Automating infectious disease surveillance and public health reporting

Automated access to timely and accurate clinical information is essential for infection and sepsis management, public health surveillance, and emergency response. However, notifiable disease reporting is largely dependent on electronic laboratory data transfer and follow-up case interviews that are subject to delayed reporting and manual chart extraction that hinder timely case investigation and interfacility communication. We investigated whether an artificial intelligence (AI)-based abstraction, combining interoperable electronic health record data with locally hosted large language models (LLM) in a secure cloud environment, could automatically generate high quality public health reporting forms for reportable public health diseases. Using two case studies (multidrug-resistant organism (MDRO) surveillance and dengue fever case identification and reporting), we integrated structured data and unstructured data from Fast Healthcare Interoperability Resources (FHIR) to generate automated case-report forms. MDRO transfer forms were reviewed by subject matter experts (board-certified infectious disease physicians) who reported no major errors in fidelity. Dengue fever case investigation forms were validated by County of San Diego Public Health Services staff confirming LLM had captured essential information with high concordance as well as additional important context that was not present in the submitted human-completed forms. These findings demonstrate the feasibility of interoperable, AI-based workflows for case-report generation using an interoperable architecture. Further, these findings highlight opportunities to automate and scale syndromic surveillance efficiently, accelerating clinical and public health response, while supporting reliable interfacility communication.

Rapid access to resistance profiles, exposure histories and presumptive diagnoses is critical for timely recognition of reportable illness and precision intervention in emergency departments (ED) and acute care settings, yet often unavailable at the point of care (1, 2). For example, multidrug-resistant organisms (MDRO) transfer forms are informative of antimicrobial selection yet are typically manually completed and transmitted, resulting in delay (3). Similarly, pathogen prevalence data from public health surveillance is essential for detecting infectious disease threats and guiding timely, empiric intervention (4). However, traditional surveillance methods are limited by their dependence on laboratory-based reporting, the inability to reach cases, and fragmented information from medical records. These factors produce response delays that limit the ability of both public health organizations and emergency medicine physicians to respond efficiently to emerging events (5). Syndromic surveillance was introduced to address these gaps by monitoring symptoms, clinical presentations, and preliminary diagnostic data for key syndromes likely to be of public health significance much earlier than laboratory confirmation. Two decades of experience have shown that syndromic surveillance can improve situational awareness and support earlier outbreak detection (6).

Despite its promise, many syndromic surveillance systems continue to face challenges related to variable data quality and inconsistent completeness across reporting sites (7, 8). Workflows often require extracting information from unstructured clinical notes or free-text chief complaints, which is labor intensive and difficult to standardize (9, 10). Prior evaluations have highlighted the need for improved data management processes, novel analytic tools, and greater standardization to enhance the utility of syndromic surveillance systems (11, 12). During the COVID-19 pandemic, these limitations were magnified as public health agencies struggled to receive timely and complete information for case investigation and reporting (13). In current County of San Diego Public Health workflows, reporting typically begins with an electronic laboratory notification and is followed by medical record requests, receipt of records by e-fax, and manual chart review before form completion. In practice, this process may take approximately 4–6 working days, particularly when records are not immediately available and chart abstraction must be performed manually.

Recent advances in artificial intelligence, healthcare interoperability, and cloud computing have created new opportunities to speed and automate these processes. Fast Healthcare Interoperability Resources (FHIR)-based data exchange and secure computing environments now make it possible to integrate clinical features in real-time (14–16). In parallel, large language models (LLMs) have advanced natural language processing in healthcare. Recent studies have demonstrated the ability of these models to interpret complex clinical notes and accurately extract key medical entities directly from unstructured electronic health records (17–20). For example, our prior work proved that LLMs deployed on consumer-grade hardware can accurately abstract complex clinical quality measures, such as the 63-step severe sepsis and septic shock management bundle (SEP-1) (21). Several studies have shown that enhanced electronic case reporting through automation can accelerate the transfer of critical information to public health agencies, although many existing systems cannot incorporate the richer clinical context (e.g., travel history, lifestyle risk factors, exposures) available in unstructured data (22, 23).

To evaluate the feasibility of this artificial intelligence (AI)-driven abstraction approach for public health, we selected MDRO transfer forms and dengue fever case investigations due to their increasing prevalence, real impact on ED and public health workflows, and prioritization by County of San Diego Public Health to address contemporary implementation needs. Beyond the public health demand, these use cases effectively demonstrate the system's capabilities. MDRO surveillance requires combining structured lab results with unstructured clinical notes, such as the patient's isolation status and wounds. This is crucial for safe patient transfers and preventing outbreaks, especially since standard hospital communications frequently fail to adequately convey MDRO status and required precautions (3). Similarly, dengue fever reporting depends on recent travel history and environmental exposures. Because these details are often buried in clinical notes rather than structured fields, automated text extraction would be essential for early detection (24).

