Digging deep into weighted patient data through multiple-level patterns

Large data volumes have been collected by healthcare organizations at an unprecedented rate. Today both physicians and healthcare system managers are very interested in extracting value from such data. Nevertheless, the increasing data complexity and heterogeneity prompts the need for new efficient and effective data mining approaches to analyzing large patient datasets. Generalized association rule mining algorithms can be exploited to automatically extract hidden multiple-level associations among patient data items (e.g., examinations, drugs) from large datasets equipped with taxonomies. However, in current approaches all data items are assumed to be equally relevant within each transaction, even if this assumption is rarely true. This paper presents a new data mining application targeted to patient data analysis. It tackles the issue of extracting generalized rules from weighted ∗Corresponding author. Tel.: +39 011 090 7084. Fax: +39 011 090 7099. Email addresses: [email protected] (Elena Baralis), [email protected] (Luca Cagliero), [email protected] (Tania Cerquitelli), [email protected] (Silvia Chiusano), [email protected] (Paolo Garza) Preprint submitted to patient data, where items may weight differently according to their importance within each transaction. To this aim, it proposes a novel type of association rule, namely the Weighted Generalized Association Rule (W-GAR). The usefulness of the proposed pattern has been evaluated on real patient datasets equipped with a taxonomy built over examinations and drugs. The achieved results demonstrate the effectiveness of the proposed approach in mining interesting and actionable knowledge in a real medical care scenario.