Mining ambulance call notes to flag panic attack clusters in crowds
When tens of thousands pour into Sydney harbour for New Year's Eve, or a packed MCG roars through an AFL Grand Final, the pulse of a city shifts in ways public-health teams rarely measure. Most monitoring systems are tuned to infectious disease, not to the quieter but very real wave of acute anxiety episodes that can ripple through a stressed crowd. New methods of reading free-text dispatch notes captured by ambulance services are starting to close that gap, giving analysts a near-real-time view of panic attack clusters that would otherwise stay buried in paperwork.
The shift is partly technical and partly cultural. Emergency medical dispatchers in Australia type short narratives alongside codes and timestamps, and those narratives contain words a well-designed text-mining pipeline can pick out: hyperventilation, chest tightness, panic, can't breathe, racing heart. By scanning thousands of those micro-stories every hour, surveillance teams can detect a sudden concentration of similar complaints at a specific venue, often before a hospital emergency department has even processed the first admission.
Reading between the dispatch codes
Every ambulance callout in Australia produces a structured record: location, time, dispatch category, and outcome. What often goes unused is the free-text field where the dispatcher writes what the caller actually said. That paragraph is messy, full of typos, abbreviations like "pt panicking" or "SOB anxious", and half-finished sentences dictated while the call-taker is reassuring someone in tears. For years it was treated as administrative noise. Researchers now treat it as a goldmine of behavioural signal, because people in acute distress describe their symptoms in remarkably consistent language, whether they are at a music festival in Brisbane, a footy match in Parramatta, or a packed train platform at Flinders Street.
The text is too unstructured for a simple keyword search to work reliably. "Panic" might describe the patient, the crowd, or the caller's own reaction. "Can't breathe" might mean asthma, a heart event, or pure anxiety. A useful pipeline therefore combines named-entity recognition, negation handling, and a precise lexicon of anxiety-related terms tuned against historical dispatch notes. The output is a probability score per case rather than a binary flag, which lets analysts tune sensitivity for their own risk appetite.
Once each call is scored, the next job is geographic and temporal aggregation. A handful of probable anxiety events spread across a city is normal background. Eight probable events clustered within a few hundred metres of the same festival gate in an hour is not. That aggregation step is what turns individual narrative snippets into a public-health signal and connects ambulance data to the rest of the surveillance stack.
Building the pipeline from call to cluster
A practical mining pipeline has four moving parts. First, an ingestion layer pulls dispatch notes from the ambulance service's electronic patient care record, usually through a secure API or nightly batch. Second, a pre-processing stage cleans the text: lowercasing, removing identifying information, expanding abbreviations, and splitting dictated run-on sentences. Third, a classification model, often a fine-tuned transformer or a more conservative support-vector machine, assigns each note a label such as "likely panic attack", "possible cardiac", "unclear", or "non-anxiety".
The fourth layer is the clustering engine. It groups cases that share a similar label, fall within a defined time window, and lie inside a spatial radius around an event venue. This is where local geography shapes the output. Setting the radius too small misses clusters that spill into surrounding streets; setting it too wide dilutes real signals into background noise. During a test run linked to a Gold Coast schoolies week, analysts found that a radius of around 400 metres captured most of the relevant calls without dragging in unrelated noise from neighbouring suburbs.
The same pipeline can be repurposed for other acute presentations. By swapping the lexicon and retraining the classifier, the same architecture can scan for heat-related illness at an Australian Open afternoon session, for crowd crush indicators during a City2Surf run, or for substance-related events at a music festival. That reusability is what makes call-note mining attractive to health departments that already struggle to keep up with bespoke dashboards for each event.
Why panic attacks in crowds deserve their own signal
Panic attacks are common, usually self-limiting, and rarely fatal, which is one reason they have slipped under the radar of traditional surveillance. Yet in mass-gathering settings they matter for three reasons. They can overwhelm on-site medical tents and stretch ambulance crews thin at the exact moment they are most needed for serious cases. They can be a precursor to stampede behaviour if a frightened crowd misreads a localised incident as a wider threat. And emerging research suggests that repeated panic episodes in a community can be an early marker of psychosocial stress, including the lingering effects of a public-health emergency or economic shock.
