FRAILTY SCREENING IN MAINTENANCE HEMODIALYSIS USING ROUTINELY COLLECTED DIALYSIS DATA: A MULTICENTER MACHINE LEARNING STUDY

Frailty is common in maintenance hemodialysis and is associated with adverse outcomes, yet scalable screening tools for busy clinics remain limited. We evaluated machine learning (ML) models to screen for frailty using only routinely collected dialysis data.

We conducted a multicenter cross-sectional study of maintenance hemodialysis patients across 19 centers (April 2018–March 2019). Frailty was defined as Japanese Cardiovascular Health Study (J-CHS) score ≥3. Thirty-two candidate predictors were screened with the Boruta algorithm; selected variables included serum creatinine (Cre), cardiothoracic ratio (CTR), height, dry weight (DW), C-reactive protein (CRP), total cholesterol, age, blood urea nitrogen (BUN), dialysis vintage, chloride (Cl), albumin (Alb), uric acid (UA), hematocrit (Ht), single-pool Kt/V, and normalized protein catabolic rate (nPCR). We trained five models: logistic regression (LR), random forest, LightGBM, CatBoost, and a soft-voting ensemble (VE). Data were randomly split into 70% development and 30% test sets; within the development set, 20% was used for probability calibration. Missing data were handled by multiple imputation, and class imbalance by class weighting. The screening threshold was defined as the F1-optimal probability cut-point determined on the development set and held fixed for evaluation on the test set. Metrics were area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), recall, negative predictive value (NPV), and positive predictive value (PPV). Clinical utility was assessed by decision curve analysis (DCA) versus treat-all and treat-none strategies. Feature importance was summarized across models to identify common top contributors.

We analyzed 1,274 outpatients (frailty n=338, 26.5%). The dataset was split into a development set (n=892; non-frailty 655 [model-building 524; calibration 131], frailty 237 [model-building 190; calibration 47]) and a test set (n=382; non-frailty 281, frailty 101). On the test set, VE achieved the highest AUROC (0.682; 95% CI, [0.620–0.739]). The highest AUPRC was observed for LR (0.456 [0.380–0.541]), followed by VE at 0.448 [0.377–0.534]. At the screening threshold, LR yielded the highest recall (0.882 [0.803–0.950]) and NPV (0.902 [0.862–0.941]), while VE showed the highest PPV (0.376 [0.334–0.419]). DCA indicated higher net benefit for VE and LR across a broad range of probability thresholds, notably 0.20–0.30. The most influential predictors across models were Cre, Cl, CTR, age, height, and CRP.

ML models based on routinely available dialysis data can assist frailty screening in outpatient hemodialysis care. VE offered the best overall discrimination, while LR provided high recall and NPV, supporting use when missing frailty would be most harmful. Importantly, these models rely solely on data routinely collected in clinical practice, requiring no additional data acquisition. Consequently, they may enable simple, low-burden frailty screening even for busy dialysis-unit staff. Looking ahead, we will pursue external validation to confirm generalizability and to inform seamless integration into clinical workflows.