Predicting Kyasanur forest disease in resource-limited settings using event-based surveillance and transfer learning

Abstract

In recent years, the reports of Kyasanur forest disease (KFD) breaking endemic barriers by spreading to new regions and crossing state boundaries is alarming. Effective disease surveillance and reporting systems are lacking for this emerging zoonosis, hence hindering control and prevention efforts. We compared time-series models using weather data with and without Event-Based Surveillance (EBS) information, i.e., news media reports and internet search trends, to predict monthly KFD cases in humans. We fitted Extreme Gradient Boosting (XGB) and Long Short Term Memory models at the national and regional levels. We utilized the rich epidemiological data from endemic regions by applying Transfer Learning (TL) techniques to predict KFD cases in new outbreak regions where disease surveillance information was scarce. Overall, the inclusion of EBS data, in addition to the weather data, substantially increased the prediction performance across all models. The XGB method produced the best predictions at the national and regional levels. The TL techniques outperformed baseline models in predicting KFD in new outbreak regions. Novel sources of data and advanced machine-learning approaches, e.g., EBS and TL, show great potential towards increasing disease prediction capabilities in data-scarce scenarios and/or resource-limited settings, for better-informed decisions in the face of emerging zoonotic threats.

Figure 1

Schematic representation for KFD data preparation and TL: (a) Rolling window framework used to create time lags of input features for each prediction horizon: nowcast, one month ahead forecast, and two months ahead forecast; (b) Expanding window approach for yearly retraining of models between October 2016 to October 2019 (Note: The starting date of the model training depended on the year KFD was first reported in the location); and (c) Various TL techniques used for KFD regional-level prediction.

Figure 2

Spatiotemporal distribution of KFD in India (Oct 2010-Oct 2019): Region and district-wise breakdown of KFD outbreak locations (top) along with the time series distribution (stacked area graph) of monthly KFD cases in affected regions (bottom). The green, red, and blue colors represent Goa and Maharashtra, Karnataka, and Kerala regions, respectively. The map was generated using QGIS 3.28.8 software (www.qgis.org).* Chamarajanagar district is present in Karnataka state. However, it was grouped in the Kerala region due to the temporal and spatial proximity of the outbreak with other outbreaks in the region.

Figure 3

The KFD monthly predictions in India. (a) The time series results of the best-performing technique, XGB-weather and EBS model (Oct 2016—2019). The shaded areas represent 95% prediction intervals for the respective prediction horizons. (b) The top 10 features for XGB models based on average permutation feature importance (negative RMSE scores) for nowcasts, 1-month ahead, and 2-months ahead forecasts for weather-only (top) and weather plus EBS models (bottom).

Figure 4

Regional-level KFD monthly predictions in India. The time series result of the best performing XGB models, (a) Karnataka: baseline model, (b) Kerala: direct transfer model, (c) Goa and Maharashtra: direct transfer model. The vertical dotted line indicates the first outbreak season in the region. The shaded areas represent 95% prediction intervals for the respective prediction horizons.