A machine-learning model developed for young and middle-aged people receiving maintenance hemodialysis classified whether they were currently in paid work with strong results in a separate patient group, according to a Chinese study. The reported current return-to-work rate was 40.23%. In this research, return to work, or RTW, meant paid full-time or part-time employment at the time of the survey. The model therefore addressed a snapshot of work status, rather than a person's chances of working later.
That distinction matters when reading the findings. Because the research was cross-sectional, the model does not establish that confidence about returning to work, social support, dialysis time or any other measured feature caused employment. It also does not predict future or sustained employment. Its output is a current-status classification, not a promise about what happens next.
A model built from a regional hospital sample
Researchers recruited 793 patients through convenience sampling at eight tertiary hospitals in Guangdong between March 2023 and May 2024, and 768 were included in the analysis. Six Guangzhou and Foshan centers supplied the 565-person modeling set. The remaining 203 patients came from two Shenzhen centers and formed a non-overlapping external validation group held back until the final evaluation.
Five approaches were compared: logistic regression, support vector machine, random forest, XGBoost and AdaBoost. The researchers used nested cross-validation, a repeated training and testing process, with 10 outer folds and five inner folds. A grid search selected model settings by maximizing AUC, a score that summarizes how well a model separates the two work-status groups.
Before settling on the final model, the team used LASSO, a technique for paring down a list of variables, to reduce 22 candidates to 10. The retained inputs were gender, age, residence, chronic comorbidities, dialysis time period, return-to-work self-efficacy, family functioning, social support, social isolation and social functioning. The selection process identifies what the model uses; it does not, by itself, show that any one factor changes a person's employment prospects.
The separate test produced similar results
In the validation comparison, the random forest had the highest AUC at 0.946, with a 95% confidence interval, the study's uncertainty range, from 0.892 to 0.997. Its validation average precision, a summary of precision-recall performance, was 0.906, with a 95% confidence interval of 0.884 to 0.928. The maximum decision-curve net benefit was 0.382.
Internal validation produced an AUC of 0.939, with a reported interval of 0.916 to 0.962. Accuracy was 0.874, sensitivity was 0.829 and specificity was 0.901. The reported learning-curve gap was about 0.064, and no apparent overfitting or underfitting was reported.
The separate check came from the 203-patient external cohort. There, the random forest had an AUC of 0.918, with a 95% confidence interval from 0.870 to 0.963, and an accuracy of 0.901. Sensitivity, the share of RTW cases identified, was 0.928; specificity, the share of non-RTW cases identified, was 0.877. The Brier score, a way of comparing predicted probabilities with observed outcomes, was 0.088, with a 95% confidence interval from 0.061 to 0.121. The analysis reported positive decision-curve net benefit across thresholds from 10% to 80%.
A web application built around the model returns a probability for current RTW and applies a 0.5 cutoff. Probabilities below 0.5 are classified as non-RTW, while probabilities above 0.5 are classified as RTW. That output labels survey-time employment status; it is not evidence about whether employment will continue.
The model's clues are not causes
To show what shaped its results, the researchers used SHAP, an explanation method for model outputs. It ranked the contributors in this order: return-to-work self-efficacy, social functioning, chronic comorbidities, social support, social isolation, family functioning, age, dialysis time period, residence and gender. The ranking shows which inputs influenced the model's calculations most strongly; it is not a list of proven causes.
The study's boundaries are substantial. The sample came from Chinese tertiary hospitals, RTW status was self-reported and may be affected by recall bias, and the external test was drawn from two Shenzhen centers, limiting stable site-specific assessment. The researchers did not assess whether performance was fair across demographic subgroups. Convenience sampling may also limit how well the results represent other patients.
For now, the evidence supports a narrow use: classifying current work status in the studied population. The model's ability to predict future or sustained employment remains untested, and whether changing the factors it highlights improves return to work is unresolved. Those questions would require evidence beyond this cross-sectional analysis.
Paper data and sources
Original title: Development and validation of a machine learning-based return to work predictive model for young and middle-aged maintenance hemodialysis patients
Authors: Mai Q, Ma X, Yang Z et al.
Journal/Repository: Frontiers in public health
Status: Peer-reviewed
First online: 2026-10-09
DOI: Not available
Original paper · Full text