Osteoarthritis (OA) is characterized by pathological changes of the joint structure and is accompanied by pain and functional limitations. The global burden of hip and knee OA was ranked as the 11th highest contributor to worldwide disability [1]. The prevalence of OA increases with age. In Germany, almost 50% of all women and 30% of all men suffer from OA [2].

The promotion of physical activity in general, and exercise in particular, is highly relevant for patients with OA, as it has been shown to decrease pain and improve physical functioning [3]. Furthermore, physical inactivity is a particular characteristic of critical OA patients, with higher all-cause mortality compared to that of the general population, and is even more pronounced in people with severe walking disabilities [4]. Accordingly, international and national guidelines strongly recommend physical exercise as a core treatment in the management of hip and knee OA [5-8]. However, only a small to moderate proportion of patients with knee and hip OA meet physical activity guidelines [9]. An essential barrier for regular physical exercise could be that people with knee or hip OA suffer from pain at the weight-bearing joints and may further believe that physical activity is not beneficial or is even harmful to their joint [10]. Accordingly, OA patients experience fear of worsening their symptoms through exercise or by executing the exercises incorrectly [11,12].

Supervision seems to be an effective means to promote safe and correct exercise techniques, especially in the initial stages of the disease, and to ensure the right exercise dosage for each individual according to their physical ability and dosage principles from a training science perspective [13,14]. It is recommended that initial supervision and knowledge transfer on activity pacing be guided by a health care provider [7]. However, there is a considerable discrepancy between activity-related recommendations and practice in health care: 63% of patients with hip or knee OA aged 60 years or older that were part of a large German statutory health insurance company were treated with medication, whereas only 41% received physical therapy, including physiotherapy and exercise therapy [15]. Therefore, exercise-related advice from health care providers to community-dwelling people is not yet satisfactory. This is especially true for information related to modifying exercise behavior [16,17].

As a consequence, exploring alternate cost-efficient forms of delivery modes for people with limited access to therapeutic services appears to be indicated [13]. In this context, digital apps could be particularly suitable, since geographical proximity and time synchronicity between a therapist and patient are no longer necessary requirements for therapeutic interventions.

Digital interventions offer various types of interfaces and degrees of supervision. First, blended physiotherapy partially replaces face-to-face appointments with a digital app comprising instructions for graded activity, complementary unsupervised exercises, and behavior-change techniques. The physiotherapist plays an active role in patient interaction with regard to exercise content, as well as in supervision of intervention progress [18]. Second, the intervention can be delivered without the local presence of a physiotherapist, but with permission for regular Skype video sessions and individual support of the home training program in terms of exercise selection, instruction, and dosage adjustment. Exercise videos are also provided [19]. Other interventions still build upon personal supervision, yet without face-to-face contact. Web-based intervention programs are used that provide information, exercise instructions, an online physiotherapist (synchronous and asynchronous chats, and telephone consultations), and education regarding factors of relevance to OA, including lifestyle [20,21]. Third, there are fully automated apps that do not involve any interaction with a real person [22,23]. In the complete absence of personal supervision, a digital app should support the patients’ tasks required for the correct and safe execution of exercises, as well as adequately control physical load and handling of pain as well as possible.

In recent years, a large number of digital exercise interventions have emerged, but there is a lack of studies comparing apps or digital interventions with human-delivered approaches in the field of chronic diseases. Hsu et al [24] compared a person-to-person and a digital-assisted approach for successful aging in older people. The results indicated that healthy behavior improved for the person-to-person group, although not significantly. However, the ability to search for health information improved in the digital-assisted group. Some other studies compared guided (email or telephone) and unguided mental health interventions. Guided interventions were significantly superior to unguided interventions [25]. However, none of the above-mentioned studies included patients with OA. Furthermore, systematic reviews have shown that the theoretical foundation of digital interventions is often not very well-developed [26].

In the present study, we evaluated a fully automated digital app that was developed on the basis of a paper-based home exercise program, which emerged as one of the important features of an evidence-based exercise intervention for patients with hip OA [27,28]. The app was designed to be used without personal supervision and comprises a selection of videos with hip-specific exercises, information on dosage principles, as well as individualized feedback mechanisms (for further details see the Methods section).

With regard to the effectiveness of exercise interventions in general, and for digital interventions in particular, a theory-based approach has been shown to optimize the targeted effects [29-31]. The digital app examined in this study was developed as part of the framework of the model of physical activity-related health competence (PAHCO) [32]. This model follows the general objective of exercise interventions for patients with chronic diseases to promote individual competence for an increasingly independent realization of regular physical exercise [33]. From this health educational perspective, it is particularly important to empower the patients to adequately master the requirements of specific exercises and to enable them to carry out physical exercise in the best possible health-effective and low-risk manner, taking into account their own physical condition.

