Brettenny, WHolloway, Jennifer PFabris-Rotelli, IDudeni-Tlhone, NontembekoAbdelatif, NontembekoLe Roux, Wouter JManjoo-Docrat, RDebba, Pravesh2026-01-132026-01-132025-121012-94052190-7668https://doi.org/10.1007/s13370-025-01401-xhttp://hdl.handle.net/10204/14581In both the management and modelling of epidemic outbreaks, the ability to determine the start of a wave of infections is of vital importance. Not only does this advantage the modelling of the outbreak, but, if done in real-time, can assist with a nation’s response to the disease. In this study, a bidirectional long-short-term-memory (Bi-LSTM) network is used to determine the start and end of the COVID-19 waves experienced in the district and metropolitan municipalities of Gauteng, South Africa, from 2020-2022 as well as the waves of the cholera outbreaks occurring in the Beira area of Mozambique between 1999 and 2005, in real-time. The problem of real-time scaling of the data prior to the first wave of an epidemic is addressed using globally available real-time information from first waves experienced in other countries and independent territories alongside the observed South African data. The use of the Bi-LSTM predicted starting dates is demonstrated for the second waves of COVID-19 infections experienced in Gauteng in 2020/21. Using the predicted starting dates, spatial-SEIR models are used to predict hospitalisations as a result of COVID-19 infections in each of the district and metropolitan municipalities of Gauteng. The fitted Bi-LSTM model demonstrates effectiveness in predicting the start and end dates of epidemic waves in real-time, allowing for pre-emptive disease modelling and predictions of spread. Moreover, it is shown that the use cases for the fitted model are not limited to COVID-19 studies, but can also be applied to other disease outbreaks that follow similar wave patterns.AbstractenTime series modellingCOVID-19Disease predictionWave detectionSpatial epidemiologyArticlen/a