Due to the high number of comorbidities, it is essential to accurately predict which PD patients are at high risk of developing MACE. Many studies have focused on building predictive tools for CKD patients, predicting cardiovascular events (AUC = 0.60-0.74), renal failure (AUC = 0.79-0.84), and all-cause mortality (AUC = 0.70-0.82).10. Some studies have focused on predicting the 5-year risk of ESRD events in patients with type 2 diabetes, with AUCs ranging from 0.86 to 0.92.11, 12, 13, 14. In our study, we developed and validated a machine learning model to predict the risk of MACE in PD patients. To contextualize model performance, we benchmarked the model against published risk tools for cardiovascular events in CKD/dialysis patients. These comparisons confirm that our ML approach provides better discrimination than traditional regression-based tools and common CKD risk scores.
Our ML analysis revealed age as an independent risk factor for MACE in PD patients. PD is considered to be patient-centered and requires good self-care abilities, which may decline with age.15,16. However, the widespread use of PD support programs has greatly expanded eligibility for treatment of the elderly and physically frail. In our cohort, PD support was not systematically implemented and chronological age remained a strong predictor of MACE. Aging is associated with structural and functional changes in the heart, further increasing MACE risk. Further research is needed to integrate such programs and assess their cardiovascular effects in older PD patients.
PD patients often have significant comorbidities. Charlson Comorbidity Index (CCI)17 Ideal for evaluating complications in PD patients. CCI remains the most widely used to quantify comorbidity burden in this setting. Among the conditions captured by CCI, diabetes, heart failure, and cerebrovascular disease were the three most frequent conditions in our cohort. In our study, a history of congestive heart failure emerged as a particularly important predictor of MACE, consistent with previous reports. Of note, a CCI score ≥ 5 has been identified as the optimal threshold for predicting cardiovascular and all-cause hospitalizations in PD patients.18highlighting the clinical utility of comorbidity assessment for risk stratification and resource allocation.
Our study revealed significant differences in calcium and PTH levels between the two groups (p< 0.05). Both play important roles in the progression of vascular calcification. Brian Kestenbaum et al.19 found that hyperphosphatemia and secondary hyperparathyroidism were associated with HF, myocardial infarction, and all-cause mortality in patients with CKD. Furthermore, hyperphosphatemia has been established as an independent risk factor for mortality in CVD and CKD patients.20, 21, 22However, unlike previous studies, our analysis did not detect a statistically significant association between serum phosphorus and MACE, probably due to limited sample size.
The ML algorithm in our study also shows that BMI is an independent risk factor for mace in PD patients. Kaya et al.twenty three We found that while BMI was negatively correlated with dialysis adequacy, patients with inadequate dialysis had a significantly increased risk of cardiovascular and cerebrovascular disease, anemia, and other complications. In PD patients, BMI increased and nutritional status improved, but the adequacy of PD decreased with increased body surface area, offsetting the benefit of increased BMI.
Hypertriglyceridemia and HDL-C are important factors for MACE in PD patients. In our study, the ML model predicting risk factors for MACE is consistent with related reports. CKD patients often have lipid metabolic disorders, hyperlipidemia, atherosclerosis, cardiovascular disease, and cerebrovascular disease.twenty fourAmong cardiovascular risk factors, lipid abnormalities account for the majority of the high mortality rate in PD patients.
Hypoalbuminemia has been identified as a major risk factor for cardiovascular disease and infections in PD patients, and this finding is consistent with the results of our ML model. Mehrotra et al. twenty fiveshowed that blood albumin was an independent predictor of death in PD patients. USRDS data further show that in patients with ESRD, the risk of cardiovascular mortality increases by 39% for every 10 g/L decrease in serum albumin, and that patients with malnutrition have a 27% increased risk of cardiovascular mortality compared with well-nourished individuals.26.
EGFR and urine output are important indicators of residual renal function in PD patients. Previous studies have suggested that decreased urine output is associated with fluid overload, which contributes to myocardial hypertrophy, cardiac hypertrophy, and decreased function. 27. Daily urine output in MACE patients was significantly lower than in non-MACE patients (754.4 ± 510.3 mL/day vs. 917.6 ± 500.9 mL/day), confirming the role of urine volume as a predictive risk factor in the ML model. Serum creatinine emerged as a significant predictor of 5-year MACE in our model. This may reflect its dual role as both an inverse indicator of residual renal function and a marker of dialysis adequacy. The 68% MACE incidence observed in our cohort is higher than some previous PD studies, and this difference is primarily due to our broad composite definition. In particular, hospitalizations due to cardiovascular diseases account for 70% of all events, among which heart failure associated with fluid overload is a frequent and unique complication of PD. Inadequate ultrafiltration or nonadherence to fluid restriction can lead to acute decompensated heart failure, which can lead to hospitalization even in the absence of atherosclerotic events.
Hypertension is a significant risk factor for increased cardiovascular morbidity and mortality in dialysis patients 28. According to the ISPD Cardiovascular and Metabolic Guidelines, a blood pressure goal of less than 140/90 mmHg is recommended for PD patients with persistent hypertension.29Hypertension was present in 91.4% of our cohort, and this number was slightly lower than other PD cohorts. This small difference may reflect the use of baseline data collected at the start of dialysis when residual renal function is well preserved and volume overload is less severe, both factors likely to reduce hypertension. Moreover, retrospective underdiagnosis cannot be completely excluded. Nevertheless, the observed prevalence remains high and consistent with registry data, reinforcing that hypertension is an important therapeutic target in this population.
In this study, we developed an ML model to predict MACE in PD patients using baseline clinical and comorbidity data. Of note, our analysis revealed that risk factors such as albumin were less influential than age, BMI, and PTH in predicting MACE in PD patients. The best performing models (Random Forest and XGBoost) consistently identified age as the most important factor for both short-term (1-year) and long-term (5-year) predictions, with HDL-C contributing to short-term risk and dialysis adequacy markers (creatinine, creatinine, eGFR) driving long-term outcomes. The time-dependent importance of these factors (5-year and 1-year models are more predictive) emphasizes the need for long-term monitoring. In particular, the role of PTH in calcium-phosphorus metabolism proved to be important in both time frames. The strong association of PTH supports the current KDIGO guidelines, which target levels below 300 pg/mL. On the other hand, the protective effect of HDL-C suggests that lipid management remains important despite ESRD. The clinical value of our findings extends beyond predictive accuracy to the identification of key modifiable risk factors. By integrating these models into clinical workflows, clinicians can enable timely interventions to improve patient outcomes and reduce healthcare costs. Future research should focus on implementing these predictive tools in real-world settings and developing management strategies targeted at high-risk patients.
