Survival analysis for predicting fitness app user churn

Introduction

The popularity of fitness applications is on the rise due to the widespread access to smartphones, internet connectivity, and mobile applications (1). When downloading fitness applications, their motivations vary, ranging from weight loss to muscle gain (2,3). Therefore, fitness applications effectively address various motivations and concerns, providing users with the necessary motivation to exercise while offering valuable information such as workout progress and feedback on exercise performance (1). Fitness applications can be conveniently accessed, and there is an industry demand to predict users churn in order to enhance user retention. In this line, the most effective way to understand user churn is to understand the causes of stopping using the application, including, for example, the demographic features of the users.

While previous studies have explored retention in fitness centres and digital health platforms, most studies relied on static models, such as logistic regressions or descriptive statistics, offering limited insight into how user engagement behaves over time. Other studies have shown that demographic variables (i.e., gender or age) along with app usage features can influence user retention and satisfaction in fitness apps (4,5). Similarly, specific app features, including notifications, personalized plans or goal settings, play a central role in long-term engagement, particularly among sedentary users (6,7).

Research on user churn has been conducted in multiple fields of studies, such as medicine, finance, or retail (8). While research on fitness center churn has grown, relatively little attention has been directed toward understanding user churn in digital fitness applications, which are increasingly popular for supporting health and fitness goals. This study addresses this gap by analyzing user churn in the fitness app Mammoth Hunters using survival analysis techniques to identify factors contributing to user dropout, thereby informing strategies to enhance user engagement and retention. In particular, we aim to examine not just who churns, but when users are most likely to disengage, using survival methods and incorporating a range of covariates. Previous studies (9), have begun to position retention within the fitness industry, yet few have applied advanced statistical methods to the specific domain of mobile fitness apps. Additionally, broader literature on sports and exercise (10) offers insights into motivations and drop-out behaviors in various physical activity settings, underscoring a need for nuanced approaches to retention across platforms. However, the research on survival analysis in the fitness applications or gaming industry (11,12) remains underexplored. Thus, survival analysis applied to the prediction of users’ churn in fitness applications can be valuable for improving user engagement and retention in similar fitness applications (13).

The research data for this study is drawn from the Mammoth Hunters application, a smartphone-based fitness app designed to provide users with accessible, personalized workout programs. The app was chosen for this study due to its unique combination of customizable training plans, real-time feedback, and progression tracking, which are features often cited as key motivators for long-term user engagement in fitness applications.

Mammoth Hunters allows users to receive tailored workout plans based on individual fitness levels, preferences, and goals, adapting to factors such as experience, physical condition, and objectives (e.g., strength, endurance, or weight loss). These features align with those of other popular fitness apps, which also offer customizable plans and instructional content. However, Mammoth Hunters distinguished itself by incorporating instructional videos that provide step-by-step guidance on exercise techniques, along with real-time feedback that allows users to track their progress and adjust their routines accordingly. This approach fosters both user engagement and a sense of achievement, as users can see improvements over time, an aspect commonly highlighted in user retention studies on fitness apps.

This study aims to predict when users might drop out of an application, evaluate their adherence, and identify user groups likely to remain engaged over time. To achieve this, survival analysis methods (techniques traditionally used in medical research) were employed, with a focus on user censoring to distinguish adherent users—those who remain engaged without churning. Censoring, in this context, refers to users whose dropout event has not occurred by the study’s end. Thus, we aim to apply survival and cure models to digital fitness context. This approach enables more accurate estimation of engagement over time, and demonstrates how survival analysis can capture dynamic user behaviours in app-based environments. The analysis includes Kaplan-Meier estimates to visualize retention trends, parametric survival models for detailed predictions, and cure models (both mixture and non-mixture) to account for long-term user adherence. The study hypothesizes that applying these survival analysis techniques can identify key factors influencing user churn and offer actionable insights for enhancing user retention strategies. By uncovering patterns in user behavior, this research aims to inform strategies that foster sustained engagement and application success.

Methods

Study group

The study group comprised 3,034 users of the fitness application Mammoth Hunters who used the application in the time period from 01/11/2021 to the moment of closing the application—27/05/2022. This study was approved by the Ethics and Research Committee of the Faculty of Psychology, Education Sciences and Sport Blanquerna of the Ramon Llull University (No. 2223013D). Participants gave informed consent to participate in the study. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. Prior to their involvement, they were presented with a comprehensive document outlining all aspects of the informed consent. To signify their agreement, participants acknowledged and accepted the terms by clicking at the end of the document. The data were accessed for research purposes on 25/05/2022.

