The current deadly outbreak of Ebola in Africa is the latest in a string of diseases to grab headlines in recent years — from COVID-19 to mpox and hantavirus. Outbreaks are easier to control if they can be detected early and contained. But many emerging, evolving or unfamiliar diseases fall through the cracks of current surveillance systems.
Better diagnostics could have limited this Ebola outbreak
Epidemics are tracked using three approaches. First, event-based tracking monitors local information sources — media reports, social-media posts, community alerts and health-worker reports — to detect early signals of an outbreak. Second, genomic sequencing is used to confirm a pathogen’s identity1,2. And third, syndromic surveillance looks for people at risk of being infected, by identifying and monitoring individuals who present with characteristic symptoms.
But these strategies don’t account for the fact that an emerging pathogen might present with unusual or atypical symptoms. The presentation of symptoms can change over time, as a pathogen evolves3, and they can also vary between people, who might have different levels of immunity. And co-infections that present with similar symptoms to those caused by a pathogen can obscure diagnoses4.
This gap in surveillance has caused problems time and again.
For example, during the early 2020 wave of SARS-CoV-2, initial diagnoses relied on the presence of fever and respiratory distress5,6. In many countries, it took months for another common symptom, loss of smell, to be accepted as an indicator of COVID-19, allowing individuals to bypass isolation protocols. New variants of the virus brought with them different symptoms: Omicron was associated with cold- and influenza-like symptoms more often than it was with the loss of smell or taste7.
COVID’s origins: what we do and don’t know
In the 2022 mpox outbreak, which affected mainly Europe and North America, some individuals with the infection presented with genital and anal sores, rather than the widespread rash associated with classic mpox. These people were initially misdiagnosed with other conditions, including sexually transmitted infections, potentially delaying recognition and public-health action8.
Similarly, Nipah virus is proving hard to diagnose in clinics in Bangladesh and India because its symptoms can shift between brain-tissue inflammation (causing confusion and seizures) and breathing difficulties9.
Unless clinicians recognize that unusual symptoms warrant investigation, even the most advanced tracking systems cannot keep up. Here we set out five steps to improve health systems’ abilities to recognize emerging pathogens. Together, these steps form a loop in which clinical signals are detected, validated, translated into surveillance and diagnostic guidance, then rapidly fed back to front-line clinicians. We call this approach ‘symptom preparedness’.
Flag clusters of unusual cases
Formal mechanisms are needed for clinicians to act on the intuition that they are dealing with something unusual, such as several people presenting with the same unrecognized set of symptoms. National public-health agencies and health ministries should provide reporting procedures for documenting such ‘clinical drift’ through existing digital surveillance tools, such as the online platform District Health Information Software 2 (DHIS2).
Credit: Tony Karumba/AFP via Getty
To prevent false-positive alerts, a second senior clinician or hospital epidemiologist would need to verify that the cluster is atypical. Clinicians could then run an algorithm on DHIS2, or an equivalent tool, that uses statistical checks to confirm that these symptoms are unlikely to be caused by routine seasonal illnesses. Once verified, these reports would automatically trigger localized digital alerts to national public-health intelligence centres, where epidemiologists would cross-reference regional data before formal surveillance measures are activated10.
Use AI to flag atypical cases
Machine-learning models that are trained on temporally and geographically labelled clinical data might help physicians to detect atypical cases in the future, by identifying people whose symptoms differ from patterns normally observed in the same population. The models should not provide a diagnosis, but could flag cases that warrant further examination. In this way, artificial-intelligence tools would augment rather than replace clinical judgement11,12.
Such AI systems are being developed — although they are currently in their infancy13. Implementation would require reliable Internet connectivity, adequate data quality, testing and evaluation in the intended local setting, ongoing performance monitoring and privacy safeguards.
Low-resource settings might lack the facilities needed to run an AI model. But health workers could submit anonymized clinical information to a country-wide centralized AI system through a phone app, SMS-based reporting system or existing electronic registries. The AI system would return an alert when needed.
Adopt tiered case definitions
Public-health authorities should transition away from using only standardized clinical criteria provided by agencies such as the World Health Organization to classify disease cases. Instead, they should adopt dynamic, tiered case definitions that distinguish suspected, probable and confirmed cases.
AI biohazard hype must not divert attention from the real risks we face now
Committees of epidemiological specialists would stratify diagnostic criteria into tiers and update them dynamically — on a weekly or rolling basis — as real-time clinical data emerge10. For example, the standard symptoms of mpox conventionally included a generalized rash. If a dynamic diagnosis system had been available in 2022, public-health authorities could, within days of recognizing atypical clusters, have expanded the ‘suspected case’ definition to include localized genital sores. This rapid adaptation could have helped to identify misdiagnosed individuals sooner.
