Weighted keywords over trained models: a transparent signal-detection engine for One Health surveillance

Every surveillance system has to make the same call, over and over: of all the items it just pulled in, which ones are worth a person looking at? Big institutions answer that with trained classifiers and a roster of analysts. Working alone, I had neither. This paper describes how Scout — a One Health surveillance system I built and put on ordinary cloud hosting — makes that call instead, and why I think the plain, transparent approach it uses suits a one-person operation better than a more elaborate one would. Scout sorts each item by what it is. If the item carries text, it gets scored against a hand-built list of about 360 terms, each tagged with a weight of 3, 2, or 1 according to how much the term should worry you; those weights come from the WHO R&D Blueprint, the WOAH disease list, the US CDC bioagent categories, and ICTV naming. Add up the weights of the terms that match, and if the total clears a cut-off and at least one of the matched terms is an actual pathogen, the item is flagged. Items that are just numbers — case counts — skip the text scorer and go to a second channel that watches for sudden jumps against a rolling baseline. I walk through the scoring, the cut-off and the one real incident that made me lower it, and how flagged items end up as dated reports and alerts. I also report what the engine did over a fifty-five-day run: 1,681 signals from real public-health events, including a six-week record that traced a single Ebola outbreak from official alert through to the research response. The point is not that the engine beats anything, but that a single person can read every rule it follows, see why it produced a given result, and correct it when it is wrong. Whether it surfaces events earlier than other surveillance systems is a separate question, and one I have not yet answered.