Abstract BACKGROUND AND AIMS Metabolic acidosis (HCO3 < 22 mmol/L) is associated with an increased hazard of death in patients with chronic kidney disease (CKD) [1, 2]. It is further associated with muscle wasting, bone disease, hypoalbuminemia, increased inflammation, protein malnutrition, and progression of CKD [3]. Hospital Civil de Guadalajara is an urban public hospital serving a population of low socioeconomic status for whom HCO3 measurements are not included in their healthcare services and create costs around 70% of which our population cannot afford. We aimed to develop a model allowing us to predict HCO3 in CKD patients and tested its diagnostic performance. METHOD Patients at CKD stage 3–5 from our pre-dialysis clinic were included in this retrospective, cross-sectional study which served as a model development. Creatinine, weight, height, calcium, phosphorus, sodium, potassium, chloride, urea, presence of diabetes, total body water calculated by the Watson equation and eGFR calculated with the CKD-EPI formula were utilized. We performed uni- and multivariate linear regression between measured serum HCO3 and the available variables. A hypothesis-driven step-wise approach for parameter inclusion based on the Akaike Information criterion and optimization of the coefficient of determination was employed to construct the final model predicting serum HCO3. Diagnostic performance for detecting metabolic acidosis using the estimated HCO3 (eHCO3) was evaluated by receiver operating characteristics curve (ROC) analysis and computation of sensitivity, specificity and the positive and negative predictive values (PPV, NPV). Post-test probabilities at different pre-test probabilities (prevalence) were assessed using Bayes’ theorem. RESULTS We screened 1089 patients and included 284 patients [62 (21 IQR), 56% male, 69±14kg, 27% at CKD stage 3 and 34% and 40% at stage 4 and 5, respectively] in the analysis. The final model [eHCO] = 21.47 + 0.61 Ca + 0.43 Na + 0.95 TBW-0.61 Pi-0.51 K-0.53 Cl-0.008 Urea + 0.95 If diabetic had an adjusted R2:0.57, RMSE 2.4 on 275df, P-value: <0.001. The mean difference between HCO3 and eHCO was 0 (95% CI −4.9 to 4.9). Diagnostic performance quantified as AUC ROC 0.87 (95% CI 0.83–0.91). At a prevalence of metabolic acidosis of 56% with a chosen threshold of 20 mEq/L, the model has a sensitivity of 56% and specificity of 95% and a PPV of 94% and NPV of 63%, at a prevalence of 56%. Simulation of the post-test probabilities using Bayesian analysis is shown in Figure 1. CONCLUSION At a considerably high specificity of 95%, we can make predictions with a relatively low risk of treating subjects without metabolic acidosis with bicarbonate. Given the cost saving for the patients and the added insights and diagnostic capability, we believe that the prediction of serum HCO3 by biochemical parameters as done in our model will prove useful in settings with limited resources such as our uninsured urban population in Mexico. Validation of our model is the next step and planned future research.
Read full abstract