Clinical Prediction Models for Life Expectancy of Patients on Hemodialysis

The life expectancy of patients being treated with maintenance hemodialysis (MHD) is varied. However, current tools for measuring life expectancy focus primarily on near-term mortality, which can affect care decisions.

Researchers, including Benjamin A. Goldstein, PhD, developed predictive models for near-term mortality and long-term survival on MHD. Their data came from the electronic health record systems of midsize, nonprofit dialysis providers and included 42,351 patients over 11 years. The data included demographics, laboratory results, vital signs, and service utilization information.

For each patient month, the researchers determined near-term mortality (death within the next 6 months) and long-term survival (survival over >5 years while receiving MHD or after a kidney transplant). They used least absolute shrinkage and selection operator logistic regression and gradient-boosting machines to predict each outcome, compared the results to time-to-event models spanning both time horizons, and examined the performance of decision rules at different cut points.

All models achieved area under the receiver operating characteristic curve ≥0.80 and optimal calibration metrics in the test set. The long-term survival models performed significantly better than near-term mortality models. Time-to-event models performed comparably to binary models. By applying different cut points from the first to 90th percentile of the predictions, the models could achieve a positive predictive value (PPV) of 54% for near-term mortality; however, sensitivity was only 6%. They could achieve a PPV of 71% for long-term survival with 67% sensitivity.

While predictive modeling built on readily available clinical data holds promise as a clinical decision support tool, the researchers’ retrospective models would need to be prospectively validated before use.

Source: American Journal of Kidney Diseases