Adaptive forecasting of antiretroviral therapy demand using machine learning in India’s national HIV programme

Abstract

Antiretroviral therapy (ART) stock-outs interrupt treatment, increase the risk of virologic failure and drug resistance, and erode the population-level benefits of viral suppression. India’s National AIDS Control Organization (NACO) manages one of the world’s largest public ART programmes, where regimen transitions, evolving formulations, changing treatment guidelines, and procurement-driven fluctuations in drug consumption complicate forecasting. We developed an end-to-end, regimen-specific forecasting workflow to support procurement planning during such periods of instability.

We analyzed monthly national ART consumption data from January 2013 through December 2024. A privacy-preserving synthetic dataset was used for pipeline development, followed by final evaluation on real national consumption time series. We compared three model classes, comprising six models: (1) classical models (Holt–Winters and ARIMA), (2) transformer models (TimesFM, which is a large pre-trained time-series foundation model, and its variant with logarithmically transformed values), and (3) hybrid models (two variants of a hybrid ARIMA-TimesFM residual model).

While the forecast horizon of 18 months remained constant, the train-test period varied across real and synthetic data, as real data were only available until February 2024. For synthetic data, models were trained through June 2023 (test window was July 2023-December 2024), while for real data, models were trained through August 2022 (our test window was September 2022-February 2024). We reported signed percentage deviation to preserve whether models tended to over-predict or under-predict, and selected models by the smallest absolute deviation. We then derived a regimen-specific model-error buffer, applied only to held-out under-prediction, and deployed the workflow through a no-code dashboard.

Forecasting performance was determined using signed percentage deviation (SPD), wherein positive deviation represents under-prediction and negative deviation represents over-prediction. Performance varied across regimens, indicating that no single approach was best-performing for all formulations. On synthetic benchmark data, the smallest absolute deviations ranged from 0.39% for adult AZT+3TC to 5.37% for adult DRV/r.

On real consumption data, classical methods remained competitive for some series, whereas transformer and hybrid models produced better predictive outcomes for others. For instance, for adult AZT+3TC, the Hybrid 70th percentile achieved an SPD of −2.02%, in contrast to the error range of [−15.7, 8.87] for other models. For adult RTV booster, Holt–Winters achieved an SPD of 1.68%, in contrast to the error range of [14.94%, 17.25%] for the other models.

Several formulations, particularly low-volume and transition regimens, nevertheless remained difficult to forecast accurately, underscoring persistent operational uncertainty. This was especially evident across the three pediatric regimens, where all models deviated systematically in the same direction — a more concerning pattern than mere error magnitude alone. For pediatric ABC+3TC, all models over-predicted within a narrow band of [−82.74, −67.43], while for pediatric AZT+3TC and LPV/r 100/25, all models under-predicted, with ranges of [24.93, 63.73] and [18.32, 52.07], respectively.

These findings support a portfolio approach to forecasting in national HIV programmes. Rather than replacing established public-health procurement systems, regimen-specific model selection, directional error reporting, and cautious model-error buffering can strengthen decision support during regimen transitions and other periods of unstable demand.