Mission imputable: Effects of missing data processing on infectious disease detection and prognosis

Missing data in medical datasets poses significant challenges for developing effective AI/ML pipelines. Inaccurate imputation can lead to biased results, reduced model performance, and compromised clinical insights. Understanding how different imputation methods affect AI/ML model performance is crucial for ensuring accurate clinical findings.

This study systematically investigates the effects of different imputation methods on AI/ML model performance and their clinical implications.

We investigate the impact of six different missing data strategies on the performance of common classification algorithms for analyzing medical data. The performance was evaluated based on sensitivity and specificity metrics for the tasks of predicting COVID-19 diagnosis and patient deterioration. We also perform feature analysis to understand the clinical implications that the choice of imputation method has.

The findings reveal that the effects of imputation depend on the clinical setting. For a general screening cohort (Einstein Data4u), multivariate imputation by chained equations (MICE) yielded the best performance in clinical settings, resulting in a 26% improvement in sensitivity compared to baseline methods and unmasking critical viral coinfections. Conversely, in an intensive care cohort (MIMIC-IV), complete-case analysis initially showed higher raw predictive metrics. However, further analysis demonstrates that this stems from selection bias driven by informative missingness (MNAR), where testing patterns are intrinsically tied to patient severity. Thus, while imputation recovers diagnostic signals in sparse screening data, it serves as a crucial tool for reducing bias in high-acuity settings.

This study demonstrates the critical impact of missing data imputation on AI/ML model performance and the resulting clinical insights. Our findings underscore the importance of selecting appropriate imputation techniques tailored to the specific characteristics of medical data to ensure accurate and reliable AI/ML predictions. By utilizing a rigorous cross-validation pipeline and a systematic comparison, we provide insights for selecting the most appropriate imputation methods for clinical decision-making applications.