The PAHCO model considers three subcompetences, each of which specifically helps in coping with demands that arise during the initiation and maintenance of physical exercise in a health-enhancing way. Movement competence relates to motor demands and is observable based on high movement quality while performing physical exercises. Movement competence is mainly based on motor abilities and skills but also requires the confidence to accomplish the task with one’s own abilities (corresponding to task self-efficacy [34]). Control competence enables people to gear their own physical exercise to optimize health benefits and minimize health risks. It is mainly based on action-related knowledge but also requires the ability to perceive and interpret body signals (eg, to adjust intensities based on muscle soreness). Self-regulation competence ensures the required regularity of physical exercise. Thus, in theoretical terms, this subcompetence is closely related to social-cognitive theories of health behavior that address motivational and volitional determinants of exercise behavior (eg, outcome expectancies, behavioral self-efficacy; for an overview see Biddle et al [35]). The subcompetences can be considered as proxies for regular physical exercise, which in turn leads to the desired health benefits [32]. The promotion of these subcompetences should therefore be an important aim of an exercise intervention. Although pain, physical functioning, fear of movement, and self-efficacy have already been the focus of previous studies on the effectiveness of digitally assisted training interventions [23,36,37], there is a lack of studies related to proxies of health outcomes such as movement quality, control competence, and self-regulation.

Based on the aforementioned knowledge deficits, the aim of this crossover study was to evaluate whether exercise instruction and guidance of one training session via a fully automated tablet computer-based app results in comparable benefits in subcompetences of PAHCO in comparison to one session supervised by a physiotherapist in subjects with hip OA. Our hypotheses were as follows: (1) movement quality of exercises guided by the app is not inferior to the movement quality under supervision by a physiotherapist; (2) the effect of the app on control competence is not inferior to the effect of the physiotherapist-guided intervention; and (3) the effects of the app on exercise self-efficacy (as a prerequisite of self-regulation competence) are not inferior to the effects of the physiotherapist-guided intervention.

This study was designed as a randomized 2×2 crossover trial. The participants were randomly assigned to one of two exercise treatment sequences with an allocation ratio of 1:1. The AP sequence started with a training session instructed by a fully automated tablet computer-based digital app, followed by an intervention supervised by a physiotherapist, whereas the PA sequence started with the physiotherapist and the second training intervention was conducted with the app. The washout phase between the two interventions was set to range between 3 and 5 weeks. This time period seemed sufficient to washout relevant treatment effects of a single exercise. Ethical approval was obtained from the ethical committee of Tuebingen University Hospital. The study was registered in the German clinical trial register (DRKS00015759).

Community-dwelling individuals with diagnosed hip OA were recruited via advertisements in regional newspapers, as well as by emails sent to employees of the University of Tuebingen and Tuebingen University Hospital. In addition, flyers were distributed by orthopedic surgeons and physiotherapists. Interested participants were asked to call staff members for further information. According to ethical guidelines, the subjects were fully informed about the positive effects of exercise therapy for hip OA. They were also informed about the structure and details of the app (ie, feedback mechanisms) and the research questions. The screening for eligibility took place in the context of this phone call. Eligible people were then randomly allocated to one of the two treatment sequences and the dates for the two treatment sessions were scheduled. The treatment sequence was blinded until the first treatment session. Both treatments took place at Tuebingen University Hospital. Inclusion and exclusion criteria for participants are described in Textbox 1.

The sample size was predefined to a minimum of N=40 without a sample size calculation.

A computer-generated randomized order list in blocks of 5 entries for each of the 2 treatment sequences (AP and PA) was created prior to the start of the study by the study personnel. Eligible participants were randomly assigned to the sequences at the end of the initial phone call. The study staff member entered the name of the caller into the list consecutively in the order of the calls. The list was visible to the study personnel with no further concealment. If randomized subjects canceled the first treatment session, the randomization slot was opened again and used for the first new incoming call of an eligible subject. The order in which raters assessed the movement quality of each participant was randomized for each rater and test day separately using the internet tool random.org [42].

Participants, assessors (raters), and statisticians were not blinded to the treatment sequence or type of intervention. However, participants did not know which treatment sequence they were assigned to before the start of the first training session. The two raters of movement quality were not blinded to treatment sequence or intervention; however, they were not included in any other part of the study, such as data collection or data analysis.

Recruitment started in November 2018. Sixty-eight people contacted the study staff by phone, and 59 were deemed eligible and received a follow-up email, including written study information, details on measuring time points, and travel directions. However, five potential participants cancelled the first training appointment due to physical problems or overlapping appointments. Finally, 54 people completed the first training session. Seven participants could not attend the second session. Further details on participants flow are depicted in Figure 1. Data collection started on December 3, 2018 (first patient in) and was completed on January 16, 2019 (last patient out). The individual time period between T1 and T2 ranged between 27 and 42 days, with an average interval of 34.7 days.