The database contained demographic data of the users. There were 3 time-range approaches and 2 censoring approaches for every time approach. First time approach (T1) was the first 4 weeks of workouts for every user. Second time approach (T2) was the first 8 weeks of workouts for every user. Third time approach (T3) was setting a time interval from 1/11/2021 to the end of the 8th week of 2022.

There is also a need to define censoring in the survival analysis. In traditional survival analysis, typically employed in medical contexts, the occurrence of an event (such as death) and censoring of data are essential components. Censoring occurs when the event of interest fails to manifest within the follow-up period, leaving the outcome beyond that point uncertain.

In terms of fitness application data analysis, the event of interest is user drop out. The time and censoring approaches are shown in Figure 1. The first censoring approach, applied to all time approaches (T1C1, T2C1, T3C1) defines the event as a dropout when the user stopped using the app for even 1 week. This means the user is required to complete at least one workout per week, to be considered still adherent with the app. In the T1C2 approach, the user was censored if they had the application installed for at least 4 weeks, completed at least 2 workouts during this 4-week time period, and worked out during the 4th week. In the T2C2 approach, the user was censored if they had the application installed for at least 8 weeks, completed at least 6 workouts within that time, and workout during the final (8^th) week. On the T3C2 approach, the user was censored if they worked out during the 8th week of 2022 and completed workouts during at least 75% of the weeks they had the application installed. If a user did not meet at least one of the conditions in a given censoring approach, the event was considered to have occurred, and the user is categorized as having dropped out.

Figure 1 Schema of the time and censoring approaches. T1C1/T2C1/T3C1, first censoring approach for Time 1, 2, or 3; T1C2/T2C2/T3C2, second censoring approach for Time 1, 2, or 3.

Upon registration, the following self-reported demographic variables were collected: gender, age, height, weight, activity level, goal, body type, body fat, newsletter subscription, notification setting, training days setting, and body mass index (BMI).

Statistical analysis and algorithms

A statistical analysis was performed using the R Project for Statistical Computing, version R 4.1.1, and Python version 3.10, with the usage of Jupyter Notebook. Categorical variables were expressed as frequencies and percentages, while basic descriptive statistics [minimum, maximum, interquartile range (IQR), median, mean, standard deviation (SD), kurtosis and skewness] summarized continuous variables. Normality of distributions was assessed with the Shapiro-Wilk test. Group differences in categorical data were examined using the χ² test and Fisher’s exact test. The Mann-Whitney U test was applied to compare non-normally distributed continuous variables. For multiple comparisons, the Kruskal-Wallis test and the Dunn’s post hoc test with Benjamini-Hochberg correction was employed. Statistical significance was set at P<0.05. The packages that were used for the survival analysis were lifelines (14) (for Python) and cuRe (15) (for R). Kaplan-Meier estimates (curves) were used to assess the probability of user survival over time. For comparison of survival curves, the log rank test was used. Models taken into account were parametric survival models (created in Python) and cure fraction models (created in R). All of the data variations were split into train (75%) and test (25%) sets.

Kaplan-Meier estimates

The Kaplan-Meier estimator is commonly used in research because it is stable and does not rely on parametric assumptions when estimating and plotting survival probabilities over time (16). First, the event of interest must be clearly defined and occur at a specific time point. Second, the survival probability of all observations is assumed to be independent of when they entered the study. Finally, censored observations are expected to have similar survival prospects to those who remain under observations.

With the Kaplan-Meier estimator it is possible to approximate survival function from the collected data. It is given by formula:

S^(t)= ∏i:ti≤t(1−dini)[1]

Where t_iis the time when at least one event happened, d_i—number of events that happened at time t_iand n_i—number of objects that had not experienced the event.

Parametric models

To calculate the survival parametric models [considered as accelerated failure time (AFT) models] we had three distributions taken into consideration: Weibull, LogNormal and LogLogistic. These models are continuous-time models, in that they assume continuous parametric distribution for the probability of failure over time. The general survival function can be described as:

S(t)=1−F(t)[2]

where F(t) is the event distribution. For the considered models the survival functions look like:

Weibull distribution:

S(t)=exp(−θ1tθ2)[3]

LogNormal distribution:

S(t)=1−Φ(log(t)−θ1θ2)[4]

Where Φ(⋅) is a normal distribution function,

LogLogistic distribution (7):

S(t)=11+(t/θ1)θ2[5]

In all of the cases θ₁, θ₂ >0 and t ≥0.

