Assessing the impact of missing value mechanisms on anomaly detection in healthcare wearable data

Remote health monitoring using wearable devices has transformed healthcare by enabling continuous observation and early intervention. However, these systems frequently suffer from missing data, which can introduce bias and impair clinical decision-making. The uncertainty surrounding the cause of missing values further complicates data analysis. This paper investigates the impact of different missing data mechanisms (i.e., MCAR, MAR, and MNAR) in the healthcare wearable data anomaly detection task. Using heart rate and step count data from patients with respiratory illnesses, we assess the performance of an anomaly detection method under varying missingness conditions. Our findings demonstrate that more complex mechanisms (MAR and MNAR) significantly degrade detection performance, even at low missing rates, highlighting the importance of developing robust imputation strategies tailored to the nature of missingness.