Triage by Transcript

Leveraging AI for Real-Time Epidemic Detection and Outbreak Response

In an era where artificial intelligence is redefining the contours of innovation, one frontier remains largely untapped: the use of Large Language Models (LLMs) to listen, understand, and act on the voices of public concern in real time. Imagine a system that listens to 911 calls, doctor’s office voicemails, or hospital triage conversations—not just to transcribe them, but to detect the early tremors of an unfolding health crisis. This isn’t science fiction, it’s a near term reality. With today’s advancements in AI and machine learning, we are on the cusp of realizing such a system. But first, let’s unpack what makes it possible.

Beyond the Buzzwords: How LLMs Understand Language

Large Language Models (LLMs) are advanced machine learning models trained on vast amounts of human language data. They use neural networks, specifically transformer architectures, to predict the next word in a sequence. Through this deceptively simple task, they learn the intricacies of grammar, context, meaning, and even tone. From answering complex questions to writing essays, LLMs simulate human-like understanding, making them versatile tools for knowledge work, customer service, content creation, and now, possibly, public health surveillance.

While LLMs are not conscious and do not "understand" language in a human sense, their ability to recognize patterns in unstructured text (or transcribed speech) allows them to function as intelligent filters and interpreters of real-world information.

What Makes Transformers Different

At the heart of every large language model is a revolutionary neural network design called the transformer architecture. Introduced by Google researchers in 2017, transformers replaced older, slower models by allowing machines to “pay attention” to every word in a sentence at once, instead of processing words in order, like a human reading a book line by line. This attention mechanism enables the model to understand relationships between words, even if they're far apart. For instance, in the sentence “She gave the patient her diagnosis,” transformers help the model determine whether “her” refers to the doctor or the patient, based on context clues.

Transformers also allow for highly parallel processing, making it possible to train massive models on billions of words efficiently. This architecture is what gives LLMs their uncanny ability to understand nuance, follow complex instructions, and generate coherent responses. It’s not just about memorizing language; it’s about modeling the relationships and patterns within it.

From Text to Triage: The Case for Real-Time AI Listening

What if LLMs could be used not just for static analysis, but for dynamic, real-time listening? In the context of healthcare, this would mean transcribing incoming calls at clinics, hospitals, and emergency centers and using LLMs to extract vital information: symptoms, urgency, emotional tone, and geographic data.

The idea is to build a system that:

·       Transcribes audio into text using speech-to-text engines.

·       Applies an LLM to identify health concerns, behavioral patterns, or common complaints.

·       Detects spikes or anomalies in symptom reporting (e.g., "fever," "shortness of breath," "rash")

·       Flags trends across geographic locations to notify health officials before a full-blown crisis emerges.

Such a system could significantly reduce response times during emerging outbreaks or environmental health events. Call it syndromic surveillance upgraded by AI and tuned to the human voice.

Where It’s Already Happening: AI in the Wild

While this vision is ambitious, it’s not entirely hypothetical and several initiatives are already underway.

AI-Powered Emergency Call Analysis: Researchers have developed systems that transcribe and analyze 911 calls in real time, using LLMs for named entity recognition and categorizing the severity of emergencies. One notable example is LogiDebrief, a system deployed in Metro Nashville that uses AI and temporal logic to automate the review of emergency calls. It has processed over 1,700 real-world calls, resulting in enhanced dispatch performance, reducing response times.

Another standout is Corti, a Denmark-based real-time AI assistant that listens in on emergency calls to detect signs of cardiac arrest and other medical conditions—often faster than human dispatchers. Corti leverages speech analysis and transformer-based understanding of tone and urgency and has been adopted in several European EMS systems with strong outcomes. In the U.S., Exacom’s Prizym is currently being piloted to transcribe and triage emergency calls. This prototype model classifies urgency levels, recommends dispatch protocols, and supports non-clinical call takers with intelligent triage support.

BEACON Project (Boston University): BEACON is an open-access infectious disease monitoring tool that uses AI to synthesize information from multiple sources, assess threat levels, and generate multilingual reports for public health officials. Though not call-specific, its structure offers a useful blueprint.

CDC and AI: The CDC is actively exploring AI for public health forecasting, including NLP on food borne illness reports and opioid overdose trend analysis. They're also experimenting with nontraditional data sources like audio and images.

These use cases reveal that while full-scale real-time epidemic listening systems aren’t widely deployed yet, the foundational technology is here, and already being tested.

Enhancing Clinical Response Through EHR Integration

A compelling next step for this technology is integrating the structured insights gathered from LLM analysis directly into Electronic Health Records (EHRs). By feeding early symptom patterns or flagged risk indicators into patient charts in near real time, providers can begin building tailored care plans or activating response protocols before patients even walk through the door. This preemptive layering of intelligence not only accelerates diagnosis but also activates clinical decision support tools, enabling interventions for high-risk individuals (such as the elderly, immunocompromised, or those with chronic conditions) before their condition escalates.

