Gradient Boosting Machine Learning Accurately Predicts Digital Health Platform Customer Churn

Researchers evaluating user retention across telehealth and mobile medicine platforms have demonstrated that advanced ensemble gradient boosting algorithms can reliably forecast customer churn by capturing complex behavioral and engagement patterns. In a study published in the journal Discover Health Systems, investigators Safar Fazli and Mohammad Javad Mohammadi examined the predictive capabilities of three leading gradient boosting frameworks—XGBoost, LightGBM, and CatBoost—using real-world operational records from an active digital healthcare enterprise. Customer attrition represents an acute operational challenge for remote healthcare organizations, threatening recurring enterprise revenue while simultaneously disrupting the continuous clinical care essential for managing long-term patient outcomes. The researchers found that following systematic hyperparameter optimization using the Optuna framework, all three ensemble architectures achieved strong classification metrics on anonymized platform data, with the extreme gradient boosting algorithm, XGBoost, demonstrating the highest overall accuracy at 84.1 percent alongside superior sensitivity in detecting at-risk users who ceased platform activity.

The operational environment of digital healthcare platforms diverges significantly from standard e-commerce and commercial subscription services, creating distinct analytical considerations when evaluating user attrition. Digital health platforms encompass diverse clinical touchpoints, ranging from online physician appointments and virtual consultations to at-home diagnostic testing services. In conventional software businesses, customer disengagement primarily triggers financial attrition; however, in digital healthcare, customer churn directly severs clinical continuity and patient adherence to ongoing health regimens. Moreover, user behaviors within healthcare applications are shaped by urgent medical necessity, personal emotional anxiety, and privacy considerations rather than casual consumer convenience. Consequently, standard financial records and demographic profiles offer limited visibility into impending departures. Instead, granular behavioral metrics reflecting interaction intensity, including average session duration, application login frequency, and appointment adherence, serve as vital leading indicators of whether patients perceive sustained clinical utility, institutional reliability, and operational trust in the underlying digital health ecosystem.

To evaluate predictive algorithms within this specialized healthcare environment, the researchers compiled a secondary operational dataset directly from the internal information systems of an established digital health organization. At the company’s explicit request, its corporate name was withheld and all identifying details remained confidential to protect proprietary commercial standing. The operational dataset captured twelve consecutive months of patient interactions between April 2024 and April 2025 across mobile applications and desktop web interfaces. The initial extraction yielded 3,500 customer records, each representing an individual user interacting with the platform’s medical and administrative services. After excluding profiles with missing target labels, the researchers retained 3,343 validated customer observations for empirical modeling. Customer churn was defined as a binary status marking complete service cessation and total non-interaction for three consecutive months following a patient’s last recorded platform event. Within this final cohort, 46 percent of records represented churned customers while 54 percent represented active users, establishing a moderately imbalanced class distribution reflective of real-world attrition patterns.

The research team structured a comprehensive feature pipeline combining behavioral, demographic, and transactional variables to characterize user interactions accurately. Behavioral features included average session duration, logins per month, active days per month, update frequency, and retention time. Demographic records captured customer age, geographical residency across Iranian provinces, and device type categorized as personal computer or mobile phone. Transactional and clinical data tracked total amount paid, treatment type, ongoing treatment status, prescription status, follow-up appointments, and diagnosis counts. Data preprocessing addressed missing values by imputing numerical features with the median to avoid outlier skewness and filling categorical gaps using mode imputation to preserve baseline category distributions. To ensure algorithmic compatibility with numerical inputs, categorical features were converted via one-hot encoding for XGBoost and LightGBM, whereas CatBoost utilized its native encoding mechanism. Class imbalance was mitigated during training using stratified cross-validation alongside model-specific weighting, including scale positive weight in XGBoost and LightGBM, LightGBM’s unbalance setting, and CatBoost’s balanced class weight configuration.

Model development implemented three distinct gradient boosting architectures evaluated through a Stratified 5-Fold Cross-Validation scheme with a fixed random seed, ensuring reproducible partitions where 80 percent of records formed training subsets and 20 percent served as validation folds. XGBoost constructs decision trees using level-wise expansion supported by L1 and L2 penalties to prevent overfitting. LightGBM adopts a leaf-wise growth mechanism combined with Gradient-Based One-Side Sampling and Exclusive Feature Bundling, accelerating computation across high-dimensional features. To maximize predictive accuracy beyond baseline settings, the researchers employed the Optuna framework for hyperparameter optimization using Bayesian Tree-structured Parzen Estimators. The optimization protocol systematically tuned tree depth, number of estimators from 100 to 2000, learning rates between 0.01 and 0.3, subsampling ratios, column sampling, and regularization coefficients, focusing primarily on maximizing minority-class Recall and F1-Score to enhance churn detection sensitivity.

