Adaptive selection of MICE algorithm parameters for pulmonary function value prediction models

Abstract Artificial intelligence-based medical prediction models are often affected by incomplete datasets containing missing values, which can significantly reduce model reliability and predictive performance. Among various imputation approaches, Multiple Imputation by Chained Equations (MICE) is widely used in medical data analysis because it models each variable using the conditional distributions of the remaining variables while reflecting uncertainty in the imputation process. Chronic obstructive pulmonary disease (COPD) is a chronic respiratory disorder characterized by persistent airflow limitation, requiring early diagnosis and continuous monitoring for improved prognosis. In this study, a pulmonary function test dataset collected from 598 subjects over a two-year period was used to investigate the effects of time-series sampling and missing-data imputation on prediction performance. An adaptive selection method was proposed for the initialization strategy and maximum iteration count of the MICE algorithm based on the statistical characteristics of representative pulmonary function indicators, including FEV1, FVC, FEF25–75%, and PEF. Different missing rates were artificially introduced, and the performances of Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) and multivariate iterative regression-based MICE were compared. Experimental results demonstrate that the proposed adaptive MICE strategy maintains stable predictive performance under increasing missing rates and provides reliable predictive information by effectively reflecting multivariate correlation structures.