This is a prospective cohort study with clinical data. The flowchart of this cohort study is presented in figure 1. We collected preoperative data between one and two weeks before TKA surgery. Follow-up measurements were performed three days, two weeks, six weeks and one year after surgery. Data collection was performed as part of routine care. We reported this study in accordance with the STROBE statement for the reporting of observational studies. The medical ethical review board of the Medisch Spectrum Twente (MST), Enschede, The Netherlands approved the study (Kh 13-06). All patients provided written informed consent prior to enrolment in the study.
Participants and setting
Participants were recruited from the MST community hospital in Enschede, the Netherlands between February 2011 and December 2014. Patients who followed a high-intensity physiotherapy program after TKA were included. The decision whether a patient was eligible to participate in the program was made by the orthopaedic surgeon and the physiotherapist together. A requirement was that patients are able to maintain the high-intensity physiotherapy program. Therefore, the following inclusion criteria were mandatory: 1) 18 years or older and diagnosed with primary osteoarthritis; 2) admitted for TKA surgery; 3) preoperatively independent in activities of daily living; 4) no comorbidity that hindered doing exercises; 5) no mental disorders as reported by the patient; 6) physically able and willing to perform a 10-day high-intensity physiotherapy program; and 7) signed informed consent (see Appendix 1). The high-intensity physiotherapy program is explained in Appendix 1.
Surgery and physiotherapy
After inclusion in the study by the orthopaedic surgeon and the physiotherapist, the preoperative assessment took place. Thereafter, all participants received a TKA procedure. Five orthopaedic surgeons performed all TKA surgeries following the same surgical procedure. The number of TKA procedures per year carried out by each surgeon varied from 50 to 70. Participants started their rehabilitation during the three day hospital admission. The orthopaedic surgeon and the physiotherapist checked a second time if inclusion in the high-intensity program was possible (no complications which hindered following a 10-day high-intensity program). After discharge and definitive inclusion, participants stayed at a resort where they followed a 10-day high-intensity physiotherapy program, which is explained in Appendix 1. After ten days all participants returned home. The program was available for all patients, independent of their social and economic status. Patients were advised about the continuation of physiotherapy after the high-intensity program, dependent on their physical status. 
The following variables were collected pre-operatively: patient characteristics (i.e. age, gender, body mass index [BMI], number of comorbidities), performance-based measure (i.e. Timed Up and Go [TUG] and self-reported measure (i.e., physical functioning measured with the Knee Osteoarthritis and Outcome Score Activities of Daily Living scale [KOOS-ADL], and pain measured with the Visual Analogue Scale [VAS]).
Physical functioning and pain are core outcome measures for people undergoing TKA surgery, which was confirmed by patients and orthopaedic surgeons. Mizner et al advised using both performance-based measure and patient-reported measure for measuring physical functioning. Therefore, performance-based physical functioning (TUG), self-reported physical functioning (KOOS-ADL) and pain (VAS) were used as outcome measures. The recovery trajectories were based on measurements preoperatively, and three days, two weeks and six weeks after surgery. All outcome measures were also measured at one year after surgery.
The TUG is a multi-activity measure. Patients were asked to stand up from an armchair (with a seat height of 46 cm), walk three metres, turn and walk back to the armchair without assistance. For scoring the TUG, the time in seconds to complete the task was measured. The instructions were to walk safely, but as fast as possible. The test was assessed twice and the lowest time score was used as outcome measure.
Self-reported physical functioning was measured with the KOOS. The KOOS is composed of five separately scored subscales: pain, symptoms, ADL, activities in sports and recreation, and knee-related quality of life.[28,29] Answers were given using a Likert scale, and each question was answered with a score from 0 to 4. A normalised score from 0 to 100 was calculated for each domain (100 indicates no symptoms/pain, and 0 indicates extreme symptoms/pain). The KOOS has excellent reliability and good content and construct validity when used for short- and long-term follow-up of knee injury.[30,31] It has been validated for people with TKA.[30-32] The score per subscale was determined and only the subscale ADL was used to measure outcome.
The VAS was used to measure pain. Patients were asked to mark on a 100-millimeter line their pain rating during the last week, where 0 corresponded to no pain and 100 corresponded to worst imaginable pain.
Patient baseline and preoperative characteristics were described as median [interquartile range (IQR)] or number of patients (percentage). The subpopulations of patients based on the postoperative performance-based physical functioning and self-reported physical functioning outcome trajectories and pain outcome trajectories during six weeks were identified using latent class mixed model (LCMM). The LCMM finds potential latent profiles in heterogeneous populations. It combines a latent class model to identify homogenous latent classes of subjects and a mixed model to describe the mean trajectory over time in each latent group, while taking into account the individual correlation between repeated measures. Each subpopulation has its own physical functioning (TUG, KOOS-ADL) or pain (VAS) growth parameters. We fitted the models using not only linear functions of time but also quadratic functions to allow nonlinear mean trajectories over time. The optimal number of classes was determined using a forward procedure, starting with one class and no subpopulations in the study sample. Then one class was added for each model. To evaluate if the model with one added class improved, three steps were taken: 1) The Bayesian information criterion (BIC) was used.[10,13, 33] The BIC considered the likelihood of the model and the number of parameters in the model. A lower BIC value indicates a better model fit and is a guidance to decide the optimal number of classes;[10, 13, 33] 2) Patients were assigned to their most likely class based on a posterior probability of ≥0.7.
The association of the identified groups with the one year outcomes was analysed using multivariable regression analysis. 191 patients scored a VAS 0 (no pain) after one year, so we decided to dichotomize the VAS score into ‘no/very low pain’ and ‘pain’. Patients with a VAS score of 0-20 were categorized as ‘no/very little pain’ and patients with a VAS of ≥21 were categorized as ‘pain’. This cut-off point was also used in other studies using the patient acceptable symptoms state (PASS).[36,37] Dichotomizing was only done for the one year VAS scores.
The one year responses were regressed on the identified subgroups and on a set of relevant baseline covariates: Age, gender and BMI. Linear regression models were used for the one year TUG and KOOS scores, and a logistic regression model for the one-year dichotomized VAS outcome. The association was determined with the regression coefficient and the 95% confidence interval (linear regression models) for TUG and KOOS and the odds ratio and the 95% confidence interval (logistic regression model) for the VAS of pain during the last week. The overall fit of the models were assessed using the total variance explained, the for the linear regression models and Nagelkerkes for the logistic regression model.
Statistical analysis was performed with R software version 3.4.4 and IBM Statistical Package for the Social Sciences (SPSS 25.0) The ‘lcmm’ R package was used to perform the latent class analysis.