Cure models

Cure models are survival models that assume a fraction of patients is cured and will not experience the event of interest again. In the context of user workout data, the cure fraction can be defined as individuals who will continue using the application as long as it remains functional. The survival function of the cure model can be represented in two ways—mixture or non-mixture. For mixture cure model (17), the formula is:

S(t|z)=π(z)+[1−π(z)]Su(t|z)[6]

Where Su(t|z) is survival function for the uncured fraction and π(z) is a cure fraction. For non-mixture cure models (17), the survival function has a formula:

S(t|z)=exp{log[π(z)][1−Su(t|z)]}[7]

where again, Su(t|z) is survival function for the uncured fraction and π(z) is a cure fraction. Cure fraction can be defined by different parametric forms, like logistic or loglog (15). They are given by formulas:

π(z)=exp(β′z)1+exp(β′z)[8]

logπ(z)=1−exp(−exp(β′z))[9]

For the survival function of the uncured fraction, it is possible to use parametric form (e.g., Weibull or LogNormal).

Evaluation of models

Model performance has been assessed using Akaike Information Criterion (AIC), the concordance index and Cox-Snell residuals. For models that included covariates, the one with lower AIC value was considered the best fit. The concordance index reflects how accurate the model ranks predicted event times (14). This statistic can be viewed as a generalization of the area under the curve and is interpreted similarly. A value of 0.5 corresponds to random predictions, a value of 1.0 indicates perfect concordance and, a value of 0.0 reflects perfect anti-concordance (multiply predictions with −1 to get 1.0).

Typically, survival models achieve concordance index values ranging 0.55 and 0.75. Although these numbers may appear suboptimal, it should be noted that even well-specified models are influenced by noise in the data, which makes very high values difficult to reach (18). Cox-Snell residuals are given by the formula:

rCi=−log(S^(ti))[10]

Where S^(t) is an estimated survival function. If the model is perfectly fitted to the data, then rCi should have exponential distribution exp(1). That means, that rCi should have mean equal to 1.

AIC will be used in parametric AFT and cure models. The concordance index will be used to check the fit of the parametric AFT models and Cox-Snell residuals will check if the cure model is correctly fitted to the data. Models with variables that have the lowest AIC and where all of the variables are statistically significant are considered the best models. For every time approach and every censoring approach, such models were created, only the best ones were chosen and later developed.

Results

Statistical analysis for all of the users

The mean age of the users of the application was 40±9 years and the median age was 41 years (IQR, 34–46 years). Women represented 38.5% of the total population (1,168 users); however, this percentage is subject to variation depending on the time and censoring approaches. For example, in T1C1 19% [218] of women were censored (did not churn) while 81% [950] experienced the event (they churned). In T3, the study population decreased because some users did not sign up for the application. Almost 56% of the users set their activity level as active. The mean value of body fat for users was 23.53%±8.19% and the median was 20% (IQR, 20–30%). The average BMI fell within the normal range at 24.23±3.74 kg/m², while the median was slightly lower at 23.67 (IQR, 21.72–25.95) kg/m². The frequency table (Table S1) and descriptive statistics (Table S2) are included in Figure S1.

Detailed frequency tables for smaller subgroups (based on 3-time approaches and 2 censoring approaches) are shown in Table 1. There were positive correlations between height and weight (ρ=0.66, P=0.03), weight and body fat (ρ=0.47, P=0.03), weight and BMI (ρ=0.82, P=0.02) and BMI and body fat (ρ=0.63, P=0.04).

Kaplan-Meier estimates

For the T1C1, there were significant differences between genders (male vs. female, P=0.002), activity level groups (active vs. sedentary, sedentary vs. very active, P=0.003; active vs. very active, P=0.01), goal groups (antiaging vs. lose, gain vs. lose, P=0.002), notification settings (True vs. False, P=0.02). For T1C2 the significant differences occurred for the variables gender (male vs. female, P=0.002), activity level (active vs. sedentary, sedentary vs. very active, P=0.002), goal (antiaging vs. lose, gain vs. lose, P=0.003) and notification settings (True vs. False, P=0.004). The curves for the approach T1C1 are shown in Figure 2 and Figure S1A.

