CLINICAL APPLICATION OF A STATISTICAL LEARNING APPROCH FOR PREDICTING DRY WEIGHT IN HEMODIALYSIS USING BIG DATA ANALYTICS

In this study, hemodialysis medical big data from the management platform of our medical center were utilized to construct an accurate assessment model for dry weight in maintenance hemodialysis (MHD) patients using statistical learning techniques. The model's accuracy in estimating dry weight was further validated through internal and prospective validation procedures.

Patients who underwent maintenance hemodialysis (MHD) at our center during January 1, 2019, and February 8, 2021, and whose electronic health records were available on the hemodialysis information platform. Eligible patients were aged 18 to 85 years old  and had undergone hemodialysis for more than three months. Only those maintaining stable dialysis status, including individuals with preserved residual renal function, were evaluated. Patients with fewer than 30 weeks of hemodialysis treatment data or substantial missing clinical information were excluded. Participants were randomly assigned in a 3:1 ratio to either the model development group or the internal validation group. Baseline characteristics, clinical parameters, biochemical laboratory results, and other relevant diagnostic and therapeutic data were retrospectively collected through the hemodialysis big data platform. Key variables such as the date of hemodialysis session, pre-dialysis body weight, post-dialysis body weight, and ultrafiltration volume were extracted. A statistical learning approach was employed to develop a predictive model for dry weight in hemodialysis patients, enabling estimation of ultrafiltration volume per session, with internal validation conducted to assess prediction accuracy. For external prospective validation, data from MHD patients treated at our center between August 1, 2021, and December 31, 2021, were collected to evaluate the predictive performance in real-time clinical settings.

A total of 131 patients were enrolled in this study, comprising 97 patients in the modeling group and 34 patients in the internal validation group. An additional 97 patients were included in the prospective validation cohort. No statistically significant differences were observed in gender, age, clinical characteristics, or baseline laboratory parameters (except for hemoglobin levels) across the three groups. The results of both internal and prospective validation demonstrated high prediction accuracy: the median absolute errors between predicted ultrafiltration volume (preUFVol) and actual ultrafiltration volume (actUFVol) were 141.30(85.74, 222.55)mL and 145.85(83.67, 233.24)mL（P=0.457）, respectively. The coefficient of determination (R²) was 0.940 for the internal validation and 0.938 for the prospective validation, with corresponding root mean square errors (RMSE) of 177.97 and 189.02. The proportion of predictions within ±300 mL of the actual value (P300) was 87.94% and 85.75%, respectively (P =0.059), indicating strong predictive performance. Bland-Altman analysis revealed good agreement between preUFVol and actUFVol in both validation sets (internal validation: bias=−9.987, SD =178.8; prospective validation: bias =6.646, SD=189.6). Subgroup analyses according to disease stability, presence of residual renal function, and diabetic status showed no significant differences in model performance, suggesting consistent accuracy across diverse patient subgroups.

Based on hemodialysis medical big data, our study employs statistical learning techniques to develop an accurate evaluation model for dry weight in MHD patients. The proposed model demonstrates high predictive accuracy and offers an intelligent, personalized decision-support tool for optimizing dry weight management and precise ultrafiltration volume assessment in MHD.