Baseline evaluations conducted under default parameter configurations demonstrated that all three algorithms provided acceptable classification capabilities, while revealing notable variations in their sensitivity to customer attrition. XGBoost attained the highest default accuracy at 82.5 percent, recording 83 percent precision for retained users and 82 percent precision for churners. In terms of recall, default XGBoost correctly identified 80 percent of actual churn instances and 84 percent of non-churn instances, maintaining an average precision of 82 percent across classes. CatBoost followed closely, achieving an overall accuracy of 82.3 percent with balanced precision and recall scores of 82 percent for both classes. Conversely, LightGBM recorded an accuracy of 81.7 percent, displaying an uneven recall distribution that identified 85 percent of retained users but only 78 percent of churn cases. Furthermore, confusion matrix analysis revealed that all three default models generated more false negatives than false positives, underscoring an inherent baseline tendency across untuned gradient boosting models to miss true churners.

Hyperparameter optimization using the Optuna framework elevated the predictive performance and stability of all three models, resulting in closely aligned metrics across the ensemble algorithms. Tuned XGBoost demonstrated the best overall performance, increasing its accuracy to 84.1 percent with a low cross-validation standard deviation of 1.2 percent. Optimized XGBoost achieved 86 percent precision for non-churners, 82 percent precision for churners, an elevated churn recall of 85 percent, an F1-Score of 83 percent, and an Area Under the ROC Curve of 0.91. LightGBM’s accuracy rose to 83.8 percent with a 1.3 percent standard deviation, achieving 84 percent recall across both classes and the highest discriminative ability with an AUC of 0.92. CatBoost reached 83.8 percent accuracy, matching LightGBM’s 84 percent recall, but displayed a higher cross-validation standard deviation of 2.0 percent and an AUC of 0.91. The researchers observed that CatBoost experienced smaller relative gains post-tuning compared to XGBoost and LightGBM, attributing this stability to CatBoost’s robust default configuration, internal regularization, and built-in categorical encoding.

Global feature importance analysis revealed the primary behavioral and operational variables driving model decisions, while highlighting variations caused by differing algorithmic importance metrics. In XGBoost, where importance is calculated using split frequency, Device Type emerged as the most critical predictor, indicating that platform access modality via personal computer or mobile application creates distinct user experiences that heavily influence retention. Following Device Type, Prescription status and behavioral features—specifically Average Session Duration, Update Frequency, and Logins Per Month—ranked as the most influential variables. Conversely, LightGBM, which computes feature importance based on total loss reduction gain, identified Average Session Duration, Active Days Per Month, and Logins Per Month as its leading predictors, followed by Total Amount Paid. CatBoost, utilizing permutation importance, similarly ranked Average Session Duration as its most decisive feature. Across all models, behavioral indicators of usage intensity consistently outranked demographic attributes, demonstrating that user engagement depth represents the foundational empirical driver of churn prediction.

To evaluate the directional impact of individual features on attrition risk, the research team conducted a game-theoretic SHAP analysis. In the XGBoost model, Update Frequency exhibited the widest spread of SHAP values, exerting the strongest single impact on model output. Across all models, higher values for Average Session Duration, Logins Per Month, Active Days Per Month, App Retention Time, Follow-Up Appointments, and Number of Diagnoses consistently exerted a negative influence on churn probability, indicating that frequent, prolonged platform interactions substantially promote customer retention. Mobile platform access and older age also correlated with reduced attrition. Conversely, an unexpected inverse relationship emerged regarding Update Frequency across all three algorithms: higher update frequencies were associated with increased churn probability, whereas lower update frequencies corresponded to higher retention. The authors emphasized that SHAP values reflect associative predictive contributions rather than verified causal mechanisms, hypothesizing that frequent software updates might introduce operational bugs or disruptive interface changes that undermine user trust and provoke attrition.

Integrating these empirical findings with the Customer Lifecycle framework provides healthcare managers with actionable insights for designing stage-specific retention interventions. In lifecycle theory, customer churn represents an observable weakening within the retention and loyalty stages rather than an abrupt termination. Declining trends in session duration, monthly logins, and active days serve as early warning signals indicating diminishing user satisfaction and clinical engagement. Meanwhile, transactional factors such as ongoing prescriptions and total expenditure create switching costs that stabilize patients within the retention phase. The authors noted several study limitations, including reliance on a cross-sectional dataset from a single private digital health provider and the absence of formal statistical significance testing between model accuracy scores. To advance digital health analytics, the researchers recommended that future investigations evaluate sequential deep learning architectures such as recurrent neural networks and transformers, utilize multi-organizational datasets to enhance generalizability, and apply customer segmentation techniques like K-means clustering to further clarify the mechanisms driving digital health attrition.

Subject of Research: Machine learning prediction of customer churn in digital health platforms

Article Title: Leveraging gradient boosting machine learning models to predict customer churn in digital health platforms

Article References: Fazli, S., & Mohammadi, M. J. (2026). Leveraging gradient boosting machine learning models to predict customer churn in digital health platforms. Discover Health Systems, 5(1), Article 51. https://doi.org/10.1007/s44250-026-00387-y

Image Credits: AI Generated

DOI: 10.1007/s44250-026-00387-y

Keywords: Machine Learning, Digital Health, XGBoost, LightGBM, CatBoost, Customer Churn, Explainable AI, SHAP, Healthcare Analytics, Leveraging, gradient, boosting