(257) Evaluation of different ARIMA models incorporating the impact of COVID-19 for the forecasting of the prescription rates of antibiotics in the UK between 2013 to 2023

Background: Accurate forecasting the usage of medicines of public health interest, such as antibiotics, may allow stakeholders and regulators to coordinate for ensuring the availability of those medicines meet the necessary amount to supply the population. However, the COVID-19 pandemic altered drastically the trends and behaviour of daily life which difficulties forecasting after it.

Objectives: To compare different ARIMA models for the forecasting of usage of antibiotics.

Methods: Monthly incidence rates [IR] (per 100,000 person-years) of 6 antibiotics (including amoxicillin, amoxicillin and clavulanate, azithromycin, cefuroxime, clarithromycin and phenoxymethylpenicillin) from Jul 2013 to May 2023 were computed using all individuals present in CPRD GOLD (UK) mapped to the Observational Medical Outcomes Partnership Common Data Model. AutoRegressive Integrated Moving Average (ARIMA) models were fitted to the obtained IR up to Oct 2022 to test the forecast of the latest 6 month of data (Nov 2022 - May 2023, test period). We compared the mean absolute error (MAE) and mean absolute percentage error (MAPE) accuracy metrics of a 6 vs 12-month validation (referred them as 6m and 12m ARIMA models, respectively), as well as the addition of a model term for the effect of COVID-19 pandemic between Jan 2020 and Dec 2021 (referred them as ARIMAX models). To compare the predictive ability of the models, the average of the ratios of 6m/12m models (eg. 6m MAE in ARIMA/12m MAE in ARIMA) as well as ARIMA/ARIMAX (eg. 6m MAE in ARIMA/6m MAE in ARIMAX) was calculated.

Results: Between 2013 and 2023, IR estimates of antibiotics use with an indication for bacterial infections frequent in winter, such as amoxicillin (IR [95%CI] ranged 5990 [5907-6074] to 31,203 [30,993-31,414]) or azithromycin (ranged 592 [571-613] to 1475 [1427-1523]), show a seasonality pattern with peaks around Nov-March months. After fitting the ARIMA, the validation periods indicate that forecasting with 6-month validation may be more accurate compared to 12-month (average ratio 6m/12m MAE and MAPE: 0.66 and 0.64, respectively), which was corroborated with the test period (0.98 and 0.96, respectively). When comparing ARIMA vs ARIMAX, their average ratio in the validation periods (MAE: 0.96, MAPE: 0.95) seem to indicate ARIMA would be more reliable. However, when the test period was evaluated, the ratios favoured the ARIMAX (average of 1.08 for both, MAE and MAPE).

Conclusions: Use of 6-months for validating the forecast of ARIMA models produced more accurate predictions in the following 6 months than using 12-month validation. The addition of a term that modelled the effect of COVID-19 pandemic between Jan 2020 and Dec 2021 is recommended for time series analyses including this period.