Impact of Data Quality Degradation on the Robustness of Clinical Regression Models: A Study on ICU Length of Stay

Data quality (DQ) directly impacts the reliability of clinical machine learning systems. This study evaluates how degradations across five DQ dimensions (completeness, coherence, validity, precision, and uniqueness) affect regression models predicting ICU length of stay using the MIMIC-IV dataset. Controlled degradations were applied to evaluation data, while models were trained on clean inputs to reflect realistic deployment conditions. Prediction error increases consistently with degradation, with coherence having the strongest impact, in line with clinical intuition. Differences across model families further indicate that robustness depends in part on model architecture.