CORKI: A Correlation-Driven Imputation Method for Partial Annotation Scenarios in Multi-label Clinical Problems

Multi-label classification tasks are relevant in healthcare, as data samples are commonly associated with multiple interdependent, non-mutually exclusive outcomes. Incomplete label information often arises due to unrecorded outcomes at planned checkpoints, varying disease testing across patients, collection constraints, or human error. Dropping partially annotated samples can reduce data size, introduce bias, and compromise accuracy. To address these issues, this study introduces CORKI (Correlation-Optimised and Robust K Nearest Neighbours Imputation for Multi-label Classification), a data-centric method for partial annotation imputation in Multi-label data. This method employs proximity measures and an optional weighting term for outcome prevalence to tackle imbalanced labels. Additionally, it leverages different modalities of correlation that consider not only variable values but also missingness patterns. CORKI’s performance was compared with a domain-knowledge-based rule system and the standard sample-dropping approach on three public and one private cardiothoracic surgery datasets with diverse missing label rates. CORKI yielded performances comparable to those of the domain-knowledge approach, establishing itself as a reliable method, while being highly generalizable. Moreover, it was able to maintain imputation accuracy in demanding partial annotation scenarios, presenting drops of only 5% for missing rates of 50%.