Baseline characteristics are presented in Table 2. Subjects in the PA group had a worse baseline condition for hip-related quality of life compared with that of the AP group. Participants in the PA group also had lower values for overall exercise-specific self-efficacy, as well as the cope and schedule subdimensions, in comparison with those of the AP group at baseline. Other baseline characteristics did not differ between participants allocated to the two treatment sequences.

Digital home training programs that help people support their training routines are urgently needed in the current world of decreasing physical activity. This is particularly true for patients suffering from OA. It is well known that exercises are efficient to decrease pain and increase physical functioning in OA [3,5-8]. To enable patients to train in a health-promoting and low-risk range, the PAHCO model provides a breakdown of subcompetences, each of which helps in the long-term realization of health-effective exercise [32,33]. A main interventional implication of the PAHCO model is to combine exercise practice, transfer of knowledge, and processing of personal experience of movement [40].

To the best of our knowledge, there is no app specifically designed for hip OA patients that combines the transfer of knowledge, exercise instructions, and processing of personal experience with movement. Furthermore, there are no studies comparing the effectiveness of these interventions provided by humans and apps. Therefore, the aim of this study was to investigate whether digital exercise instruction and guidance leads to benefits in PAHCO’s subcompetences in a manner that is comparable to physiotherapist care for patients with hip OA.

Movement quality of exercise execution was one important outcome in our study. Exercising in groups of two, in which the partner is alternately practicing and observing, shows important effects on learning success, especially for practical implementation with patients [44-47]. In our study, the physiotherapist served as an active corrective, but less as a role model, because she did not actively execute the exercises. In contrast, the app provides permanent illustrations of how to conduct the exercises via the virtual training partner and additional acoustic signals. The exercises selected for the study differed in terms of their starting position, used musculature, and complexity. Compared to the other two exercises, the two strength exercises were characterized by a more error-prone starting position and execution of movement, whereas the mobility exercise and balance tasks were simpler. Movement quality was inferior in the training session with the app compared with that guided by the physiotherapist, except for the mobility exercise. The difference between the physiotherapist and app for the balance exercise was smaller than the critical effect size value of 0.2, and the criterion for noninferiority was only missed because the 95% CIs exceeded the noninferiority margin. The effects of exercise for mobility and balance in the intervention with the app were close to those with the physiotherapist. However, both strength exercises showed large effects in favor of the physiotherapist-assisted intervention. It can therefore be concluded that not all exercises with different focuses and difficulties are suitable for being guided by a training app only. To avoid comprehension difficulties and poor movement quality, as well as potentially related impairments of training efficacy and risk of harm, we recommend that patients receive supervised instructions when initiating training, especially for more complex exercises. This is in line with results of other investigations, in which initial face-to-face contact with individuals before a new intervention showed positive changes in physical activity outcomes. This is also recommended in guidelines for the nonpharmacological treatment of OA [7,48].

Self-efficacy is an important key to modifying behavior [37]. People with high self-efficacy set higher goals, invest more effort into the pursuit of their goals, and will be more likely to keep trying when barriers and setbacks arise [49]. Self-efficacy increases when the individual receives positive feedback after successfully completing a task [50]. The results from this study revealed that the app is inferior to a physiotherapist in outcomes related to exercise-specific self-efficacy. In particular, the overall scale for exercise-specific self-efficacy as well as its subscales coping and scheduling showed large effects in favor of the physiotherapist. Coping is related to barriers for carrying out exercises, such as “I am able to exercise when I lack energy” or “…don’t feel well,” and scheduling is related to the ability to integrate exercises into daily routines. It can be concluded that face-to-face supervision of a single training session by a physiotherapist cannot be equally replaced by a digital app with a pedagogical agent and video-based training instructions. This was confirmed by Danbjørg et al [12], who reported that the participants missed personal contact with the physiotherapist. However, it should not be disregarded that self-efficacy increased even after training with the app. Litman et al [51] also reported an improvement in self-efficacy using digital apps, but also regarded this type of intervention as a complementary and supporting measure.

We found carryover effects for all outcomes related to the self-administered exercise-specific self-efficacy score, but only for one movement quality outcome. The subjective confidence of being able to show a desired behavior seems to be influenced by a personal supervisor much more than the objective ability to perform a movement. This finding underlines the importance of using both subjective and objective outcome measures if PAHCO is to be improved by a special intervention.

