Machine Learning Models for Predicting Zoonotic Disease Hosts and Preventing Ebola Outbreaks

In April 2014, a team of ecologists, veterinarians, and an anthropologist traveled to the Guinean village of Meliandou to investigate the origin of an Ebola outbreak, focusing on a two‑year‑old boy named Emile as the index case and exploring the role of insect‑eating bats as possible reservoirs.

The researchers interviewed villagers, sampled local primates, and collected bats, but the suspected bat roost had burned down before definitive conclusions could be drawn.

Recognizing the difficulty of pinpointing zoonotic hosts through field work alone, the author—a disease ecologist at the Cary Institute—applied computer modeling and machine‑learning techniques to predict which wildlife species might carry future pathogens.

Using a classification‑and‑regression‑tree (CART) algorithm, the model was trained on known host data, assigning binary labels (1 = known zoonotic host, 0 = unknown) and incorporating over 50 biological traits such as body size, metabolic rate, reproductive rate, and geographic range.

Iterative tree building and boosting generated thousands of trees; an ensemble of these trees improved classification accuracy, achieving about 90 % accuracy on a held‑out test set of rodents.

The model identified 58 previously undocumented potential zoonotic hosts and highlighted two geographic hotspots—mid‑western United States and a belt across Central Asia and the Middle East—where surveillance of rodents like the northern grasshopper mouse and the reed vole could be prioritized.

Field validation quickly confirmed two new rodent hosts: the red‑backed vole (Myodes gapperi) carrying echinococcosis and the Anatolian vole (Microtus guentheri) as a leishmaniasis reservoir.

The approach also flagged several bat species worldwide as possible carriers of hemorrhagic fevers such as Ebola and Marburg, raising questions for public‑health officials about surveillance in regions without recorded outbreaks.

Advantages of the machine‑learning method include handling incomplete data, mitigating sampling bias, and uncovering complex trait interactions without predefined rules, thereby providing actionable intelligence for ecologists and health agencies.

Overall, the study demonstrates how AI can transform disease‑ecology research from reactive investigation to proactive prediction, offering a template for anticipating and preventing future zoonotic disease emergence.