing forward, and weight comfortably distributed between both legs. The alignment device was aligned with the centre of the tibial tuberosity and neck of the talus before reading the smartphone inclinometer value. Participants briefly marched on the spot between each measurement to reset their standing position, and the mean of three measurements was calculated.

iv)

Gait speed was measured using photoelectric timing gates positioned 4 m apart in the middle of a 10 m walkway. Participants completed three passes along the full length of the walkway, and the mean speed across all trials was calculated.

v)

Foot progression angle was measured as participants walked across a 3 m length of medical exam paper while donning wet nylon socks that left an imprint of the paper. Foot progression angle was measured as the angle between the foot axis (heel centroid to tip of 2nd toe) and a line corresponding to walking direction. The mean of three trials was calculated.

Fig. 1

Visualizations of the collection of the data using clinically-accessible methods: a Ankle/subtalar eversion range of motion was measured using goniometry; b Frontal plane tibial angle was measured using a smartphone inclinometer with an external alignment device collinear to the long axis of the smartphone (left). The participant position during measurement of frontal plane tibial angle is shown on the right; c Foot progression angle was measured using a wet sock as the participant walked across a length of exam paper

Participants then underwent 3D gait analysis using motion capture technology. Forty-seven retroreflective markers were affixed to the skin over anatomical landmarks on the pelvis and lower body, including: the sacrum, and bilaterally over the anterior and posterior superior iliac spines, lateral femoral epicondyles, lateral malleoli, posterior aspects of the calcanei, and heads of the second metatarsals. Tracking markers were placed bilaterally on the lateral thighs and shanks as rigid plates (four markers each), bilaterally on the anterior thighs and shanks, as well as on either side of the posterior calcanei markers forming a heel triad. Ten of the forty-seven markers were affixed bilaterally only during a static calibration trial, including: greater trochanters, medial femoral epicondyles, medial malleoli, and first and fifth metatarsal heads.

Participants completed walking trials across a 10 m walkway, with kinematic data collected at 100 Hz from 14 cameras (Motion Analysis Corp., Santa Rosa, CA) synchronized with two floor-embedded force platforms (AMTI Inc., Watertown, MA) sampling at 2000 Hz. All gait testing always occurred with the FLAT condition first, with the order of the other two conditions (WEDG or WEDG + V-ARCH) systematically randomized following a Williams design (2 × 2; AB BA) to minimize any condition ordering effect. The mean of five successful walking trials represented the laboratory-derived gait data for each insole condition. A walking trial was deemed successful only when index foot completed the stance phase entirely within the boundaries of the force platform, and the walking speed was maintained within ± 5% of the self-selected walking speed established during walking trials with FLAT.

Data processing

Using a previously-published biomechanical model [26] inverse dynamics calculations combined segmental kinematic and force data to calculate 3D joint forces and moments using commercially available software (Visual 3D; C-motion, Rockville, MD). The KAM impulse (Nm/kg*sec) was selected as the outcome of interest for defining biomechanical response as it is representative of the cumulative loading across the duration of the stance phase of gait and can be obtained with acceptable test–retest measurement reliability [27]. We also identified the following 3D gait values from the FLAT trials to serve as predictor variables: rearfoot excursion (°, frontal plane range between initial contact and the time of peak eversion), frontal plane knee angle (°, mean value between 25 and 50% of stance), gait speed (m/sec), and foot progression angle (°, mean value between 15 and 50% of stance).

Statistical analysis

The primary objective of the study was to identify biomechanical responders to LWI use, and this was achieved using logistic regression. The dependent variable used in all logistic regression modelling was the binary biomechanical responder classification and for each LWI condition (WEDG and WEDG + V-ARCH) separately. We created regression models for a variety of KAM impulse response thresholds (decreases of 2%, 6%, or 10% compared to the FLAT condition) using both Akaike Information Criterion (AIC) and forced-entry modelling (see Supplementary File 1 for description). The 2% threshold corresponds with the threshold used by Felson et al. [17] shown to be associated with a greater likelihood of pain improvement, while the 6% threshold approximates the average KAM reduction shown in meta-analyses [5, 6]. The additional 10% threshold was chosen as an incremental increase from 2 and 6%, such that anyone predicted to be a responder at this threshold would be likely to experience a greater reduction in the KAM with LWIs.

Two separate model lists were created that included demographic, anthropometric, and disease-related outcome predictor variables and outcomes derived either using clinically-accessible methods or laboratory-derived 3D motion capture technology. A summary of these models can be found in Table 1.

Table 1 Predictor variables separated by category of participant information. Note that 2 separate groups of models were used: Demographic, anthropometric, and knee OA descriptor variables were used in all models, while the gait, posture, and movement characteristics were separated by whether they were derived using clinically accessible or laboratory-derived approaches. Note that the laboratory-derived data were from the FLAT condition trials

In the initial step of the analysis, predictor variables were checked for collinearity with Spearman rank correlations. Since a 1.0 m/s unit change in gait speed is larger than can be meaningfully interpreted as a predictor variable, clinically-accessible and laboratory-derived gait speed values were multiplied by 10 for use in the prediction models. By doing so, the scale of interpreting gait speed from prediction model outputs was improved, such that a single unit difference in gait speed represented a change of 1.0 dm/s (10 cm/s), rather than 1.0 m/s. Univariable logistic regression models were fit and the odds ratios for each predictor variable were explored for the directionality and strength of relationship between the predictor variable and biomechanical response.

Multivariable logistic regression models were fit using two levels of variable selection. At the first level, a forward selection and backwards elimination stepwise process selected a set of possible predictor variables from each pool of clinically-accessible or laboratory-derived predictor variable inputs at α = 0.30. This level of variable pre-selection eliminated any predictor variable that was unlikely to remain as a significant predictor in the final model, and also determined the order of predictor variable entry for the next level of variable selection. The second level of variable selection used the AIC approach to determine the final set of variables in each model. Predictor variables were entered into the logistic model in a stepwise fashion until the addition of a predictor variable increased the AIC from the previous step. Only predictor variables that were entered into the model before the AIC increased were included in the final prediction model.

The omnibus effect of each model was determined to be significant if the model likelihood ratio was significant and the Hosmer and Lemeshow goodness of fit test statistic was non-significant at p > 0.05. The predictive utility of each model was assessed via the area under the curve (AUC) of the receiver operating characteristic (ROC) curve, c, along with its 95% confidence limits. The odds ratio for each predictor variable indicated the factor by which the odds of being classified as a biomechanical responder changed per unit increase in the predictor variable. An odds ratio > 1.0 represented a greater odds of a biomechanical responder classification, and < 1.0 represented a lower odds of biomechanical responder classification in each particular combination of response threshold and LWI condition. A predictor variable was considered significant in the final AIC-selected model if the p-value for its odds ratio was p < 0.05. Predictor variables with an odds ratio that had a p-value 0.05 < p < 0.10 were considered predictor variables of interest. All analyses were performed using the statistical software SAS v9.4 (SAS Institute, Cary, NC).