Figure 2 Kaplan-Meier survival curves for T1C1 groups by gender (A), activity level (B), goal (C), and notification settings (D). T1C1, first censoring approach for Time 1.

For T1C1 (Figure 2) and T1C2 (Figure S1B), the figures show the probability of not dropping out for selected features. In the case of gender, women have lower probability of not dropping out compared to men (as seen in both censoring approaches). A similar trend can be observed for activity level in both approaches, where the lowest probability of not dropping out is seen in the sedentary group, and the highest in the very active group. However, it is noteworthy that, regarding notification settings, the group that does not turn on notifications has a higher chance of survival (i.e., not dropping out). This observation may be due to a significant imbalance in the data.

For the T2C1, there were significant differences between gender (male vs. female, P=0.002), activity level groups (active vs. sedentary, sedentary vs. very active, P=0.001; active vs. very active, P=0.01), goal (gain vs. lose, antiaging vs. lose, P=0.003) and notification settings (True vs. False, P=0.003). For T2C2, the significant differences between groups align with those observed in the first censoring approach. However, in this analysis, body type emerges as a significant factor (mid vs. strong, P=0.01; strong vs. thin, P=0.02).

The Kaplan-Meier curves, for T2C1 (Figure S1C) and T2C2 (Figure S1D) show results similar to those for T1. The main difference is that, in the first censoring approach (T2C1), most groups at the eighth week had a probability of not dropping out equal to 0, meaning there was no chance that any user would remain engaged beyond that point.

For T3C1 (Figure S1E) the significant differences between groups in variables were for gender (male vs. female, P=0.002), activity level (active vs. sedentary, sedentary vs. very active, P=0.004), goal (antiaging vs. lose, gain vs. lose, P=0.004) and notification settings (True vs. False, P=0.004). For T3C2 (Figure S1F), the significant differences were between groups in gender (male vs. female, P=0.002), activity level (active vs. sedentary, sedentary vs. very active, P=0.005; active vs. very active, P=0.04), goal (antiaging vs. lose, gain vs. lose, P=0.005) and notification settings (True vs. False, P=0.005). The conclusions are similar to T1 and T2, but for the T3C2 survival probabilities are stable from 12.5 weeks. Similarly, to T2, for T3C1 approach most of the survival probabilities are approaching probability 0.

Parametric models

Multiple parametric models were created, each considering different variables for each time and censoring approach. Initially, all variables were included in the models, and statistically insignificant variables were systematically removed one by one until only statistically significant variables remained. The models were created using Weibull, LogNormal and LogLogistic distributions, and were compared based on their AIC values to determine the most suitable model for each approach. Table 2 shows AIC values and concordance indexes for the models.

The chosen models are all from the second censoring approach, as the AIC is consistently lower compared to the first approach. The best model (the one with the highest concordance index) is T3C1. However, it could be appreciated that there are no high differences between values of concordance indexes. Table 3 presents a summary of the top-performing models tested on the training set. While for the best models for T2C2 and T3C2 have the same predictors, the T1C2 model has fewer predictors. To understand the results, it is crucial to examine the exponential of the coefficient [exp(coef) in Table 3] and compare it to number 1 (100%). To provide a clearer illustration, we have examined two cases for T1C2 model. On the one hand, the exponent of coefficient for activity level sedentary is equal to 0.866. The difference between 1 and 0.866 is −0.134, thus, setting activity level at sedentary (while signing up), decreases the probability of not dropping out by 13.4%, when all the other variables are constants. However, on the other hand, the setting users’ gender as male, has the coefficient exponent equal to 1.15. The difference between 1.15 and 1 is 0.15, indicating that selecting male as the gender, while keeping all other variables constant, result in a 15% increase in the median survival time (i.e., the probability of not dropping out increases). Upon analyzing the summary, it is clear that the coefficients consistently have the same signs across all time approaches. Specifically, in each time approach setting the activity level to “sedentary” decreases the probability of not dropping out.

To better illustrate users’ predicted survival curves for 5 random users from the test set, all time approaches are presented in Figure 3.

Figure 3 Predicted survival curves for five randomly selected users. (A) T1C2. (B) T2C2. (C) T3C2. T1C2/T2C2/T3C2, second censoring approach for Time 1, 2, or 3.

