Computational and Human Intelligence Methods for Constructing Practical Risk Prediction Models: An Application to Cardio-Renal Outcomes in Non-Diabetic CKD Patients

The current investigation aimed to develop a novel approach for risk prediction modeling of clinical outcomes in common diseases based on computational and human intelligence techniques with no a priori input on risk factors using real-world individual patient-level data from administrative claims. Bootstrapping multivariable Cox regression and ant colony optimization were employed to develop time-to-first-event risk prediction models of cardio-renal outcomes in patients with non-diabetic chronic kidney disease (CKD) as a demonstration case. A cohort of 504,924 non-diabetic CKD stage 3 or 4 patients enrolled from 2008 to 2018 were identified in the US administrative de-identified claims database, Optum Clinformatics (R) Data Mart. Initial set of potential risk factors was derived from patient-level data at baseline and included more than 540,000 variables. Risk prediction models of hospitalization for heart failure, worsening of CKD stage from baseline and a renal composite outcome of end-stage kidney disease, kidney failure or need for dialysis in non-diabetic CKD stage 3 or 4 were built. Final model optimization was conducted using human intelligence to combine clinically similar features and build equivalence classes to ensure that risk factors included in the final model were routinely collected and easily interpretable by healthcare providers. Demonstrated validity of our approach in non-diabetic CKD offers opportunities for application in other therapeutic areas, with the potential to improve overall prognosis and decrease the clinical and economic burden of common diseases. The approach enables developing practical prediction models for risk estimation in routine clinical practice.