Building on our prior work in abstracting complex clinical quality measures (21), we hypothesized that the same approach could support accurate abstraction of public health surveillance and case reporting forms, using the MDRO transfer form and dengue investigation forms as representative pilots.

Study cohorts

We extracted Electronic Health Record (EHR) data in FHIR format for two separate cohorts within the UC San Diego Health (UCSDH) system. The MDRO cohort (N = 21) included all adult inpatients at UCSDH Hillcrest and La Jolla Medical Centers with a positive culture for drug-resistant organisms during the first week in May 2025. The dengue cohort included all patients with a positive dengue serology and molecular result recorded between January 2022 and December 2024 (N = 18). This study received the UCSDH Institutional Review Board (IRB) approval with a waiver of informed consent (IRB #800257).

Interoperability architecture

At UCSDH, Müller et al. previously developed iBiogram, a prototype patient-specific antimicrobial clinical decision support system that integrates over one million historical isolates with patient-level EHR data to support individualized antibiogram generation (25). Our current system operates within a Health Insurance Portability and Accountability Act (HIPAA)-compliant Amazon Web Services (AWS) cloud environment that provides secure, scalable infrastructure for hosting the LLM. This environment functions as a highly secure, isolated network with limited egress as described by Boussina et al. (14, 21). Structured data are retrieved from the Patient, Observation, ServiceRequest and DiagnosticReport FHIR resource types. Unstructured notes are gathered via DocumentReference resources.

Within this environment, we deployed the locally hosted Llama 3.1 8B Instruct model. We selected this specific small language model (SLM) because it is compact, cost-effective, and capable of being hosted on a single virtual machine within our institution's HIPAA-compliant cloud environment (Supplementary Table 1). No external Application Programming Interface (API) calls are made to third-party LLM providers, ensuring that no protected information leaves the institutional firewall. Furthermore, this model was fine-tuned to follow instructions, making it suitable for performing the specific abstraction tasks required by our prompts (Supplementary Table 2). To extract clinical context from unstructured documentation, we implemented a retrieval-augmented generation (RAG) approach. Clinical notes were segmented into 2,000-character chunks and embedded using the Instruct model. For each inference, the top 12 most relevant chunks (k = 12) were retrieved to augment the prompt. The model was used in a few-shot prompting strategy with a temperature of 0.1 to populate predefined public health reporting fields. Structured EHR elements such as microbiology results, isolation status, devices, wounds, and encounter transitions were retrieved through FHIR APIs. The maximum context length used for inference was 8,192 tokens, and this limit was not reached in the evaluated cases. The workflow was executed on an AWS g5.2xlarge instance, with total processing time of approximately 6 min for the evaluated cases and a compute cost of $1.20 per hour. No task-specific fine-tuning was performed (Figure 1 and Supplementary Table 3).

Study validation and statistical analysis

Automated MDRO forms were reviewed by board-certified infectious disease physicians to assess correctness, completeness, and concordance with the source case records. Automatically generated dengue investigation forms were independently evaluated by San Diego County Public Health personnel. Major errors were defined as omission or factual incorrectness of key data elements required for the target form. Since this study was designed as a feasibility evaluation of workflow-integrated form generation, assessment focused on correctness, completeness, and concordance with human review. For the dengue cohort, we report the proportion of forms in which the key travel exposure field was correctly captured, together with exact 95% Clopper-Pearson confidence intervals. Analyses were descriptive given the small size and pilot study design.

Automated MDRO transfer form generation

From May 5 to May 9, 2025, all eligible inpatients with microbiology-confirmed MDROs were processed through the automation abstraction workflow. The system retrieved structured data from FHIR resources including demographics, medical history, diagnoses, details of current antimicrobial therapy, culture source, collection date, organisms, and carbapenemase sensitivity. Additional structured clinical elements such as the presence of wounds, venous and arterial lines, supportive airways, and isolation status were also extracted. Unstructured information from clinical notes was incorporated to capture the reason for transfer, symptoms, objective findings and patient status. These structured and unstructured components were combined to generate an MDRO interfacility transfer form for each patient (Figure 2). Reviewer assessment found 100% concordance between the generated MDRO transfer forms and the source case records for the evaluated cases, with no major fidelity errors identified.