For T3C2, the greatest decrease in the probability of not dropping out occurs between the first and sixth week, after which the probability becomes more stable. In contrast, for T1C2 approach, the survival probability of not dropping out decreases in a linear fashion. Taking user 432 as example, the probability of not dropping out at the end of the follow-up period is approximately 0.35 for T1C2 and around 0.2 for T2C2. The predicted median survival times for this user are 2.26 weeks for T1C2, 2.66 weeks for T2C2, and 3.7 weeks for T3C2. The means between predicted and actual survival times are 1.02, 1.94 and 3.32 weeks, respectively.

Cure models

To select the best possible model with variables, the approach was the same as previously described for parametric models. The non-mixture versions with LogNormal distribution of the uncured fraction proved to be the best for all of the time and censoring approaches. Table 4 shows AIC and mean Cox-Snell residuals values for each time approach for the best possible model. By AIC value, the best model is for T1C2, but looking at the residuals, the best model is T2C2 approach.

The interpretation of coefficients for cure models slightly differs from parametric models, because interpretation is only for the probability of being in the cure fraction. Parameters (coefficients) are already in logistic form. For instance, on the one hand, for T1C2 the body fat coefficient is 0.96. This implies that an additional percentage (at the moment of downloading the app) decreases the odds of being cured by 4%. Moreover, male gender enhances the likelihood of being cured by 73%. On the other hand, by selecting female gender, the odds of being uncured are 0.58 times higher than when selecting male gender. Interestingly, for all 3-time approaches, the coefficients representing body fat exhibit the same coefficient value and P value. The details are given in Table 4.

The probability that a random individual in the test set would not stop using the app as long as it works (so would be cured) is respectively 66%, 41% and 47% and the fraction of users (in the test set) that would not stop using the app is respectively 0.47, 0.13 and 0.13, so in the test set the cure fraction is very low for T2C2 and T3C2 (Figure 4).

Figure 4 Hazard and survival curves across the first (T1), second (T2), and third (T3) time approaches. (A-C) Hazard curves of 5 randomly selected users for the first (A), second (B), and third (C) time approaches. (D-F) Survival curves of 5 randomly selected users for the first (D), second (E), and third (F) time approaches. (G-I) Hazard curves of the uncured fractions for the first (G), second (H), and third (I) time approaches.

The hazard curves for 5 randomly selected users from the test set for each time approach is given in Figure 4A-4C). For all the time approaches, the hazard curve increases and subsequently, decreases. That means, the probability of dropping out raises till the peak (respectively 1.3 weeks, circa 2 weeks and 2 weeks) and later decreases (in case of T1C2 and T3C2 it approaches 0, in the T2C2 it approaches 0.2). Figure 4D-4F shows the survival curves of 5 randomly selected users, and it could be observed that at the end of each follow-up period, the probabilities of not dropping out approached approximately 0.4, 0.2, and 0.15. The curves decrease most rapidly in the first 2, 4, and 5 weeks respectively, and then stabilize.

Figure 4G-4I presents the hazard curves of the uncured fraction for all time approaches. Notably, for T1, the hazard only increases, never decreases, meaning the probability of dropping out in the uncured fraction reaches 1 by the second week. In contrast, for T2 and T3, the probability of dropping out increases until the second week and then remains stable for the reminder of the follow-up period.

Discussion

The objective of this research was to assess the suitability of survival analysis for predicting user churn times in fitness applications. The results demonstrated significant differences among the Kaplan-Meier curves for various groups based on key variables. Of particular interest, men exhibited a higher probability of survival compared to women, while enabling notification settings was associated with decreased likelihood of survival, as observed by other studies (11,19). Further analysis of parametric and cure models confirmed these differences, providing valuable insights into which variables should be included in predictive models and highlighting their relative significance.

To illustrate this better, consider two examples of users from the T2C2 survival model, who represent high and low survival probabilities: users 432 and 2179. User 432, with the highest survival probability, is a male, standing at 184 cm and weighing 81 kg, with a BMI of 23.9 kg/m². He is very active, with the goal of gaining muscle, self-identifying his body type as “strong”, and setting his body fat percentage at 20%. Moreover, he is subscribed to the newsletter, has his notification settings enabled, and has committed to training 4 days a week. This user was censored. Conversely, user 2179, with the lowest survival probability, is a male standing at 174 cm and weighing 85 kg, with a BMI of 28.1 kg/m². He set the activity level as active with the goal of losing weight, identifies himself with a medium body type, and has a body fat percentage of 35%. Like user 432, he is also subscribed to the newsletter and has his notification settings enabled. However, his training frequency is lower, with only one day per week. Thus, the difference in demographic characteristics between the users explains why user 2179 exhibits a considerably lower survival probability compared to user 432. User 2179’s characteristics include obesity, higher weight, shorter stature, 15% more body fat, a different fitness goal and a lower frequency of training days.