Additionally, EHR-integrated LLM tools could automatically surface relevant treatment protocols, suggest diagnostic tests, or recommend patient outreach based on trends detected across similar calls. In this way, AI doesn’t replace the clinician; it augments their situational awareness and catalyzes action.

Fortifying Health Infrastructure and Supply Chains

While an EHR or Practice Management integration is essential, the value of this data extends beyond the clinic. With proper routing, real-time health signal detection from LLMs can feed directly into local health department dashboards, public health surveillance tools, and emergency preparedness systems. By identifying regional spikes in symptoms or distress calls, agencies can rapidly assess which communities need testing centers, medical personnel, or public information campaigns.

Even more critically, this intelligence can be tied into pharmaceutical and clinical supply chain logistics. A sudden uptick in respiratory complaints in a metro area, for instance, could trigger automated reallocation of inhalers, antivirals, or PPE to nearby pharmacies and hospitals, ensuring supplies match projected demand. This moves public health from reactive stockpiling to predictive, data-informed distribution.

Challenges to Confront

Of course, several hurdles remain. Privacy and HIPAA compliance is a significant concern, as voice data in healthcare is highly sensitive. Any AI-powered system must incorporate strict encryption protocols, data anonymization, and stringent access controls to ensure compliance. Another challenge lies in the quality of the input data: callers may be emotional, panicked, or incoherent, making it difficult for models to interpret chaotic or unstructured speech without misclassification or loss of nuance.

False positives and false negatives also pose serious risks. Poorly tuned systems could flood public health systems with unnecessary alerts or miss critical signals altogether. Integration complexity is another barrier—real-time analysis systems must fit seamlessly within existing hospital, EMS, and public health infrastructures without introducing latency or workflow disruptions. Finally, there’s the issue of bias and equity. AI models must be trained on diverse datasets to perform reliably across different accents, dialects, and cultural communication styles to ensure they serve all communities fairly and effectively.

Overcoming the Obstacles: From Risk to Readiness

While the challenges are real, they are far from insurmountable. Privacy and HIPAA compliance can be addressed through a combination of end-to-end encryption, automated de-identification of sensitive data, and edge computing, where transcription and analysis happen locally before any data is transmitted. This approach minimizes exposure while enabling real-time, HIPAA-aligned processing. Additionally, incorporating consent-based data capture practices and transparent governance frameworks can help build public trust in the system.

To improve performance in emotionally charged or chaotic call scenarios, LLMs can be paired with speech recognition models fine-tuned on medical and emergency audio. Incorporating sentiment analysis and speaker diarization (the process of determining who spoke when) helps the model distinguish distress from actionable information. False positives can be reduced through ensemble modeling, where multiple algorithms cross-validate findings, while setting confidence thresholds ensures that only high-certainty alerts are prioritized.

For smooth integration into existing healthcare and public health systems, APIs and HL7/FHIR-compatible interfaces (Hooray for Interoperability!!)should be employed to maintain interoperability with EHRs, hospital systems, and public health dashboards. A modular system design also allows for easy adaptation across diverse environments (i.e., urban and rural settings, emergency response and Telehealth workflows alike). Finally, to mitigate bias, AI systems must be trained on linguistically diverse datasets and subjected to ongoing performance audits across different demographic groups.

Looking Forward: Toward an Ambient Public Health Intelligence

The future of epidemic detection may well be ambient, where AI systems passively and constantly listen, learn, and alert us before catastrophe strikes. By combining LLMs with real-time data streams, predictive models, and verified health information sources, we can create intelligent co-pilots for public health.

These systems could:

·       Flag abnormal symptom clusters within a single city block or other predetermined geographic areas.

·       Suggest triage adjustments in overwhelmed ERs, secondary or tertiary care facilities.

·       Offer real-time updates to the CDC, WHO or even State/Municipal Health Departments based on incoming and outgoing phone interactions.

·       May even be helpful in debunking misinformation by recognizing common myths as they’re repeated in conversation, helping to define population health agendas and outreach.

A Call to Listen—Literally

As healthcare systems around the world grapple with increasingly complex and fast-moving threats, the need for intelligent, anticipatory infrastructure has never been more urgent. LLMs may not be conscious, but they are exceptionally good listeners and, in a future defined by connected systems and proactive care, that may be exactly what we need.

Real-time AI listening isn’t just a technological challenge. It’s an ethical, clinical, and human opportunity to hear what the world is telling us, before it’s too late. Now is the time for technologists, clinicians, and policymakers to collaborate on building the frameworks that make real-time AI listening safe, equitable, and effective.

The infrastructure isn’t the barrier—our imagination is.