⚠ EARLY-WARNING ALARM — fused signal score has crossed the threshold
Official case counts are the most reliable disease-surveillance signal a health department has, and also the slowest: infection, incubation, symptom onset, a decision to seek care, testing turnaround and reporting all have to happen before a single confirmed case shows up in the official tally. Long before that pipeline completes, though, people are already searching their symptoms online, buying fever reducers and cough remedies, and — closer to the end of the delay — being admitted to hospital. Each of these behaviours leaves a noisy digital trace that runs ahead of, or closer to, the true underlying epidemic curve. This simulation generates a synthetic "true" outbreak that nobody can observe directly in real time, derives three realistic observable signals from it with different lags and noise levels, and lets you build a live early-warning alarm by fusing them together.
As the simulated "today" advances day by day, only the data up to that day is ever revealed — exactly like a real dashboard. A weighted, z-scored combination of each signal's recent trend produces a single fused score; when it crosses your alarm threshold, the simulation records the day. Once the hospital-admissions signal, on its own, would have crossed an equivalent threshold, the simulation reports the gap between the two — the number of days of lead time signal fusion actually bought you.
Why do search trends and pharmacy sales lead official case counts?
People typically search for symptoms in the first day or two of feeling unwell, well before they decide to seek formal care, get tested, and have that test result flow into a public case count. Over-the-counter pharmacy purchases follow a similar but slightly later pattern, once symptoms are established enough to prompt a pharmacy visit. Hospital admissions and confirmed case counts sit at the end of the pipeline: incubation period, symptom onset, decision to seek care, testing turnaround and reporting lag can together add a week or more. This simulation builds the search signal by shifting the true underlying infection curve backward by 10–14 days with heavy noise, pharmacy sales by 5–7 days with moderate noise, and hospital admissions by a 3–5 day lag with the least noise of the three — reproducing the real ordering of care-seeking behaviour.
What does it mean to "fuse" several noisy leading indicators into one early-warning score?
Signal fusion means combining several imperfect observations of the same underlying process into a single decision statistic, so noise in any one signal is partly averaged out while the shared true trend reinforces itself. This simulation computes a short rolling slope for each of the three signals, converts each slope into a z-score (standard deviations above its own recent history), and takes a weighted sum of the three z-scores using the slider weights. The result is a single fused score each day. Because search trends genuinely lead the outbreak they get the highest default weight; hospital admissions, the most reliable but slowest signal, get the lowest weight.
Why does combining noisy leading indicators beat waiting for a single lagging one?
A very noisy but early signal used alone triggers false alarms constantly, because its noise floor is comparable to the real outbreak signal. A clean but late signal is reliable, but by the time it moves most of the useful warning time is gone. Averaging several partially independent noisy signals reduces the combined noise roughly by the square root of the number of signals, while the shared true epidemic trend adds up directly — so the fused score reaches a statistically confident deviation from baseline sooner than the noisy signals individually, without inheriting the full delay of the clean-but-late signal. This is the same logic behind ensemble forecasting and sensor fusion generally.
Every day of lead time is a day to pre-position testing capacity, alert clinicians to a specific presentation, activate surge staffing plans, order additional supplies, or issue public advisories before hospitals are under strain. Historical syndromic-surveillance systems have often detected respiratory-disease upswings one to three weeks after infections were already spreading; catching the same signal even a week earlier can be the difference between a coordinated response and a reactive scramble. The simulation quantifies this directly: it compares the day the fused score first crosses the alarm threshold against the day the hospital-admissions signal alone would have crossed an equivalent threshold, and reports the gap in days.
Any statistical alarm trades sensitivity against false-alarm rate, governed here by the threshold slider. A low threshold catches real outbreaks earlier but also fires on ordinary noise, costing credibility and resources each time responders mobilise for nothing. A high threshold cuts false alarms but delays genuine detection, sometimes past the point where early intervention still matters. Because search and pharmacy data are especially noisy and can be driven by non-epidemic causes such as media coverage or unrelated seasonal illness, the weights and threshold in a real system are tuned against a long history of past outbreaks and non-outbreaks, and even then some false alarms are accepted as the cost of catching true ones early.
For each signal the simulation takes a rolling 7-day slope: (value today minus value 7 days ago) divided by 7. It keeps an expanding, causal record of that slope from early in the outbreak up to "today" and computes the slope's mean and standard deviation using only data seen so far, never looking into the future, exactly as a live dashboard would have to. Each day's slope is expressed as a z-score: (slope minus historical mean) divided by historical standard deviation. The fused score is the weighted average of the three signals' z-scores, using the normalised slider weights. An alarm fires the first day the fused score exceeds the threshold, and that crossing day is compared against the day the hospital-admissions z-score alone would cross the same threshold.
The CDC's ILINet and the FluSight / COVID-19 Forecast Hub ensembles combine outpatient visit data, hospital admissions, wastewater surveillance, and increasingly search and retail data into blended nowcasts, much like the fusion score here. Google Flu Trends (2008–2015) tried to use search volume alone to estimate flu activity in near-real time and notably over-predicted the 2013 season, partly because it was never fused with slower but more reliable ground-truth signals and its model was not retrained as search behaviour drifted — a cautionary tale for exactly the false-alarm risk described above. Modern systems generally follow the multi-signal fusion approach this simulation demonstrates: never trust one noisy leading indicator alone, but don't ignore it either.