Dengue case investigation form automation

All UC San Diego Health patients with a positive dengue antibody laboratory result between January 2022 and December 2024 (N = 18) were evaluated using the same automated abstraction workflow. Generated investigation forms were reviewed by County of San Diego Public Health Services (Figure 3). Among the 18 forms, 17 correctly captured the travel exposure history field, corresponding to a correctness proportion of 94.4% (95% CI, 72.7%–99.9%). An error analysis was conducted on the single discordant case, identifying that the error was an omission of specific travel information. Conversely, two cases contained additional context that clinicians had manually omitted: one case in which travel duration in India and timing of subsequent fever onset were extracted by the system but missing in the original case report form, and another in which the system captured additional concerning symptoms that had been omitted.

In this study, we provide preliminary evidence that locally hosted LLMs combined with interoperable EHR data may support automated clinical and public health reporting. Based on expert comparison and review of clinical records, both case studies (MDRO transfer forms and dengue investigation forms) showed high reviewer-assessed correctness and concordance with source records, supporting the feasibility of operational use in these pilot settings. These findings suggest that automated abstraction can reduce manual workload and improve the timeliness of reporting (22, 23).

For the MDRO pilot, the generated forms captured key pathogens, carbapenemase sensitivity, culture source and date, and current antimicrobial therapy, including medication dose, frequency, and treatment dates. Unstructured notes contributed clinically relevant context including the reason for transfer, symptoms, and changes in mental status. An expert reviewer confirmed that the generated forms matched the underlying record. This shows that structured EHR data and unstructured note extraction can support reliable interfacility communication. For dengue, the system correctly captured travel histories in most cases, and in some instances, included additional useful context. These findings suggest that LLMs can extract epidemiologic details that may be missed during manual review. In current public health workflows, reporting often relies on laboratory notification followed by manual record retrieval and chart review, which can introduce delays. By contrast, the FHIR-based abstraction workflow can assemble structured and unstructured clinical information and generate draft reporting forms in near real time once the relevant data are available.

These results illustrate the broader potential of AI-enabled abstraction within existing interoperability infrastructure. By integrating structured and unstructured data through FHIR resources and secure cloud-based processing, this system offers clear advantages for public health. Compared with workflows that depend on laboratory confirmation followed by manual record retrieval and review, this approach has the potential to support earlier identification of symptom patterns and exposure-based insights, accelerating the response to emerging outbreaks (22). Faster and more complete reporting may contribute to improved community safety by supporting more timely public health responses to infectious threats.

Beyond public health reporting, these findings point to a broader model of health system vigilance in which the emergency department serves as the front line. More timely and complete information for transfer partners, as demonstrated by the MDRO transfer from, can improve clinical decision-making while protecting other patients from drug-resistant infections. Similarly, rapid surveillance and reporting of critical infectious diseases can accelerate detection and ease the coordination burden on public health agencies, freeing clinician time for direct patient care. Although this study focused on two use cases, the underlying architecture could be adapted to other settings, including inpatient wards, long-term care facilities and outpatient clinics.

Despite these strengths, this study has several important limitations. First, both pilots involved small cohorts and were conducted at a single health system, limiting generalizability. These findings should therefore be interpreted as feasibility demonstrations rather than validation of a fully generalizable system. Second, this study emphasized workflow-level feasibility rather than formal natural language processing (NLP) benchmarking. We therefore relied on expert and public health reviewer assessment of correctness, completeness, and concordance rather than on an annotated gold-standard dataset or direct comparison with alternative language models. Third, the abstraction workflow was optimized for the documentation practices, FHIR infrastructure, and operational context of UC San Diego Health, and may not transfer directly to institutions with different EHR systems, data standards, or reporting workflows. Future evaluations should include prospective deployment, multi-site assessment, and measurement of operational impact, including time savings and completeness of reporting.

Nevertheless, these early findings demonstrate that an interoperable and scalable architecture can support automated generation of public health case reports with reviewer-assessed fidelity in pilot use cases. As health systems continue to expand their use of FHIR and LLM infrastructure, such approaches may enable more widespread and higher quality health surveillance, improved interfacility communication, and more efficient collaboration between healthcare organizations and public health agencies.

This study provides preliminary evidence that locally hosted LLMs integrated with interoperable EHR data may support automated generation of public health case-report forms in real-world workflows. In pilot MDRO and dengue use cases, the approach showed feasibility, high reviewer-assessed fidelity, and the ability to capture clinically relevant context from both structured and unstructured data. With further prospective and multi-site validation, this interoperable AI-enabled workflow may help reduce manual abstraction burden and improve the timeliness and completeness of public health reporting.