Previous studies have analysed adherence to physical activity (20,21). In a study in 2021, adherence to fitness apps was predicted using 12 weeks of training session data. Users were grouped into activity-level clusters, and a Long Short-Term Memory (LSTM) model was applied to the time series data, achieving high accuracy in classifying adherence to fitness apps (5). In the present study we similarly compared groups based on activity level using Kaplan-Meier curves, and the differences are also significant.

The observed relationship between predictors and user survival aligns with established literature. For example, higher activity levels and more frequent training were associated with longer app usage (22). Similarly, more active users tend to maintain fitness centre membership longer. The unexpected result regarding notification settings (users who disabled notifications surviving longer) may reflect confounding factors such as user autonomy or over-notification fatigue (19). Moreover, our results and others, underscore the importance of personal and contextual characteristics (age, gender, body fat) in health outcome studies and their impact on retention and behavior (23).

Similar to other studies, this study contributes to the emerging application of survival analysis in non-medical domains by modelling user retention over time (not static) (11-13,24). This allows researchers to assess both the probability and the timing of user churn, providing deeper insight into user behaviour over time.

Modelling of parametric survival models that include demographic variables was not used in the survival analysis for fitness data. However, this study demonstrates that it is possible to predict survival with parametric survival models and highlights the significance of variables such as gender and age in the modelling process. Various studies, such as Okada et al. (25) or Kaminski & Geisler (26), have also demonstrated the significance of variables such as age and sex in survival analysis. However, it is important to note that these studies examined different scenarios and employed distinct models.

In other studies, different approaches have been considered to analyse the churn in the fitness domain, which are more commonly used than survival analysis. These include logistic regression (27), linear regression (28) or Random Forest (29). The cure models presented by (30) demonstrate their applicability in capturing customers’ recurring churn behaviour, offering a unique perspective beyond the conventional medical field, which aligns with our own results. Similarly, (31) employed nine data approaches (encompassing different follow-up time periods and sample sizes), which is similar to our study, where we considered a total of six approaches.

The influence of seasonality on physical activity levels is a critical aspect of research investigations. Numerous studies have demonstrated a consistent pattern of seasonal variations, with higher physical activity rates observed during summer compared to other seasons, notably winter (32). This seasonal fluctuation is a crucial factor that can significantly affect the interpretation and generalization of findings in studies focusing on physical activity as a primary outcome. Notably, our study spans a comprehensive 15-month period, systematically covering all four seasons. This deliberate inclusion allows for a nuanced understanding of how physical activity levels vary across different climatic conditions.

This study contributes by demonstrating the applicability of survival and cure models in digital health platforms, offering a novel analytical use for understanding app engagement and churn. The identification of key dropout predictors, such as gender, age, body fat and notification settings provide actionable insights that app developers can use to design targeted interventions, personalize user experiences, and optimize engagement strategies. The findings also serve as foundation for other digital health platforms seeking to adopt evidence-based retention models.

Strengths and limitations

The present study is one of the few that applies survival analysis (specifically Kaplan-Meier estimates, parametric AFT models and cure models) to assess user churn in a fitness app context. The use of multiple time and censoring approaches improves the robustness of the findings and provides important insights into user engagement trajectories. The relatively large and diverse user sample (n=3,034) strengths the generalizability of the results to other similar apps. Furthermore, the integration of both demographic app usage data allowed for the identification of key predictors of dropout, not only providing information about user retention, but also providing valuable foundation that other fitness app developers can use to enhance user engagement and reduce churn.

However, several limitations have been identified. First, the analysis is based on data from a single fitness app, which may limit external validity across different platforms. Second, the study relies on observational data, which may be subject to unmeasured confounding variables (i.e., user’s health, motivation, environmental factors, etc.). In addition, although the cure models offer useful estimates, the actual ‘cure fraction’ remains hypothetical and cannot be empirically validated. Finally, the model accuracy is limited by the follow-up period, event definition or statistical assumptions employed, suggesting the need for further validation using alternative survival or machine learning approaches.

