Design of suitable statistical methods for time series analysis and anomaly detection in antibiotics consumption data

Overuse and misprescription of antimicrobials in human medicine reduces the safety and efficacy of treatment, promotes the emergence and spread of resistant pathogens, and generally endangers the sustainability of the health system. Analyzing and assessing the adequacy of antibiotic consumption in a healthcare facility helps detect poor prescription habits, potential frauds, or significant milestones in the development of an epidemiological situation. Statistical techniques to detect anomalies in consumption provide an important guide to microbiology experts for implementation of most efficient coordinated measures. In this paper, statistical tools suitable for analysis and anomaly detection in time series were reviewed and tested on de-identified real data from a Czech healthcare facility obtained in collaboration with Janiga Labs, s.r.o. The LOF technique, based on neighborhood density comparison, was applied alongside the time-series decomposition based methods. The measures of relative extremity based on ESD test and the modified Z-score indicated a total of 251 possible anomalous measurement. The 15 most significant outliers identified by the proposed methods were subsequently consulted with the head of microbiology at the data providing healthcare facility. This interpretation showed that the methods successfully identified anomalies in antibiotic consumption. Adoption of techniques for automatic notification of unusual use of ATB in hospital systems could significantly improve control and prevention measures for prudent use of antimicrobials and thus support achieving effective antibiotic policy objectives.