Document resource
Objective To characterise temporal trends in antiretroviral therapy (ART) utilisation and forecast short-term changes in regimen distribution within a provincial HIV treatment programme.Design Population-level longitudinal analysis of administrative drug dispensation data combined with time-series forecasting using recurrent neural network (RNN) models.Setting The British Columbia (BC) Centre for Excellence in HIV/AIDS Drug Treatment Program (DTP), which centrally coordinates and records all publicly funded ART dispensations in BC, Canada.Participants All ART dispensations recorded in the DTP between fiscal years 2014/2015 and 2023/2024.Primary outcome measures Temporal changes in ART dispensation volume and regimen market share. Forecasts of regimen utilisation were generated using drug-specific RNN models, and predictive performance was evaluated using different measures, including mean absolute percentage error (MAPE) and mean absolute error. Forecast uncertainty was presented using 95% prediction intervals.Results ART utilisation patterns changed substantially over the study period. Dispensation of second-generation integrase strand transfer inhibitor (INSTI)-based regimens, particularly dolutegravir- and bictegravir-containing combinations, increased steadily, while utilisation of several older regimens, including generic formulations, declined in relative market share. RNN models demonstrated strong predictive performance during temporal validation, with validation MAPE ranging from 0.27% to 2.26% across regimens. For example, the model forecasting bictegravir/emtricitabine/tenofovir alafenamide utilisation showed a validation MAPE of 0.27%, indicating very small prediction error during temporal validation. Forecasts suggest that second-generation INSTI-based regimens will continue to account for the majority of ART dispensation volume within the provincial programme in the near term.Conclusions Over the past decade, ART utilisation in BC has shifted from older and generic regimens towards newer second-generation INSTI-based combinations. These changes illustrate evolving treatment patterns within a publicly funded HIV programme and may have implications for long-term programme planning as newer therapies replace existing treatment options. In addition, this study demonstrated the potential application of machine learning-based forecasting approaches for analysing pharmaceutical utilisation patterns using administrative health data. Such approaches may complement traditional pharmacoepidemiological methods and provide additional tools for monitoring treatment trends and informing health system planning.