The effect size for control competence for physical training in our study was smaller than the critical value of 0.2, and the criterion for noninferiority was only missed because of the 95% CI exceeding the noninferiority margin. Control competence for physical training is quantified using items that are directly related to the competence to control training intensity in the required way (ie, “I know how I can best increase my strength in the leg and hip area with physical training” or “I am able to adjust my training effort well to my physical condition”). These items are comparable to the exercise-specific self-efficacy subscale task to some extent, with items such as “…I am certain that I will be able to carry out exercises with the right technique.” The app was inferior to the physiotherapist in this subscale. However, only a small effect size was obtained in comparison to all other exercise-specific self-efficacy scales that showed medium or large effect sizes.

The characteristics of our sample are typical characteristics of OA patients, indicating the good generalizability of the findings. The average age and the gender distribution correspond to the risk profile of the disease [52]. The average Western Ontario and McMaster Universities Osteoarthritis Index (WOMAC) pain score attests a moderate hip OA [53]. However, the fear of movement among our participants was relatively low compared to that of other OA samples [11]. Overall baseline values further indicate that the study sample was physically active according to physical activity guidelines, and had a positive attitude toward technology [54]. These criteria may restrict the external validity of the results to a comparable population with respect to fear of movement, physical activity status, and technical affinity.

No adverse events were reported for any of the interventions. However, safety aspects are extremely relevant for fully automated computer-based interventions as there is no health care professional controlling for nonphysiological or even harmful execution of exercises. The app tested in this study used personalized closed feedback loops to adapt exercise instructions and dosage according to the user’s feedback on pain and physical exhaustion. Nonetheless, it has been shown that movement quality was inferior by using the app when it comes to more complex exercises. Future studies with longer intervention periods should therefore evaluate if minor movement competence goes along with higher pain levels, more adverse events, or poorer health outcomes. Proof of safety and medical benefit or patient-relevant structural and process improvements are mandatory aspects for approval of an “app on prescription” in Germany [55].

Results of this study show ceiling effects for the variables self-efficacy and control competence. The study population already had very good values at baseline, thereby reducing the possibility for change. To be able to assess effects in a more differentiated way, samples of future studies should have lower initial values in this outcome dimension. To investigate outcome effects in a more vulnerable population, further research should also focus on a sample that is less physically active, has more severe symptoms, feels greater barriers to technology, and has a greater fear of movement. This limitation and the above-mentioned restrictions with regard to the external validity of the study results may be caused by a potential recruiting bias, which may have affected the outcomes as well. Subjects were fully informed about the rationale and aim of the study in the context of recruitment and inclusion. It therefore cannot be ruled out that this information may have had an effect on user self-selection and expectation, and may have therefore biased the results.

This study also has some methodological issues that should be discussed. Five outcomes of the study showed carryover effects in favor of the physiotherapist, four of which were related to exercise-specific self-efficacy. Although a washout phase with a minimum of 3 weeks was conducted between training sessions, this phase was not long enough to eliminate positive treatment effects. The session with the physiotherapist induced long-lasting effects that sustained during the washout phase. As a consequence, a sensitivity analysis was conducted and results of this sensitivity analysis led to similar but even more pronounced statements related to the inferiority of the app versus the physiotherapist. We are aware that the practice of analyzing data from the first study period as if it had been obtained from a conventional parallel-group design has been shown to be potentially strongly anticonservative [56]. Yet, no conclusive answer is provided by the literature on how to proceed if analyses yield significant carryover differences. Future studies should use a parallel-group design or significantly extend the washout phase to an adequate length to eliminate carryover [57]. However, the results of our study indicate that the latter seems to be difficult to implement in studies with exercise interventions. A complete washout was also not desired in our study.

Exercise-specific self-efficacy and control competence for physical training were assessed prior to and after each intervention, and change from baseline was used as a dependent variable for each intervention. There is little agreement in the current literature as to whether or how to introduce period-specific baseline measurements into the model. The use of change from baseline is discouraged due to its poor type I error rate control and lower power than other methods, yet under the assumption of no carryover [58]. Nevertheless, change from baseline was applied in our study, as it is the most simplistic way to analyze the data, and other methods face similar problems in the presence of carryover effects [59].

Aside from carryover effects, period effects were present in almost all measures. Effects of differences between pre and post values of the training session were larger in period 1, whereas movement quality assessed while exercising was better in period 2. This seems conclusive, as the possibility of improvement within one training session may have a saturation effect partially due to ceiling effects, whereas the better movement quality in period 2 may be related to a learning effect on exercise execution from period 1. As period effects do not affect comparison between groups, this limitation does not seem crucial for the interpretation of the results.

Despite the absence of noninferiority to the physiotherapist in almost all measures of interest, exercise-specific self-efficacy and control competence for physical training also improve using an app, and movement quality is acceptable for exercises that are easy to perform. However, relevant differences in movement quality are present in challenging tasks. The digital app therefore opens up possibilities to take on the role of a supplementary tool to support the patient in independent home training for less complex exercises. Nevertheless, it cannot replace a physiotherapist with an equivalent effect.