A machine-learning model designed to flag fear of movement in older people after coronary procedures performed well in the study’s internal test and in a later group of patients. Its ability to distinguish people with and without the condition was measured by the area under the curve, or AUC. The model recorded an AUC of 0.821 in the internal validation group and 0.981 in the time-separated external group.
The findings suggest the tool could help clinical teams decide who may need a closer assessment for kinesiophobia, the study’s term for fear of movement. They do not show that the model can predict who will develop the problem later, or that it should replace clinical judgment.
What the model measured
The researchers studied older patients with coronary heart disease after percutaneous coronary intervention, or PCI. The study aimed to identify kinesiophobia present at the time of assessment, rather than forecast a future condition.
The initial analysis included 502 participants. Of those, 321, or 63.9%, met the study’s definition of kinesiophobia. The researchers divided the initial group into 352 people for model development and 150 for internal validation. A separate 150-person group was recruited in a later time window for external validation; 88 people in that group, or 58.7%, met the definition. The reported difference in prevalence across cohorts was not statistically significant, with P = 0.503.
Kinesiophobia was assessed with a 17-item questionnaire covering four dimensions. Scores ranged from 17 to 68, and a total score of at least 37 classified a participant as having the condition.
A cautious selection process
The final model used 15 predictors drawn from functional, clinical, demographic, behavioral and psychosocial information. The researchers selected them by keeping the predictors chosen by both LASSO, a method that narrows a larger set of variables, and Boruta, a method that compares variables with randomized alternatives.
Five algorithms were compared using nested cross-validation. In this process, a 10-fold outer loop estimated performance while a five-fold inner loop helped tune the models. Data preparation, feature selection, resampling and model fitting were restricted to the relevant training portions. XGBoost received the best overall rank under a prespecified scoring system that considered discrimination, calibration, prediction error and potential clinical net benefit, and it was chosen before validation.
On the development data, Logistic Regression had the highest average outer-fold AUC at 0.884, with reported variation of 0.076. XGBoost was close behind at 0.878 with variation of 0.072. The other averages were 0.879 for Support Vector Machine, 0.876 for Random Forest and 0.795 for Decision Tree.
Performance was strong, but not uniform
In the 150-person internal validation group, XGBoost had an AUC of 0.821, with a 95% confidence interval from 0.755 to 0.884. Its accuracy was 0.760, sensitivity was 0.781 and specificity was 0.722. In plain terms, the model identified many participants who met the study definition, while also correctly excluding a substantial share of those who did not.
The later external group produced a much higher AUC of 0.981, with a 95% confidence interval from 0.963 to 0.994. Accuracy was 0.920, sensitivity 0.909 and specificity 0.935. The model’s Brier score, the study’s measure of probability prediction error, was 0.062 in this group, compared with 0.170 in internal validation.
That striking external result needs careful reading. The later group was not drawn from different regions or hospitals. It came from the same three hospitals during a different recruitment period, making it a temporal rather than geographically independent test. The researchers also observed differences in the mix of patients between cohorts. Together, those features complicate interpretation of the unusually high AUC.
What the model says, and what it cannot say
An interpretability analysis showed which variables contributed most to the model’s predictions. The largest average contributions were resignation coping, exercise self-efficacy, avoidance coping, fear of disease progression, depression and self-perceived burden. The values ranged from 0.720 for resignation coping to 0.309 for self-perceived burden.
These contribution scores describe predictive associations, not causes. Because the outcome was measured concurrently with the predictors, the study cannot establish which factors lead to kinesiophobia or predict whether the condition will persist or change over time.
The authors say the model could serve as an aid for identifying patients who warrant further assessment and individualized cardiac-rehabilitation support. They also say it needs broader testing in geographically independent populations and may need recalibration before routine clinical use. For now, the tool should not substitute for professional assessment.
Paper data and sources
Original title: Development and validation of an interpretable machine learning model for concurrent identification of kinesiophobia in older patients with coronary heart disease after percutaneous coronary intervention
Authors: Xi Y, Zhang M, Han L et al.
Journal/Repository: Frontiers in medicine
Status: Peer-reviewed
First online: 2026-10-08
DOI: Not available
Original paper · Full text