Penalized competing-risk model predicts dialysis transition in multicenter retrospective cohort

A penalized Fine–Gray competing-risk model developed from a multicenter retrospective cohort of 700 adults receiving peritoneal dialysis performed better than three alternative prediction models in an independent test set. The model estimated the cumulative probability of transitioning from peritoneal dialysis to hemodialysis while treating death as a competing event rather than simply censoring it. Across the full follow-up period, the penalized model achieved a mean test-set area under the time-dependent curve of 0.689 and a mean concordance index of 0.646. Its time-specific areas under the curve were 0.667 at three years, 0.671 at six years and 0.713 at nine years in the test cohort. The findings identify body mass index, previous hospitalization for peritonitis and nutritional status as major predictors and yield a nomogram for estimating three-, six- and nine-year transition risk.

The study analyzed clinical records from three dialysis centers in Chongqing, China: Daping Hospital, Kaizhou District People’s Hospital and Yongchuan Hospital of Chongqing Medical University. Eligible participants were at least 16 years old, had chronic renal failure and had received regular peritoneal dialysis for more than one month. The source cohort contained 824 records for people who began peritoneal dialysis for the first time between August 2007 and September 2025. After excluding patients with acute kidney injury, emergency dialysis, subsequent renal recovery or kidney transplantation, as well as those lost to follow-up, 737 records remained. Detection of 37 statistical outliers produced a final analytic cohort of 700 patients, who were randomly divided into a training set of 491 and a test set of 209.

During follow-up, 320 patients transitioned to hemodialysis, 178 died and 202 remained on peritoneal dialysis without either event by the end of observation. The investigators defined a hemodialysis transition as a move to hemodialysis lasting more than 90 days without a return to peritoneal dialysis, following a 2026 position statement from the International Society for Peritoneal Dialysis. Patients who died within that 90-day observation window were classified as having died rather than having made a sustained transition. Patients with unknown outcomes at censoring were treated as censored observations, not as competing deaths. Kidney transplantation and renal recovery were excluded from the primary analysis because those categories had few events, which the researchers said could make competing-risk estimates unstable.

Because the investigators expected death to compete with the planned hemodialysis transition, they compared a cause-specific proportional hazards model, two Fine–Gray model-development strategies and a conventional LASSO-Cox model that censored death. The cause-specific model used LASSO feature selection with tenfold cross-validation, while the Fine–Gray models were built using three penalized regression methods—LASSO, smoothly clipped absolute deviation and minimax concave penalty—as well as a boosting procedure. The researchers retained variables identified in univariate screening but also allowed the feature-selection procedures to consider other potentially relevant clinical characteristics. Predictive performance was assessed with time-dependent area under the curve and concordance-index curves, while calibration and discrimination were examined at three, six and nine years.

In the training set, the penalized Fine–Gray model had a mean time-dependent area under the curve of 0.716, compared with 0.696 for the cause-specific model and 0.718 for the boosted Fine–Gray model. Its training-set concordance index was 0.662. The conventional LASSO-Cox model had the highest mean training area under the curve, 0.762, but its performance was weaker in the held-out data: its test-set concordance index was 0.573, compared with 0.646 for the penalized Fine–Gray model. The penalized model also had the highest mean test-set area under the curve at 0.689. The investigators therefore selected it as the best overall model, citing its held-out performance and its smaller predictor set relative to the boosted Fine–Gray model. This was a predictive comparison within one retrospective dataset, not evidence that changing care according to the model would improve patient outcomes.

The final penalized model associated a higher body mass index with a greater subdistribution hazard of hemodialysis transition, with a subdistribution hazard ratio of 1.05 and a 95% confidence interval of 1.00 to 1.09 per BMI unit. A history of hospitalization for peritonitis was associated with more than twice the subdistribution hazard, with a ratio of 2.15 and a confidence interval of 1.64 to 2.84. Better nutrition-related parameters were associated with lower transition hazard, with a ratio of 0.27 and a confidence interval of 0.08 to 0.90. Other significant associations in the penalized model included older age, previous urological surgery and previous hemodialysis. The investigators discussed several plausible mechanisms without proving them in this study: peritonitis may make peritoneal dialysis technically unsustainable or damage the peritoneum; malnutrition may impair immune function and tissue repair; and higher BMI has previously been associated with peritonitis, loss of residual kidney function and hospitalization.

Researchers converted the penalized model into a nomogram that assigns points to clinical variables and maps a patient’s total score to estimated cumulative incidence of hemodialysis transition at three, six and nine years. In a worked example, a 60-year-old man with a BMI of 20.4 kg/m², reduced urine output and previous hospitalization for both peritonitis and cardiovascular disease received approximately 398 points. The nomogram estimated his cumulative transition risk as 39% at three years, 60% at six years and 69% at nine years. Patients in the test cohort were divided into low-, middle- and high-risk groups using score thresholds derived from the training set. Their cumulative incidence curves differed significantly for both hemodialysis transition and death, and the high-risk group had the highest incidence of transition, although a minor early crossing occurred between the low- and middle-risk groups.

Diagnostics also exposed important limitations in the model’s statistical assumptions. The proportional subdistribution hazards assumption was violated for normal 24-hour urine output, hospitalization for peritonitis and the nutrition-related parameters. Stratifying by peritonitis hospitalization did not resolve the violation, while splitting patients at a nutrition-derived threshold produced subgroups that satisfied the assumption. However, the well-nourished subgroup showed substantial differences between training and test performance: its mean area under the curve fell from 0.763 in training to 0.609 in testing, and its concordance index fell from 0.686 to 0.579. The malnutrition subgroup performed more closely between training and testing, with area-under-the-curve values of 0.714 and 0.691 and concordance indices of 0.661 and 0.640. The authors consequently treated the subgroup models as exploratory and retained nutrition as a continuous variable in the main analysis.

The retrospective design, missing laboratory data and lack of time-varying clinical measurements constrain how broadly the results can be applied. Some clinical indicators had missing rates exceeding 30%, requiring mean, nearest-neighbor or multiple-linear-regression imputation that may have introduced bias. The dataset was drawn from three centers in one Chinese city, and the paper’s limitations section describes the work as single-center despite the multicenter data sources, an internal inconsistency that further emphasizes the need for independent validation. The model also used baseline information rather than changes in nutrition, infection or laboratory status during follow-up, and it did not establish that risk prediction led to better clinical decisions or outcomes. The authors proposed prospective multicenter validation and models capable of incorporating dynamic covariates. Until those steps are completed, the penalized Fine–Gray model and nomogram are best regarded as retrospective risk-estimation tools whose discrimination is moderate, with meaningful uncertainty around their individual predictions.

Subject of Research: Competing-risk prediction of peritoneal dialysis-to-hemodialysis transition in adults

Article Title: Optimal prediction of peritoneal dialysis to hemodialysis transition: a penalized fine-gray model

Article References: Wan, J., Wei, Q., Wu, D., Hou, S., Yv, K., Tang, R., He, J., Chen, X., Li, Y., Hu, Y., Yang, J., Li, T., & Chen, K. (2026). Optimal prediction of peritoneal dialysis to hemodialysis transition: a penalized fine-gray model. Frontiers in Medicine, 13, Article 1843381. https://doi.org/10.3389/fmed.2026.1843381

Image Credits: AI Generated

DOI: 10.3389/fmed.2026.1843381

Keywords: peritoneal dialysis, hemodialysis transition, competing risks, Fine–Gray model, clinical prediction, nomogram, peritonitis, nutritional status, retrospective cohort, model validation, Optimal, prediction