Mass-gathering medicine in Australia has long understood that the medical footprint of an event is shaped as much by minor complaints as by major trauma. A St John Ambulance review of Sydney festival events noted that anxiety and hyperventilation were among the top five reasons people sought help at first-aid posts, alongside sprains, dehydration, and minor lacerations. If a mining pipeline can flag a sudden rise in anxiety calls during the headline act at a Big Day Out-style festival, event medics can pre-position additional counsellors and quiet rooms, and organisers can dim lighting and lower stage volume on adjacent stages.
There is also a longer-term value. A persistent low-grade signal of anxiety-related calls from a particular neighbourhood, even outside event periods, can flag emerging community distress, whether linked to a housing crisis, a localised industrial closure, or the slow burn of a prolonged drought in a regional town. That broader signal is what turns a tactical tool into a strategic asset.
Connecting call-note mining to the wider surveillance picture
Ambulance call notes are only one stream in the multi-channel system that public-health teams use in Australia. Pharmacy sales data, school absenteeism records, and emergency department triage notes all carry complementary signals. A spike in pharmacy sales of short-acting anxiolytics, for instance, often precedes a measurable rise in acute anxiety presentations by 24 to 48 hours, which is the kind of lead time that lets health authorities brief event organisers and on-site medical teams in advance. Readers interested in that adjacent stream can find more on pharmacy sales data for respiratory virus spread and how those patterns were used in practice.
The real power comes from layering. When ambulance notes, pharmacy sales, school absenteeism, and sentinel GP reports all nudge upward in the same postcode around the same time, confidence in the signal rises sharply. When only one stream moves, analysts treat it as a hint rather than a confirmed trend. This kind of triangulation mirrors the approach used during the Sydney Olympics and again during enhanced monitoring around recent international fixtures held in Australian cities.
For Australian health departments, the practical challenge is governance. Free-text dispatch notes contain personal information, even after redaction, and using them for surveillance requires clear data-sharing agreements, ethics approvals, and strict access controls. The good news is that several state ambulance services have already invested in research partnerships that allow de-identified text to be analysed in secure environments, which means the hardest bureaucratic work is partially done.
Australian pilots and what the early data suggests
Early pilots in Australia have been small but informative. A collaboration between a state ambulance service and a university team analysed dispatch notes from a series of major events over a single summer, including a Brisbane Ekka day, a Sydney harbour fireworks night, and an AFL match at Docklands. The mining pipeline identified clusters of probable anxiety calls that had not been flagged by traditional dispatch-code reporting, and in two cases the cluster onset preceded a documented surge in on-site medical tent presentations by enough time for organisers to have intervened.
What the pilots also revealed is the importance of local vocabulary. Australian dispatchers and callers use terms that would not appear in a US or UK lexicon, including colloquialisms and event-specific shorthand. A note about someone "having a mare" or "freaking out" at a rave, or a "dob" in distress at a country show, would be missed by an off-the-shelf model trained only on American English. Fine-tuning on a few thousand locally labelled notes lifted the classifier's precision by a meaningful margin, which is a reminder that even the most elegant NLP pipeline still needs local grounding.
The next frontier is real-time deployment. Instead of running the pipeline overnight, several teams are testing streaming versions that update dashboards every fifteen minutes. For a country that hosts hundreds of large public events every year, from Anzac Day dawn services to regional agricultural shows, that kind of turnaround could quietly reshape how public-health teams think about the link between mental health and mass events.
If you work in public-health monitoring, emergency medical services, or event medicine and want to dig deeper into how these signals are built and combined across data sources, start by exploring the core syndromic surveillance overview as a reference point, then reach out to a research group running text-mining pilots in your state. The building blocks are largely in place; what is needed now are the local partnerships, the ethics agreements, and the willingness to treat a frightened caller in a crowd as a data point worth listening to.