Future research

Future work should address the limitations identified in this study. First, replication using data from other fitness applications would improve the external validity and enable potential comparisons. Second, incorporating behavioural data, such as user motivation, app usage or adherence could enhance the predictive power of the present churn models. Additionally, testing alternative survival models, such as conditional inference survival ensembles or exploring other cure models to achieve the optimal data. Further, validating the actual assumptions using long-term user tracking would provide more accurate survival estimates. Finally, the influence of contextual factors, such as seasonal trends, geographic variation could be integrated to provide a more holistic understanding of user adherence dynamics.

Conclusions

In summary, our findings advance user retention strategies, demonstrating that using demographic data collected during app download enables accurate churn prediction. Furthermore, our approach reveals the potential to identify the proportion of users who will continue using the app as long as it remains operational.

Acknowledgments

The authors would like to thank Mammoth Hunters and especially Oriol Roda Noguera and Nestor Sanchez for the data provided to conduct this research.

Footnote

Data Sharing Statement: Available at https://mhealth.amegroups.com/article/view/10.21037/mhealth-25-15/dss

Peer Review File: Available at https://mhealth.amegroups.com/article/view/10.21037/mhealth-25-15/prf

Funding: This work was supported by funds from Aristos Campus Mundus projects (ACM2022_25 and ACM2023_19), and Fundació Naccari-Ravà (FNR2020-23).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://mhealth.amegroups.com/article/view/10.21037/mhealth-25-15/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Ethics and Research Committee of the Faculty of Psychology, Education Sciences and Sport Blanquerna of the Ramon Llull University (No. 2223013D). Participants gave informed consent to participate in the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Zhang X, Xu X. Continuous use of fitness apps and shaping factors among college students: A mixed-method investigation. Int J Nurs Sci 2020;7:S80-7. [Crossref] [PubMed]
2. Ferreira Barbosa H, García-Fernández J, Pedragosa V, et al. The use of fitness centre apps and its relation to customer satisfaction: a UTAUT2 perspective. International Journal of Sports Marketing and Sponsorship 2022;23:966-85.
3. Alturki RM, Gay V. A Systematic Review on What Features Should be Supported by Fitness Apps and Wearables to Help Users Overcome Obesity. IJRET 2016;05:197-206.
4. Martín F, García-Fernández J, Valcarce-Torrente M, et al. Importance-performance analysis in fitness apps. A study from the viewpoint of gender and age. Front Public Health 2023;11:1226888. [Crossref] [PubMed]
5. Lyu L, Abidin NEZ, Zulnaidi H. Exploring female fitness app users' motivations and perceptions: A qualitative study. Medicine (Baltimore) 2024;103:e40076. [Crossref] [PubMed]
6. Sousa Basto P, Ferreira P. Mobile applications, physical activity, and health promotion. BMC Health Serv Res 2025;25:359. [Crossref] [PubMed]
7. Lin YP, Lee KC, Ma WF, et al. A mobile technology-based tailored health promotion program for sedentary employees: development and usability study. BMC Public Health 2025;25:1452. [Crossref] [PubMed]
8. Viljanen M, Airola A, Heikkonen J, et al. Playtime Measurement With Survival Analysis. IEEE Transactions on Games 2018;10:128-38.
9. Yeomans C, Karg A, Nguyen J. Who churns from fitness centres? Evidence from behavioural and attitudinal segmentation. Managing Sport and Leisure 2024; [Crossref]
10. Nguyen D, Kleeman NJ, Yager Z, et al. Identifying barriers and facilitators to implementing mindfulness-based programmes into schools: A mixed methods study. Appl Psychol Health Well Being 2022;14:1172-88. [Crossref] [PubMed]
11. Periáñez Á, Saas A, Guitart A, et al. Churn Prediction in Mobile Social Games: Towards a Complete Assessment Using Survival Ensembles. In: 2016 IEEE International Conference on Data Science and Advanced Analytics (DSAA) 2016:564-73. Available online: https://ieeexplore.ieee.org/document/7796943
12. Demediuk S, Murrin A, Bulger D, et al. Player retention in league of legends: a study using survival analysis. In: Proceedings of the Australasian Computer Science Week Multiconference. New York, NY, USA: Association for Computing Machinery; 2018:1-9. (ACSW ’18). Available online: https://dl.acm.org/doi/10.1145/3167918.3167937
13. Jossa-Bastidas O, Zahia S, Fuente-Vidal A, et al. Predicting Physical Exercise Adherence in Fitness Apps Using a Deep Learning Approach. Int J Environ Res Public Health 2021;18:10769. [Crossref] [PubMed]
14. Davidson-Pilon C. lifelines: survival analysis in Python. Journal of Open Source Software 2019;4:1317.
15. Jakobsen LH, Clements M, Jensen RK, et al. cuRe: Parametric Cure Model Estimation 2023 [cited 2024 Jan 8]. Available online: https://cran.r-project.org/web/packages/cuRe/
16. Medium LE. Introduction to Survival Analysis: the Kaplan-Meier estimator. 2022 [cited 2024 Jan 8]. Available online: https://medium.com/data-science/introduction-to-survival-analysis-the-kaplan-meier-estimator-94ec5812a97a
17. Peng Y, Yu B. Cure Models: Methods, Applications, and Implementation. New York: CRC Press; 2021.
18. Zakrzewska M. Modelling and analysis survival data with cure fraction. Wroclaw, Poland; 2022.
19. Gani I, Litchfield I, Shukla D, et al. Understanding "Alert Fatigue" in Primary Care: Qualitative Systematic Review of General Practitioners Attitudes and Experiences of Clinical Alerts, Prompts, and Reminders. J Med Internet Res 2025;27:e62763. [Crossref] [PubMed]
20. André N, Grousset M, Audiffren M. A Behavioral Perspective for Improving Exercise Adherence. Sports Med Open 2024;10:56. [Crossref] [PubMed]
21. Fuente-Vidal A, Blanco R, Prat R, et al. Analyzing motivation for tele-exercise in adult fitness app users. mHealth 2025;11:50.
22. Yang Y, Koenigstorfer J. Determinants of Fitness App Usage and Moderating Impacts of Education-, Motivation-, and Gamification-Related App Features on Physical Activity Intentions: Cross-sectional Survey Study. J Med Internet Res 2021;23:e26063. [Crossref] [PubMed]
23. McChrystal R, Lees J, Gillies K, et al. Participant and trial characteristics reported in predictive analyses of trial attrition: an umbrella review of systematic reviews of randomised controlled trials across multiple conditions. Trials 2025;26:84. [Crossref] [PubMed]
24. Linardon J, Firth J, Torous J, et al. Efficacy of mental health smartphone apps on stress levels: a meta-analysis of randomised controlled trials. Health Psychol Rev 2024;18:839-52. [Crossref] [PubMed]
25. Okada E, Ukawa S, Nakamura K, et al. Demographic and lifestyle factors and survival among patients with esophageal and gastric cancer: The Biobank Japan Project. J Epidemiol 2017;27:S29-35. [Crossref] [PubMed]
26. Kaminski D, Geisler C. Survival analysis of faculty retention in science and engineering by gender. Science 2012;335:864-6. [Crossref] [PubMed]
27. Clavel San Emeterio I, García-Unanue J, Iglesias-Soler E, et al. Prediction of abandonment in Spanish fitness centres. Eur J Sport Sci 2019;19:217-24. [Crossref] [PubMed]
28. Perna S, Spadaccini D, Riva A, et al. A path model analysis on predictors of dropout (at 6 and 12 months) during the weight loss interventions in endocrinology outpatient division. Endocrine 2018;61:447-61. [Crossref] [PubMed]
29. Wullems JA, Verschueren SMP, Degens H, et al. Performance of thigh-mounted triaxial accelerometer algorithms in objective quantification of sedentary behaviour and physical activity in older adults. PLoS One 2017;12:e0188215. [Crossref] [PubMed]
30. Kumar V, Leszkiewicz A, Herbst A. Are you Back for Good or Still Shopping Around? Investigating Customers’ Repeat Churn Behavior. Journal of Marketing Research 2018;55:208-25.
31. Kearns B, Stevenson MD, Triantafyllopoulos K, et al. The Extrapolation Performance of Survival Models for Data With a Cure Fraction: A Simulation Study. Value Health 2021;24:1634-42. [Crossref] [PubMed]
32. Garriga A, Sempere-Rubio N, Molina-Prados MJ, et al. Impact of Seasonality on Physical Activity: A Systematic Review. Int J Environ Res Public Health 2021;19:2. [Crossref] [PubMed]